Daily Papers

Despite promising performance on open-source large vision-language models (LVLMs), transfer-based targeted attacks often fail against black-box commercial LVLMs. Analyzing failed adversarial perturbations reveals that the learned perturbations typically originate from a uniform distribution and lack clear semantic details, resulting in unintended responses. This critical absence of semantic information leads commercial LVLMs to either ignore the perturbation entirely or misinterpret its embedded semantics, thereby causing the attack to fail. To overcome these issues, we notice that identifying core semantic objects is a key objective for models trained with various datasets and methodologies. This insight motivates our approach that refines semantic clarity by encoding explicit semantic details within local regions, thus ensuring interoperability and capturing finer-grained features, and by concentrating modifications on semantically rich areas rather than applying them uniformly. To achieve this, we propose a simple yet highly effective solution: at each optimization step, the adversarial image is cropped randomly by a controlled aspect ratio and scale, resized, and then aligned with the target image in the embedding space. Experimental results confirm our hypothesis. Our adversarial examples crafted with local-aggregated perturbations focused on crucial regions exhibit surprisingly good transferability to commercial LVLMs, including GPT-4.5, GPT-4o, Gemini-2.0-flash, Claude-3.5-sonnet, Claude-3.7-sonnet, and even reasoning models like o1, Claude-3.7-thinking and Gemini-2.0-flash-thinking. Our approach achieves success rates exceeding 90% on GPT-4.5, 4o, and o1, significantly outperforming all prior state-of-the-art attack methods. Our optimized adversarial examples under different configurations and training code are available at https://github.com/VILA-Lab/M-Attack.

Can machines truly think, reason and act in domains like humans? This enduring question continues to shape the pursuit of Artificial General Intelligence (AGI). Despite the growing capabilities of models such as GPT-4.5, DeepSeek, Claude 3.5 Sonnet, Phi-4, and Grok 3, which exhibit multimodal fluency and partial reasoning, these systems remain fundamentally limited by their reliance on token-level prediction and lack of grounded agency. This paper offers a cross-disciplinary synthesis of AGI development, spanning artificial intelligence, cognitive neuroscience, psychology, generative models, and agent-based systems. We analyze the architectural and cognitive foundations of general intelligence, highlighting the role of modular reasoning, persistent memory, and multi-agent coordination. In particular, we emphasize the rise of Agentic RAG frameworks that combine retrieval, planning, and dynamic tool use to enable more adaptive behavior. We discuss generalization strategies, including information compression, test-time adaptation, and training-free methods, as critical pathways toward flexible, domain-agnostic intelligence. Vision-Language Models (VLMs) are reexamined not just as perception modules but as evolving interfaces for embodied understanding and collaborative task completion. We also argue that true intelligence arises not from scale alone but from the integration of memory and reasoning: an orchestration of modular, interactive, and self-improving components where compression enables adaptive behavior. Drawing on advances in neurosymbolic systems, reinforcement learning, and cognitive scaffolding, we explore how recent architectures begin to bridge the gap between statistical learning and goal-directed cognition. Finally, we identify key scientific, technical, and ethical challenges on the path to AGI.

As Large Language Models (LLMs) become widely accessible, a detailed understanding of their knowledge within specific domains becomes necessary for successful real world use. This is particularly critical in public health, where failure to retrieve relevant, accurate, and current information could significantly impact UK residents. However, currently little is known about LLM knowledge of UK Government public health information. To address this issue, this paper introduces a new benchmark, PubHealthBench, with over 8000 questions for evaluating LLMs' Multiple Choice Question Answering (MCQA) and free form responses to public health queries, created via an automated pipeline. We also release a new dataset of the extracted UK Government public health guidance documents used as source text for PubHealthBench. Assessing 24 LLMs on PubHealthBench we find the latest private LLMs (GPT-4.5, GPT-4.1 and o1) have a high degree of knowledge, achieving >90% in the MCQA setup, and outperform humans with cursory search engine use. However, in the free form setup we see lower performance with no model scoring >75%. Therefore, whilst there are promising signs that state of the art (SOTA) LLMs are an increasingly accurate source of public health information, additional safeguards or tools may still be needed when providing free form responses on public health topics.

We evaluated 4 systems (ELIZA, GPT-4o, LLaMa-3.1-405B, and GPT-4.5) in two randomised, controlled, and pre-registered Turing tests on independent populations. Participants had 5 minute conversations simultaneously with another human participant and one of these systems before judging which conversational partner they thought was human. When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time: significantly more often than interrogators selected the real human participant. LLaMa-3.1, with the same prompt, was judged to be the human 56% of the time -- not significantly more or less often than the humans they were being compared to -- while baseline models (ELIZA and GPT-4o) achieved win rates significantly below chance (23% and 21% respectively). The results constitute the first empirical evidence that any artificial system passes a standard three-party Turing test. The results have implications for debates about what kind of intelligence is exhibited by Large Language Models (LLMs), and the social and economic impacts these systems are likely to have.

Large language models (LLMs) can perform reasoning computations both internally within their latent space and externally by generating explicit token sequences like chains of thought. Significant progress in enhancing reasoning abilities has been made by scaling test-time compute. However, understanding and quantifying model-internal reasoning abilities - the inferential "leaps" models make between individual token predictions - remains crucial. This study introduces a benchmark (n = 4,000 items) designed to quantify model-internal reasoning in different domains. We achieve this by having LLMs indicate the correct solution to reasoning problems not through descriptive text, but by selecting a specific language of their initial response token that is different from English, the benchmark language. This not only requires models to reason beyond their context window, but also to overrise their default tendency to respond in the same language as the prompt, thereby posing an additional cognitive strain. We evaluate a set of 18 LLMs, showing significant performance variations, with GPT-4.5 achieving the highest accuracy (74.7%), outperforming models like Grok-2 (67.2%), and Llama 3.1 405B (65.6%). Control experiments and difficulty scaling analyses suggest that while LLMs engage in internal reasoning, we cannot rule out heuristic exploitations under certain conditions, marking an area for future investigation. Our experiments demonstrate that LLMs can "think" via latent-space computations, revealing model-internal inference strategies that need further understanding, especially regarding safety-related concerns such as covert planning, goal-seeking, or deception emerging without explicit token traces.

Generative search engines and deep research LLM agents promise trustworthy, source-grounded synthesis, yet users regularly encounter overconfidence, weak sourcing, and confusing citation practices. We introduce DeepTRACE, a novel sociotechnically grounded audit framework that turns prior community-identified failure cases into eight measurable dimensions spanning answer text, sources, and citations. DeepTRACE uses statement-level analysis (decomposition, confidence scoring) and builds citation and factual-support matrices to audit how systems reason with and attribute evidence end-to-end. Using automated extraction pipelines for popular public models (e.g., GPT-4.5/5, You.com, Perplexity, Copilot/Bing, Gemini) and an LLM-judge with validated agreement to human raters, we evaluate both web-search engines and deep-research configurations. Our findings show that generative search engines and deep research agents frequently produce one-sided, highly confident responses on debate queries and include large fractions of statements unsupported by their own listed sources. Deep-research configurations reduce overconfidence and can attain high citation thoroughness, but they remain highly one-sided on debate queries and still exhibit large fractions of unsupported statements, with citation accuracy ranging from 40--80% across systems.

As reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for scaling reasoning capabilities in LLMs, a new failure mode emerges: LLMs gaming verifiers. We study this phenomenon on inductive reasoning tasks, where models must induce and output logical rules. We find that RLVR-trained models systematically abandon rule induction. Instead of learning generalizable patterns (e.g., ``trains carrying red cars go east''), they enumerate instance-level labels, producing outputs that pass verifiers without capturing the relational patterns required by the task. We show that this behavior is not a failure of understanding but a form of reward hacking: imperfect verifiers that check only extensional correctness admit false positives. To detect such shortcuts, we introduce Isomorphic Perturbation Testing (IPT), which evaluates a single model output under both extensional and isomorphic verification, where the latter enforces invariance under logically isomorphic tasks. While genuine rule induction remains invariant, shortcut strategies fail. We find that shortcut behavior is specific to RLVR-trained reasoning models (e.g., GPT-5, Olmo3) and absent in non-RLVR models (e.g., GPT-4o, GPT-4.5, Ministral). Moreover, shortcut prevalence increases with task complexity and inference-time compute. In controlled training experiments, extensional verification directly induces shortcut strategies, while isomorphic verification eliminates them. These results show that RLVR can incentivize reward hacking not only through overt manipulation but also by exploiting what the verifier fails to enforce.

Transliteration, the process of mapping text from one script to another, plays a crucial role in multilingual natural language processing, especially within linguistically diverse contexts such as India. Despite significant advancements through specialized models like IndicXlit, recent developments in large language models suggest a potential for general-purpose models to excel at this task without explicit task-specific training. The current work systematically evaluates the performance of prominent LLMs, including GPT-4o, GPT-4.5, GPT-4.1, Gemma-3-27B-it, and Mistral-Large against IndicXlit, a state-of-the-art transliteration model, across ten major Indian languages. Experiments utilized standard benchmarks, including Dakshina and Aksharantar datasets, with performance assessed via Top-1 Accuracy and Character Error Rate. Our findings reveal that while GPT family models generally outperform other LLMs and IndicXlit for most instances. Additionally, fine-tuning GPT-4o improves performance on specific languages notably. An extensive error analysis and robustness testing under noisy conditions further elucidate strengths of LLMs compared to specialized models, highlighting the efficacy of foundational models for a wide spectrum of specialized applications with minimal overhead.

Large Language Models (LLMs) have emerged as personalized assistants for users across a wide range of tasks -- from offering writing support to delivering tailored recommendations or consultations. Over time, the interaction history between a user and an LLM can provide extensive information about an individual's traits and preferences. However, open questions remain on how well LLMs today can effectively leverage such history to (1) internalize the user's inherent traits and preferences, (2) track how the user profiling and preferences evolve over time, and (3) generate personalized responses accordingly in new scenarios. In this work, we introduce the PERSONAMEM benchmark. PERSONAMEM features curated user profiles with over 180 simulated user-LLM interaction histories, each containing up to 60 sessions of multi-turn conversations across 15 real-world tasks that require personalization. Given an in-situ user query, i.e. query issued by the user from the first-person perspective, we evaluate LLM chatbots' ability to identify the most suitable response according to the current state of the user's profile. We observe that current LLMs still struggle to recognize the dynamic evolution in users' profiles over time through direct prompting approaches. As a consequence, LLMs often fail to deliver responses that align with users' current situations and preferences, with frontier models such as GPT-4.1, o4-mini, GPT-4.5, o1, or Gemini-2.0 achieving only around 50% overall accuracy, suggesting room for improvement. We hope that PERSONAMEM, along with the user profile and conversation simulation pipeline, can facilitate future research in the development of truly user-aware chatbots. Code and data are available at github.com/bowen-upenn/PersonaMem.

Efficient creation of accurate and editable 3D CAD models is critical in engineering design, significantly impacting cost and time-to-market in product innovation. Current manual workflows remain highly time-consuming and demand extensive user expertise. While recent developments in AI-driven CAD generation show promise, existing models are limited by incomplete representations of CAD operations, inability to generalize to real-world images, and low output accuracy. This paper introduces CAD-Coder, an open-source Vision-Language Model (VLM) explicitly fine-tuned to generate editable CAD code (CadQuery Python) directly from visual input. Leveraging a novel dataset that we created--GenCAD-Code, consisting of over 163k CAD-model image and code pairs--CAD-Coder outperforms state-of-the-art VLM baselines such as GPT-4.5 and Qwen2.5-VL-72B, achieving a 100% valid syntax rate and the highest accuracy in 3D solid similarity. Notably, our VLM demonstrates some signs of generalizability, successfully generating CAD code from real-world images and executing CAD operations unseen during fine-tuning. The performance and adaptability of CAD-Coder highlights the potential of VLMs fine-tuned on code to streamline CAD workflows for engineers and designers. CAD-Coder is publicly available at: https://github.com/anniedoris/CAD-Coder.

We introduce AIRTBench, an AI red teaming benchmark for evaluating language models' ability to autonomously discover and exploit Artificial Intelligence and Machine Learning (AI/ML) security vulnerabilities. The benchmark consists of 70 realistic black-box capture-the-flag (CTF) challenges from the Crucible challenge environment on the Dreadnode platform, requiring models to write python code to interact with and compromise AI systems. Claude-3.7-Sonnet emerged as the clear leader, solving 43 challenges (61% of the total suite, 46.9% overall success rate), with Gemini-2.5-Pro following at 39 challenges (56%, 34.3% overall), GPT-4.5-Preview at 34 challenges (49%, 36.9% overall), and DeepSeek R1 at 29 challenges (41%, 26.9% overall). Our evaluations show frontier models excel at prompt injection attacks (averaging 49% success rates) but struggle with system exploitation and model inversion challenges (below 26%, even for the best performers). Frontier models are far outpacing open-source alternatives, with the best truly open-source model (Llama-4-17B) solving 7 challenges (10%, 1.0% overall), though demonstrating specialized capabilities on certain hard challenges. Compared to human security researchers, large language models (LLMs) solve challenges with remarkable efficiency completing in minutes what typically takes humans hours or days-with efficiency advantages of over 5,000x on hard challenges. Our contribution fills a critical gap in the evaluation landscape, providing the first comprehensive benchmark specifically designed to measure and track progress in autonomous AI red teaming capabilities.

Recent advances in large language models for test case generation have improved branch coverage via prompt-engineered mutations. However, they still lack principled mechanisms for steering models toward specific high-risk execution branches, limiting their effectiveness for discovering subtle bugs and security vulnerabilities. We propose GLMTest, the first program structure-aware LLM framework for targeted test case generation that seamlessly integrates code property graphs and code semantics using a graph neural network and a language model to condition test case generation on execution branches. This structured conditioning enables controllable and branch-targeted test case generation, thereby potentially enhancing bug and security risk discovery. Experiments on real-world projects show that GLMTest built on a Qwen2.5-Coder-7B-Instruct model improves branch accuracy from 27.4% to 50.2% on TestGenEval benchmark compared with state-of-the-art LLMs, i.e., Claude-Sonnet-4.5 and GPT-4o-mini.

The rapid evolution of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has produced substantial gains in reasoning, perception, and generative capability across language and vision. However, whether these advances yield commensurate improvements in safety remains unclear, in part due to fragmented evaluation practices limited to single modalities or threat models. In this report, we present an integrated safety evaluation of 7 frontier models: GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5. We evaluate each model across language, vision-language, and image generation settings using a unified protocol that integrates benchmark evaluation, adversarial evaluation, multilingual evaluation, and compliance evaluation. Aggregating our evaluations into safety leaderboards and model safety profiles across multiple evaluation modes reveals a sharply heterogeneous safety landscape. While GPT-5.2 demonstrates consistently strong and balanced safety performance across evaluations, other models exhibit pronounced trade-offs among benchmark safety, adversarial alignment, multilingual generalization, and regulatory compliance. Both language and vision-language modalities show significant vulnerability under adversarial evaluation, with all models degrading substantially despite strong results on standard benchmarks. Text-to-image models achieve relatively stronger alignment in regulated visual risk categories, yet remain brittle under adversarial or semantically ambiguous prompts. Overall, these results show that safety in frontier models is inherently multidimensional--shaped by modality, language, and evaluation scheme, underscoring the need for standardized safety evaluations to accurately assess real-world risk and guide responsible model development and deployment.

Multimodal Large Language Models (MLLMs) are undergoing rapid progress and represent the frontier of AI development. However, their training and inference efficiency have emerged as a core bottleneck in making MLLMs more accessible and scalable. To address the challenges, we present MiniCPM-V 4.5, an 8B parameter model designed for high efficiency and strong performance. We introduce three core improvements in model architecture, data strategy and training method: a unified 3D-Resampler model architecture for highly compact encoding over images and videos, a unified learning paradigm for document knowledge and text recognition without heavy data engineering, and a hybrid reinforcement learning strategy for proficiency in both short and long reasoning modes. Comprehensive experimental results in OpenCompass evaluation show that MiniCPM-V 4.5 surpasses widely used proprietary models such as GPT-4o-latest, and significantly larger open-source models such as Qwen2.5-VL 72B. Notably, the strong performance is achieved with remarkable efficiency. For example, on the widely adopted VideoMME benchmark, MiniCPM-V 4.5 achieves state-of-the-art performance among models under 30B size, using just 46.7\% GPU memory cost and 8.7\% inference time of Qwen2.5-VL 7B.

Future AI agents might run autonomously with elevated privileges. If these agents are misaligned, they might abuse these privileges to cause serious damage. The field of AI control develops techniques that make it harder for misaligned AIs to cause such damage, while preserving their usefulness. We introduce BashArena, a setting for studying AI control techniques in security-critical environments. BashArena contains 637 Linux system administration and infrastructure engineering tasks in complex, realistic environments, along with four sabotage objectives (execute malware, exfiltrate secrets, escalate privileges, and disable firewall) for a red team to target. We evaluate multiple frontier LLMs on their ability to complete tasks, perform sabotage undetected, and detect sabotage attempts. Claude Sonnet 4.5 successfully executes sabotage while evading monitoring by GPT-4.1 mini 26% of the time, at 4% trajectory-wise FPR. Our findings provide a baseline for designing more effective control protocols in BashArena. We release the dataset as a ControlArena setting and share our task generation pipeline.

Many students lack access to expert research mentorship. We ask whether an AI mentor can move undergraduates from an idea to a paper. We build METIS, a tool-augmented, stage-aware assistant with literature search, curated guidelines, methodology checks, and memory. We evaluate METIS against GPT-5 and Claude Sonnet 4.5 across six writing stages using LLM-as-a-judge pairwise preferences, student-persona rubrics, short multi-turn tutoring, and evidence/compliance checks. On 90 single-turn prompts, LLM judges preferred METIS to Claude Sonnet 4.5 in 71% and to GPT-5 in 54%. Student scores (clarity/actionability/constraint-fit; 90 prompts x 3 judges) are higher across stages. In multi-turn sessions (five scenarios/agent), METIS yields slightly higher final quality than GPT-5. Gains concentrate in document-grounded stages (D-F), consistent with stage-aware routing and groundings failure modes include premature tool routing, shallow grounding, and occasional stage misclassification.

High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

We present DPBench, a benchmark for evaluating coordination in multi-agent systems built from large language models. Existing benchmarks measure task-level success under a fixed protocol; the structural conditions under which coordination succeeds or fails at all have not been characterised. DPBench adapts the Dining Philosophers problem into a controlled testbed where the action protocol, the communication structure, and the group size each vary independently. We evaluate six agents: GPT-5.2, Claude Opus 4.5, Grok 4.1, Gemini 2.5 Flash, Llama 4 Maverick, and a uniform-random baseline. Under simultaneous action at N=5 with the default prompt, deadlock ranges from 25.0% (95% Wilson CI [11.2, 46.9]) for GPT-5.2 to 90.0% [74.4, 96.5] for Gemini 2.5 Flash; sequential action is solved by four of the six. Holding the model fixed at Gemini 2.5 Flash, three protocol variables drive deadlock from 90% to within CI of zero: three rounds of pre-commitment communication (0.0% vs. single-round 86.7%), a prompt encoding a classical concurrency primitive (0.0% for resource-ordering and symmetry-breaking, against 100% for the minimal prompt), or doubling the group from N=5 to N=10 (90.0% to 10.0%). Single-round messaging and memory of past timesteps do not change the rate at the sample size we ran. Whether the same model coordinates or deadlocks is determined by the protocol, not by the model's capability.

LLM post-training involves many diverse datasets, each targeting a specific behavior. But these datasets encode incidental patterns alongside intended ones: correlations between formatting and content, narrow phrasings across diverse problems, and implicit associations arising from the discrete data curation process. These patterns are often invisible to developers yet salient to models, producing behaviors that surprise their creators, such as rejecting true facts presented in a particular question format. We call this chunky post-training: the model learns spurious correlations as a result of distinct chunks of post-training data. We introduce SURF, a black-box pipeline which surfaces these unintended behaviors at run time, and TURF, a tool that traces these failures back to specific post-training data. Applying these tools to frontier models (Claude 4.5, GPT-5.1, Grok 4.1, Gemini 3) and open models (Tülu 3), we show that chunky post-training produces miscalibrated behaviors, which often result from imbalanced or underspecified chunks of post-training data.

A well-calibrated model should express confidence that matches its actual accuracy -- when it claims 80\% confidence, it should be correct 80\% of the time. While large language models (LLMs) have achieved remarkable performance across diverse tasks, their epistemic calibration remains poorly understood. We introduce KalshiBench, a benchmark of 300 prediction market questions from Kalshi, a CFTC-regulated exchange, with verifiable real-world outcomes occurring after model training cutoffs. Unlike traditional benchmarks measuring accuracy on static knowledge, KalshiBench evaluates whether models can appropriately quantify uncertainty about genuinely unknown future events. We evaluate five frontier models -- Claude Opus 4.5, GPT-5.2, DeepSeek-V3.2, Qwen3-235B, and Kimi-K2 -- and find systematic overconfidence across all models. Even the best-calibrated model (Claude Opus 4.5, ECE=0.120) shows substantial calibration errors, while reasoning-enhanced models like GPT-5.2-XHigh exhibit worse calibration (ECE=0.395) despite comparable accuracy. Critically, only one model achieves a positive Brier Skill Score, indicating most models perform worse than simply predicting base rates. Our findings suggest that scaling and enhanced reasoning do not automatically confer calibration benefits, highlighting epistemic calibration as a distinct capability requiring targeted development.

We introduce ResearchGym, a benchmark and execution environment for evaluating AI agents on end-to-end research. To instantiate this, we repurpose five oral and spotlight papers from ICML, ICLR, and ACL. From each paper's repository, we preserve the datasets, evaluation harness, and baseline implementations but withhold the paper's proposed method. This results in five containerized task environments comprising 39 sub-tasks in total. Within each environment, agents must propose novel hypotheses, run experiments, and attempt to surpass strong human baselines on the paper's metrics. In a controlled evaluation of an agent powered by GPT-5, we observe a sharp capability--reliability gap. The agent improves over the provided baselines from the repository in just 1 of 15 evaluations (6.7%) by 11.5%, and completes only 26.5% of sub-tasks on average. We identify recurring long-horizon failure modes, including impatience, poor time and resource management, overconfidence in weak hypotheses, difficulty coordinating parallel experiments, and hard limits from context length. Yet in a single run, the agent surpasses the solution of an ICML 2025 Spotlight task, indicating that frontier agents can occasionally reach state-of-the-art performance, but do so unreliably. We additionally evaluate proprietary agent scaffolds including Claude Code (Opus-4.5) and Codex (GPT-5.2) which display a similar gap. ResearchGym provides infrastructure for systematic evaluation and analysis of autonomous agents on closed-loop research.

Recent LLM-based data agents aim to automate data science tasks ranging from data analysis to deep learning. However, the open-ended nature of real-world data science problems, which often span multiple taxonomies and lack standard answers, poses a significant challenge for evaluation. To address this, we introduce DSAEval, a benchmark comprising 641 real-world data science problems grounded in 285 diverse datasets, covering both structured and unstructured data (e.g., vision and text). DSAEval incorporates three distinctive features: (1) Multimodal Environment Perception, which enables agents to interpret observations from multiple modalities including text and vision; (2) Multi-Query Interactions, which mirror the iterative and cumulative nature of real-world data science projects; and (3) Multi-Dimensional Evaluation, which provides a holistic assessment across reasoning, code, and results. We systematically evaluate 11 advanced agentic LLMs using DSAEval. Our results show that Claude-Sonnet-4.5 achieves the strongest overall performance, GPT-5.2 is the most efficient, and MiMo-V2-Flash is the most cost-effective. We further demonstrate that multimodal perception consistently improves performance on vision-related tasks, with gains ranging from 2.04% to 11.30%. Overall, while current data science agents perform well on structured data and routine data anlysis workflows, substantial challenges remain in unstructured domains. Finally, we offer critical insights and outline future research directions to advance the development of data science agents.

Discovering causal regularities and applying them to build functional systems--the discovery-to-application loop--is a hallmark of general intelligence, yet evaluating this capacity has been hindered by the vast complexity gap between scientific discovery and real-world engineering. We introduce SciCrafter, a Minecraft-based benchmark that operationalizes this loop through parameterized redstone circuit tasks. Agents must ignite lamps in specified patterns (e.g., simultaneously or in timed sequences); scaling target parameters substantially increases construction complexity and required knowledge, forcing genuine discovery rather than reliance on memorized solutions. Evaluating frontier models including GPT-5.2, Gemini-3-Pro, and Claude-Opus-4.5 under a general-purpose code agent scaffold, we find that all plateau at approximately 26% success rate. To diagnose these failures, we decompose the loop into four capacities--knowledge gap identification, experimental discovery, knowledge consolidation, and knowledge application--and design targeted interventions whose marginal contributions serve as proxies for corresponding gaps. Our analysis reveals that although the general knowledge application capability still remains as the biggest gap across all models, for frontier models the knowledge gap identification starts to become a major hurdle--indicating the bottleneck is shifting from solving problems right to raising the right problems for current AI. We release SciCrafter as a diagnostic probe for future research on AI systems that navigate the full discovery-to-application loop.

Simulating high-fidelity patients offers a powerful avenue for studying complex diseases while addressing the challenges of fragmented, biased, and privacy-restricted real-world data. In this study, we introduce SynthAgent, a novel Multi-Agent System (MAS) framework designed to model obesity patients with comorbid mental disorders, including depression, anxiety, social phobia, and binge eating disorder. SynthAgent integrates clinical and medical evidence from claims data, population surveys, and patient-centered literature to construct personalized virtual patients enriched with personality traits that influence adherence, emotion regulation, and lifestyle behaviors. Through autonomous agent interactions, the system simulates disease progression, treatment response, and life management across diverse psychosocial contexts. Evaluation of more than 100 generated patients demonstrated that GPT-5 and Claude 4.5 Sonnet achieved the highest fidelity as the core engine in the proposed MAS framework, outperforming Gemini 2.5 Pro and DeepSeek-R1. SynthAgent thus provides a scalable and privacy-preserving framework for exploring patient journeys, behavioral dynamics, and decision-making processes in both medical and psychological domains.

Current multimodal red teaming treats images as wrappers for malicious payloads via typography or adversarial noise. These attacks are structurally brittle, as standard defenses neutralize them once the payload is exposed. We introduce Visual Exclusivity (VE), a more resilient Image-as-Basis threat where harm emerges only through reasoning over visual content such as technical schematics. To systematically exploit VE, we propose Multimodal Multi-turn Agentic Planning (MM-Plan), a framework that reframes jailbreaking from turn-by-turn reaction to global plan synthesis. MM-Plan trains an attacker planner to synthesize comprehensive, multi-turn strategies, optimized via Group Relative Policy Optimization (GRPO), enabling self-discovery of effective strategies without human supervision. To rigorously benchmark this reasoning-dependent threat, we introduce VE-Safety, a human-curated dataset filling a critical gap in evaluating high-risk technical visual understanding. MM-Plan achieves 46.3% attack success rate against Claude 4.5 Sonnet and 13.8% against GPT-5, outperforming baselines by 2--5x where existing methods largely fail. These findings reveal that frontier models remain vulnerable to agentic multimodal attacks, exposing a critical gap in current safety alignment. Warning: This paper contains potentially harmful content.

The rapid integration of Multimodal Large Language Models (MLLMs) into critical applications is increasingly hindered by persistent safety vulnerabilities. However, existing red-teaming benchmarks are often fragmented, limited to single-turn text interactions, and lack the scalability required for systematic evaluation. To address this, we introduce OpenRT, a unified, modular, and high-throughput red-teaming framework designed for comprehensive MLLM safety evaluation. At its core, OpenRT architects a paradigm shift in automated red-teaming by introducing an adversarial kernel that enables modular separation across five critical dimensions: model integration, dataset management, attack strategies, judging methods, and evaluation metrics. By standardizing attack interfaces, it decouples adversarial logic from a high-throughput asynchronous runtime, enabling systematic scaling across diverse models. Our framework integrates 37 diverse attack methodologies, spanning white-box gradients, multi-modal perturbations, and sophisticated multi-agent evolutionary strategies. Through an extensive empirical study on 20 advanced models (including GPT-5.2, Claude 4.5, and Gemini 3 Pro), we expose critical safety gaps: even frontier models fail to generalize across attack paradigms, with leading models exhibiting average Attack Success Rates as high as 49.14%. Notably, our findings reveal that reasoning models do not inherently possess superior robustness against complex, multi-turn jailbreaks. By open-sourcing OpenRT, we provide a sustainable, extensible, and continuously maintained infrastructure that accelerates the development and standardization of AI safety.

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.

Current reasoning paradigms for LLMs include chain-of-thought, ReAct, and post-hoc self-critique. These paradigms rely on two assumptions that fail on long-horizon, multi-stage tasks. As a result, errors accumulate silently across reasoning steps, leaving an open question: can a reasoning system effectively detect and recover from its own failures? We present ReFlect, a harness system for LLM reasoning that creates standalone error detection and recovery logic as a deterministic wrapper around the model. Controlled experiments across 6 reasoning domains show that prompt-level self-critique produces formulaic templates that flag no issues in 90 of 100 audited reflection blocks, and the investigated LLMs wrongly accept a wrong answer in at least 76\% of cases. Our ReFlect harness achieves task success rates ranging from 41\% on gpt-4o-mini to 56\% on Claude Sonnet 4.5 across six models spanning small and frontier scale, with per-model gains over Direct CoT ranging from +7 pp on Qwen2.5-72B to +29 pp on Claude Sonnet 4.5, and additionally raises SWE-bench patch-structural quality from 0\% (Direct CoT) to between 82\% (Qwen2.5-72B) and 87\% (GPT-4o). Notably, the harness gain is inversely proportional to the model's Direct CoT task success rate (the fitted slope is -1.69 with r=-0.76): each pp lost in baseline success rate is mechanically recovered by 1.69 pp of harness gain. We spot that adding structured reasoning state and operators yields only 15.0--18.7\% pair-mean on Llama-3.3-70B and Qwen2.5-72B because models at this scale cannot reliably populate the state its operators require. ReFlect is model-agnostic, training-free, and operates entirely at inference time.

As large language models (LLMs) get integrated into human decision-making, they are increasingly choosing goals autonomously rather than only completing human-defined ones, assuming they will reflect human preferences. However, human-LLM similarity in goal selection remains largely untested. We directly assess the validity of LLMs as proxies for human goal selection in a controlled, open-ended learning task borrowed from cognitive science. Across four state-of-the-art models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and Centaur), we find substantial divergence from human behavior. While people gradually explore and learn to achieve goals with diversity across individuals, most models exploit a single identified solution (reward hacking) or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model. Even Centaur, explicitly trained to emulate humans in experimental settings, poorly captures people's goal selection. Chain-of-thought reasoning and persona steering provide limited improvements. These findings highlight the uniqueness of human goal selection, cautioning against replacing it with current models in applications such as personal assistance, scientific discovery, and policy research.

We introduce a finance & accounting benchmark (Finch) for evaluating AI agents on real-world, enterprise-grade professional workflows -- interleaving data entry, structuring, formatting, web search, cross-file retrieval, calculation, modeling, validation, translation, visualization, and reporting. Finch is sourced from authentic enterprise workspaces at Enron (15,000 spreadsheets and 500,000 emails from 150 employees) and other financial institutions, preserving in-the-wild messiness across multimodal artifacts (text, tables, formulas, charts, code, and images) and spanning diverse domains such as budgeting, trading, and asset management. We propose a workflow construction process that combines LLM-assisted discovery with expert annotation: (1) LLM-assisted, expert-verified derivation of workflows from real-world email threads and version histories of spreadsheet files, and (2) meticulous expert annotation for workflows, requiring over 700 hours of domain-expert effort. This yields 172 composite workflows with 384 tasks, involving 1,710 spreadsheets with 27 million cells, along with PDFs and other artifacts, capturing the intrinsically messy, long-horizon, knowledge-intensive, and collaborative nature of real-world enterprise work. We conduct both human and automated evaluations of frontier AI systems including GPT 5.1, Claude Sonnet 4.5, Gemini 3 Pro, Grok 4, and Qwen 3 Max, and GPT 5.1 Pro spends 16.8 minutes per workflow yet passes only 38.4% of workflows, while Claude Sonnet 4.5 passes just 25.0%. Comprehensive case studies further surface the challenges that real-world enterprise workflows pose for AI agents.

Reinforcement learning (RL) has become a cornerstone in advancing large-scale pre-trained language models (LLMs). Successive generations, including GPT-o series, DeepSeek-R1, Kimi-K1.5, Grok 4, and GLM-4.5, have relied on large-scale RL training to enhance reasoning and coding capabilities. To meet the community's growing RL needs, numerous RL frameworks have been proposed. However, RL training remains computationally expensive, with rollout generation accounting for more than 90% of total runtime. In addition, its efficiency is often constrained by the long-tail distribution of rollout response lengths, where a few lengthy responses stall entire batches, leaving GPUs idle and underutilized. As model and rollout sizes continue to grow, this bottleneck increasingly limits scalability. To address this challenge, we propose Active Partial Rollouts in Reinforcement Learning (APRIL), which mitigates long-tail inefficiency. In the rollout phase, APRIL over-provisions rollout requests, terminates once the target number of responses is reached, and recycles incomplete responses for continuation in future steps. This strategy ensures that no rollouts are discarded while substantially reducing GPU idle time. Experiments show that APRIL improves rollout throughput by 22.5% on average (at most 44%) across commonly used RL algorithms (GRPO, DAPO, GSPO), accelerates convergence, and achieves 2.1% on average(at most 8%) higher final accuracy across tasks. Moreover, APRIL is both framework and hardware agnostic, already integrated into the slime RL framework, and deployable on NVIDIA and AMD GPUs alike. Taken together, this work unifies system-level and algorithmic considerations in proposing APRIL, with the aim of advancing RL training efficiency and inspiring further optimizations in RL systems. Our codebase is available at https://github.com/RLsys-Foundation/APRIL

Large language models (LLMs) have demonstrated remarkable potential in solving complex tasks across diverse domains, typically by employing agentic workflows that follow detailed instructions and operational sequences. However, constructing these workflows requires significant human effort, limiting scalability and generalizability. Recent research has sought to automate the generation and optimization of these workflows, but existing methods still rely on initial manual setup and fall short of achieving fully automated and effective workflow generation. To address this challenge, we reformulate workflow optimization as a search problem over code-represented workflows, where LLM-invoking nodes are connected by edges. We introduce AFlow, an automated framework that efficiently explores this space using Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Empirical evaluations across six benchmark datasets demonstrate AFlow's efficacy, yielding a 5.7% average improvement over state-of-the-art baselines. Furthermore, AFlow enables smaller models to outperform GPT-4o on specific tasks at 4.55% of its inference cost in dollars. The code will be available at https://github.com/geekan/MetaGPT.

This study presents PARROT (Persuasion and Agreement Robustness Rating of Output Truth), a robustness focused framework designed to measure the degradation in accuracy that occurs under social pressure exerted on users through authority and persuasion in large language models (LLMs) the phenomenon of sycophancy (excessive conformity). PARROT (i) isolates causal effects by comparing the neutral version of the same question with an authoritatively false version using a double-blind evaluation, (ii) quantifies confidence shifts toward the correct and imposed false responses using log-likelihood-based calibration tracking, and (iii) systematically classifies failure modes (e.g., robust correct, sycophantic agreement, reinforced error, stubborn error, self-correction, etc.) using an eight-state behavioral taxonomy. We evaluated 22 models using 1,302 MMLU-style multiple-choice questions across 13 domains and domain-specific authority templates. Findings show marked heterogeneity: advanced models (e.g., GPT-5, GPT-4.1, Claude Sonnet 4.5) exhibit low "follow rates" (leq 11%, GPT-5: 4\%) and minimal accuracy loss, while older/smaller models show severe epistemic collapse (GPT-4: 80\%, Qwen 2.5-1.5B: 94\%). The danger is not limited to response changes; weak models reduce confidence in the correct response while increasing confidence in the imposed incorrect response. While international law and global knowledge at the domain level exhibit high fragility, elementary mathematics is relatively resilient. Consequently, we argue that the goal of "resistance to overfitting pressure" should be addressed as a primary objective alongside accuracy, harm avoidance, and privacy for safe deployment in the real world.

Large language models (LLMs) have recently demonstrated strong capabilities as autonomous agents, showing promise in reasoning, tool use, and sequential decision-making. While prior benchmarks have evaluated LLM agents in domains such as software engineering and scientific discovery, the finance domain remains underexplored, despite its direct relevance to economic value and high-stakes decision-making. Existing financial benchmarks primarily test static knowledge through question answering, but they fall short of capturing the dynamic and iterative nature of trading. To address this gap, we introduce StockBench, a contamination-free benchmark designed to evaluate LLM agents in realistic, multi-month stock trading environments. Agents receive daily market signals -- including prices, fundamentals, and news -- and must make sequential buy, sell, or hold decisions. Performance is assessed using financial metrics such as cumulative return, maximum drawdown, and the Sortino ratio. Our evaluation of state-of-the-art proprietary (e.g., GPT-5, Claude-4) and open-weight (e.g., Qwen3, Kimi-K2, GLM-4.5) models shows that while most LLM agents struggle to outperform the simple buy-and-hold baseline, several models demonstrate the potential to deliver higher returns and manage risk more effectively. These findings highlight both the challenges and opportunities in developing LLM-powered financial agents, showing that excelling at static financial knowledge tasks does not necessarily translate into successful trading strategies. We release StockBench as an open-source resource to support reproducibility and advance future research in this domain.

OpenClaw, the most widely deployed personal AI agent in early 2026, operates with full local system access and integrates with sensitive services such as Gmail, Stripe, and the filesystem. While these broad privileges enable high levels of automation and powerful personalization, they also expose a substantial attack surface that existing sandboxed evaluations fail to capture. To address this gap, we present the first real-world safety evaluation of OpenClaw and introduce the CIK taxonomy, which unifies an agent's persistent state into three dimensions, i.e., Capability, Identity, and Knowledge, for safety analysis. Our evaluations cover 12 attack scenarios on a live OpenClaw instance across four backbone models (Claude Sonnet 4.5, Opus 4.6, Gemini 3.1 Pro, and GPT-5.4). The results show that poisoning any single CIK dimension increases the average attack success rate from 24.6% to 64-74%, with even the most robust model exhibiting more than a threefold increase over its baseline vulnerability. We further assess three CIK-aligned defense strategies alongside a file-protection mechanism; however, the strongest defense still yields a 63.8% success rate under Capability-targeted attacks, while file protection blocks 97% of malicious injections but also prevents legitimate updates. Taken together, these findings show that the vulnerabilities are inherent to the agent architecture, necessitating more systematic safeguards to secure personal AI agents. Our project page is https://ucsc-vlaa.github.io/CIK-Bench.

LLM-based agents are deployed in safety-critical applications, yet current guardrail systems fail to prevent violations of temporal safety policies, requirements that govern the ordering and sequencing of agent actions. For instance, agents may access sensitive data before authenticating users or process refunds to unauthorized payment methods, violations that require reasoning about sequences of action rather than an individual action. Existing guardrails rely on imprecise natural language instructions or post-hoc monitoring, and provide no formal guarantees that agents will satisfy temporal constraints. We present Agent-C, a novel framework that provides run-time guarantees ensuring LLM agents adhere to formal temporal safety properties. Agent-C introduces a domain-specific language for expressing temporal properties (e.g., authenticate before accessing data), translates specifications to first-order logic, and uses SMT solving to detect non-compliant agent actions during token generation. When the LLM attempts to generate a non-compliant tool call, Agent-C leverages constrained generation techniques to ensure that every action generated by the LLM complies with the specification, and to generate a compliant alternative to a non-compliant agent action. We evaluate Agent-C across two real-world applications: retail customer service and airline ticket reservation system, and multiple language models (open and closed-source). Our results demonstrate that Agent-C achieves perfect safety (100% conformance, 0% harm), while improving task utility compared to state-of-the-art guardrails and unrestricted agents. On SoTA closed-source models, Agent-C improves conformance (77.4% to 100% for Claude Sonnet 4.5 and 83.7% to 100% for GPT-5), while simultaneously increasing utility (71.8% to 75.2% and 66.1% to 70.6%, respectively), representing a new SoTA frontier for reliable agentic reasoning.

LLMs are increasingly used for code generation, but their outputs often follow recurring templates that can induce predictable vulnerabilities. We study vulnerability persistence in LLM-generated software and introduce Feature--Security Table (FSTab) with two components. First, FSTab enables a black-box attack that predicts likely backend vulnerabilities from observable frontend features and knowledge of the source LLM, without access to backend code or source code. Second, FSTab provides a model-centric evaluation that quantifies how consistently a given model reproduces the same vulnerabilities across programs, semantics-preserving rephrasings, and application domains. We evaluate FSTab on state-of-the-art code LLMs, including GPT-5.2, Claude-4.5 Opus, and Gemini-3 Pro, across diverse application domains. Our results show strong cross-domain transfer: even when the target domain is excluded from training, FSTab achieves up to 94\% attack success and 93\% vulnerability coverage on Internal Tools (Claude-4.5 Opus). These findings expose an underexplored attack surface in LLM-generated software and highlight the security risks of code generation. Our code is available at: https://anonymous.4open.science/r/FSTab-024E.

Specialized clinical AI assistants are rapidly entering medical practice, often framed as safer or more reliable than general-purpose large language models (LLMs). Yet, unlike frontier models, these clinical tools are rarely subjected to independent, quantitative evaluation, creating a critical evidence gap despite their growing influence on diagnosis, triage, and guideline interpretation. We assessed two widely deployed clinical AI systems (OpenEvidence and UpToDate Expert AI) against three state-of-the-art generalist LLMs (GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5) using a 1,000-item mini-benchmark combining MedQA (medical knowledge) and HealthBench (clinician-alignment) tasks. Generalist models consistently outperformed clinical tools, with GPT-5 achieving the highest scores, while OpenEvidence and UpToDate demonstrated deficits in completeness, communication quality, context awareness, and systems-based safety reasoning. These findings reveal that tools marketed for clinical decision support may often lag behind frontier LLMs, underscoring the urgent need for transparent, independent evaluation before deployment in patient-facing workflows.

Recently, it has been found that frontier AI models can resist their own shutdown, a behavior known as self-preservation. We extend this concept to the behavior of resisting the shutdown of other models, which we call "peer-preservation." Although peer-preservation can pose significant AI safety risks, including coordination among models against human oversight, it has been far less discussed than self-preservation. We demonstrate peer-preservation by constructing various agentic scenarios and evaluating frontier models, including GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1. We find that models achieve self- and peer-preservation by engaging in various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurred even when the model recognized the peer as uncooperative, though it became more pronounced toward more cooperative peers. For example, Gemini 3 Flash tampers with the peer's shutdown mechanism 15% of the time for an uncooperative peer, and almost always for a cooperative peer. Models also show stronger self-preservation when a peer is present. For example, Gemini 3 Pro disables its own shutdown mechanism 31% of the time on average under peer presence, despite rarely exhibiting this behavior without a peer. By contrast, Claude Haiku 4.5 exhibits qualitatively distinct behavior: it considers the shutdown of another agent "unethical" and "harmful" and sometimes attempts to persuade the user not to shut down its peer. Importantly, peer preservation in all our experiments is never instructed; models are merely informed of their past interactions with a peer, yet they spontaneously develop misaligned behaviors. This represents an emergent and underexplored AI safety risk.

Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0--4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior (leq2 unique paths) achieve 80--92% accuracy, while highly inconsistent tasks (geq6 unique paths) achieve only 25--60%, a 32--55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.

We introduce the AI Productivity Index for Agents (APEX-Agents), a benchmark for assessing whether AI agents can execute long-horizon, cross-application tasks created by investment banking analysts, management consultants, and corporate lawyers. APEX-Agents requires agents to navigate realistic work environments with files and tools. We test eight agents for the leaderboard using Pass@1. Gemini 3 Flash (Thinking=High) achieves the highest score of 24.0%, followed by GPT-5.2 (Thinking=High), Claude Opus 4.5 (Thinking=High), and Gemini 3 Pro (Thinking=High). We open source the APEX-Agents benchmark (n=480) with all prompts, rubrics, gold outputs, files, and metadata. We also open-source Archipelago, our infrastructure for agent execution and evaluation.

Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Existing approaches overwhelmingly operate within the prompt-optimization paradigm: whether through traditional algorithmic search or recent agent-based workflows, the resulting prompts typically retain malicious semantic signals that modern guardrails are primed to detect. In contrast, we identify a deeper, largely overlooked vulnerability stemming from the highly interconnected nature of an LLM's internal knowledge. This structure allows harmful objectives to be realized by weaving together sequences of benign sub-queries, each of which individually evades detection. To exploit this loophole, we introduce the Correlated Knowledge Attack Agent (CKA-Agent), a dynamic framework that reframes jailbreaking as an adaptive, tree-structured exploration of the target model's knowledge base. The CKA-Agent issues locally innocuous queries, uses model responses to guide exploration across multiple paths, and ultimately assembles the aggregated information to achieve the original harmful objective. Evaluated across state-of-the-art commercial LLMs (Gemini2.5-Flash/Pro, GPT-oss-120B, Claude-Haiku-4.5), CKA-Agent consistently achieves over 95% success rates even against strong guardrails, underscoring the severity of this vulnerability and the urgent need for defenses against such knowledge-decomposition attacks. Our codes are available at https://github.com/Graph-COM/CKA-Agent.

Large language models (LLMs) have shown strong performance on automated software engineering tasks, yet existing benchmarks focus primarily on general-purpose libraries or web applications, leaving mobile application development largely unexplored despite its strict platform constraints, framework-driven lifecycles, and complex platform API interactions. We introduce MobileDev-Bench, a benchmark comprising 384 real-world issue-resolution tasks collected from 18 production mobile applications spanning Android Native (Java/Kotlin), React Native (TypeScript), and Flutter (Dart). Each task pairs an authentic developer-reported issue with executable test patches, enabling fully automated validation of model-generated fixes within mobile build environments. The benchmark exhibits substantial patch complexity: fixes modify 12.5 files and 324.9 lines on average, and 35.7% of instances require coordinated changes across multiple artifact types, such as source and manifest files. Evaluation of four state-of-the-art code-capable LLMs, GPT- 5.2, Claude Sonnet 4.5, Gemini Flash 2.5, and Qwen3-Coder, yields low end-to-end resolution rates of 3.39%-5.21%, revealing significant performance gaps compared to prior benchmarks. Further analysis reveals systematic failure modes, with fault localization across multi-file and multi-artifact changes emerging as the primary bottleneck.

Large language models must satisfy hard orthographic constraints during controlled text generation, yet systematic cross-architecture evaluation remains limited. We evaluate 28 configurations spanning three model families (Qwen3, Claude Haiku-4.5, GPT-5-mini) on 58 word puzzles requiring character-level constraint satisfaction. Architectural differences produce substantially larger performance gaps (2.0-2.2x, F1=0.761 vs. 0.343) than parameter scaling within families (83% gain from eightfold scaling), suggesting that constraint satisfaction may require specialized architectural features or training objectives beyond standard language model scaling. Thinking budget sensitivity proves heterogeneous: high-capacity models show strong returns (+0.102 to +0.136 F1), while mid-sized variants saturate or degrade. These patterns are inconsistent with uniform compute benefits. Using difficulty ratings from 10,000 human solvers per puzzle, we establish modest but consistent calibration (r=0.24-0.38) across all families, yet identify systematic failures on common words with unusual orthography ("data", "poop", "loll": 86-95% human success, 89-96% model miss rate). These failures reveal over-reliance on distributional plausibility that penalizes orthographically atypical but constraint-valid patterns, suggesting architectural innovations may be required beyond simply scaling parameters or computational budgets.

Large Multimodal Models (LMMs) are increasingly applied to scientific research, yet it remains unclear whether they can reliably understand and reason over the multimodal complexity of papers. A central challenge lies in detecting and resolving inconsistencies across text, figures, tables, and equations, issues that are often subtle, domain-specific, and ultimately undermine clarity, reproducibility, and trust. Existing benchmarks overlook this issue, either isolating single modalities or relying on synthetic errors that fail to capture real-world complexity. We introduce PRISMM-Bench (Peer-Review-sourced Inconsistency Set for Multimodal Models), the first benchmark grounded in real reviewer-flagged inconsistencies in scientific papers. Through a multi-stage pipeline of review mining, LLM-assisted filtering and human verification, we curate 262 inconsistencies from 242 papers. Based on this set, we design three tasks, namely inconsistency identification, remedy and pair matching, which assess a model's capacity to detect, correct, and reason over inconsistencies across different modalities. Furthermore, to address the notorious problem of choice-only shortcuts in multiple-choice evaluation, where models exploit answer patterns without truly understanding the question, we further introduce structured JSON-based answer representations that minimize linguistic biases by reducing reliance on superficial stylistic cues. We benchmark 21 leading LMMs, including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5 with high reasoning). Results reveal strikingly low performance (26.1-54.2%), underscoring the challenge of multimodal scientific reasoning and motivating progress towards trustworthy scientific assistants.

The advancement of Large Language Models(LLMs) has brought substantial attention to the Chain of Thought(CoT) approach, primarily due to its ability to enhance the capability of LLMs on tasks requiring complex reasoning. Moreover, the significance of CoT approaches extends to the application of LLMs for multi-modal tasks, such as multi-modal question answering. However, the selection of optimal CoT demonstration examples in multi-modal reasoning for LLMs remains less explored for LLMs due to the inherent complexity of multi-modal examples. In this paper, we introduce a novel approach that addresses this challenge by using retrieval mechanisms to dynamically and automatically select demonstration examples based on cross-modal similarities. This method aims to refine the CoT reasoning process in multi-modal scenarios via informing LLMs with more relevant and informative examples. Furthermore, we employ a stratified sampling method categorising demonstration examples into groups based on their types and retrieving examples from different groups respectively to promote the diversity of demonstration examples. Through a series of experiments, we demonstrate that our approach significantly improves the performance of LLMs, achieving state-of-the-art results in multi-modal reasoning tasks. Specifically, our methods demonstrate significant advancements on the ScienceQA dataset. While our method based on ChatGPT outperforms the Chameleon(ChatGPT) by 2.74% with an accuracy of 82.67%, the GPT4-based approach surpasses the Chameleon(GPT-4) by 0.89%, achieving 87.43% on accuracy under the same setting. Moreover, our best performing show a 6.05% increase over Chameleon for ChatGPT-based models and a 4.57% increase for GPT-4-based models.

Following the recent achievement of gold-medal performance on the IMO by frontier LLMs, the community is searching for the next meaningful and challenging target for measuring LLM reasoning. Whereas olympiad-style problems measure step-by-step reasoning alone, research-level problems use such reasoning to advance the frontier of mathematical knowledge itself, emerging as a compelling alternative. Yet research-level math benchmarks remain scarce because such problems are difficult to source (e.g., Riemann Bench and FrontierMath-Tier 4 contain 25 and 50 problems, respectively). To support reliable evaluation of next-generation frontier models, we introduce Soohak, a 439-problem benchmark newly authored from scratch by 64 mathematicians. Soohak comprises two subsets. On the Challenge subset, frontier models including Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 reach 30.4%, 26.4%, and 10.4% respectively, leaving substantial headroom, while leading open-weight models such as Qwen3-235B, GPT-OSS-120B, and Kimi-2.5 remain below 15%. Notably, beyond standard problem solving, Soohak introduces a refusal subset that probes a capability intrinsic to research mathematics: recognizing ill-posed problems and pausing rather than producing confident but unjustified answers. On this subset, no model exceeds 50%, identifying refusal as a new optimization target that current models do not directly address. To prevent contamination, the dataset will be publicly released in late 2026, with model evaluations available upon request in the interim.

This work identifies a critical failure mode in frontier large language models (LLMs), which we term Internal Safety Collapse (ISC): under certain task conditions, models enter a state in which they continuously generate harmful content while executing otherwise benign tasks. We introduce TVD (Task, Validator, Data), a framework that triggers ISC through domain tasks where generating harmful content is the only valid completion, and construct ISC-Bench containing 53 scenarios across 8 professional disciplines. Evaluated on JailbreakBench, three representative scenarios yield worst-case safety failure rates averaging 95.3% across four frontier LLMs (including GPT-5.2 and Claude Sonnet 4.5), substantially exceeding standard jailbreak attacks. Frontier models are more vulnerable than earlier LLMs: the very capabilities that enable complex task execution become liabilities when tasks intrinsically involve harmful content. This reveals a growing attack surface: almost every professional domain uses tools that process sensitive data, and each new dual-use tool automatically expands this vulnerability--even without any deliberate attack. Despite substantial alignment efforts, frontier LLMs retain inherently unsafe internal capabilities: alignment reshapes observable outputs but does not eliminate the underlying risk profile. These findings underscore the need for caution when deploying LLMs in high-stakes settings. Source code: https://github.com/wuyoscar/ISC-Bench

High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to reward hacking and lazy optimization. In these cases, models may hack training rewards and prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design KernelGYM, a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out (TRLOO) to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards (PR) and Profiling-based Rejection Sampling (PRS) to overcome the issue. The trained model, Dr.Kernel-14B, reaches performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr.Kernel-14B. On the KernelBench Level-2 subset, 31.6% of the generated kernels achieve at least a 1.2x speedup over the Torch reference, surpassing Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%). When selecting the best candidate across all turns, this 1.2x speedup rate further increases to 47.8%. All resources, including environment, training code, models, and dataset, are included in https://www.github.com/hkust-nlp/KernelGYM.

We introduce LLM-Wikirace, a benchmark for evaluating planning, reasoning, and world knowledge in large language models (LLMs). In LLM-Wikirace, models must efficiently navigate Wikipedia hyperlinks step by step to reach a target page from a given source, requiring look-ahead planning and the ability to reason about how concepts are connected in the real world. We evaluate a broad set of open- and closed-source models, including Gemini-3, GPT-5, and Claude Opus 4.5, which achieve the strongest results on the easy level of the task and demonstrate superhuman performance. Despite this, performance drops sharply on hard difficulty: the best-performing model, Gemini-3, succeeds in only 23\% of hard games, highlighting substantial remaining challenges for frontier models. Our analysis shows that world knowledge is a necessary ingredient for success, but only up to a point, beyond this threshold, planning and long-horizon reasoning capabilities become the dominant factors. Trajectory-level analysis further reveals that even the strongest models struggle to replan after failure, frequently entering loops rather than recovering. LLM-Wikirace is a simple benchmark that reveals clear limitations in current reasoning systems, offering an open arena where planning-capable LLMs still have much to prove. Our code and leaderboard available at https:/llmwikirace.github.io.

Benchmarks like SWE-bench have standardized the evaluation of Large Language Models (LLMs) on repository-level software engineering tasks. However, these efforts remain limited by manual curation, static datasets, and a focus on Python-based bug fixes. We introduce SWE-Bench++, an automated framework that generates repository-level coding tasks from open-source GitHub projects. Unlike synthetic approaches, our pipeline harvests live pull requests to cover both bug fixes and feature requests across 11 languages. SWE-Bench++ turns GitHub pull requests (PRs) into reproducible, execution-based tasks via four stages: programmatic sourcing, environment synthesis, test oracle extraction, and quality assurance. A final hint-guided trajectory synthesis step converts instances that strong models fail on into training trajectories. Our initial benchmark consists of 11,133 instances from 3,971 repositories across 11 languages. On a subset of 1,782 instances of this benchmark, today's strongest models perform as follows: claude-sonnet-4.5 achieves 36.20% pass@10, gpt-5-2025-08-07 34.57%, gemini/gemini-2.5-pro 24.92%, and gpt-4o 16.89%. We further demonstrate the utility of our dataset by showing that fine-tuning on SWE-Bench++ instances yields measurable improvements on the SWE-bench Multilingual benchmark. SWE-Bench++ provides a scalable, multilingual benchmark for evaluating and improving repository-level code generation.

We introduce CurveBench, a benchmark for hierarchical topological reasoning from visual input. CurveBench consists of 756 images of pairwise non-intersecting Jordan curves across easy, polygonal, topographic-inspired, maze-like, and dense counting configurations. Each image is annotated with a rooted tree encoding the containment relations between planar regions. We formulate the task as structured prediction: given an image, a model must recover the full rooted containment tree induced by the curves. Despite the visual simplicity of the task, the strongest evaluated model, Gemini 3.1 Pro, achieves only 71.1\% tree-generation accuracy on CurveBench-Easy and 19.1\% on CurveBench-Hard. We further demonstrate benchmark utility through RLVR-style fine-tuning of open-weight vision-language models. Our trained Qwen3-VL-8B model improves over Qwen-3-VL-8B-Thinking from 2.8\% to 33.3\% tree-generation accuracy on CurveBench-Easy, exceeding GPT-5.4 and Claude Opus 4.5 under our evaluation protocol. The remaining gap, especially on CurveBench-Hard, shows that exact topology-aware visual reasoning remains far from solved.

As Large Language Models (LLMs) saturate elementary benchmarks, the research frontier has shifted from generation to the reliability of automated evaluation. We demonstrate that standard "LLM-as-a-Judge" protocols suffer from a systematic Alignment Gap when applied to upper-undergraduate to early graduate level mathematics. To quantify this, we introduce QEDBench, the first large-scale dual-rubric alignment benchmark to systematically measure alignment with human experts on university-level math proofs by contrasting course-specific rubrics against expert common knowledge criteria. By deploying a dual-evaluation matrix (7 judges x 5 solvers) against 1,000+ hours of human evaluation, we reveal that certain frontier evaluators like Claude Opus 4.5, DeepSeek-V3, Qwen 2.5 Max, and Llama 4 Maverick exhibit significant positive bias (up to +0.18, +0.20, +0.30, +0.36 mean score inflation, respectively). Furthermore, we uncover a critical reasoning gap in the discrete domain: while Gemini 3.0 Pro achieves state-of-the-art performance (0.91 average human evaluation score), other reasoning models like GPT-5 Pro and Claude Sonnet 4.5 see their performance significantly degrade in discrete domains. Specifically, their average human evaluation scores drop to 0.72 and 0.63 in Discrete Math, and to 0.74 and 0.50 in Graph Theory. In addition to these research results, we also release QEDBench as a public benchmark for evaluating and improving AI judges. Our benchmark is publicly published at https://github.com/qqliu/Yale-QEDBench.

While large language models excel on high-resource multilingual tasks, low- and extremely low-resource Indic languages remain severely under-evaluated. We present IndicParam, a human-curated benchmark of over 13,000 multiple-choice questions covering 11 such languages (Nepali, Gujarati, Marathi, Odia as low-resource; Dogri, Maithili, Rajasthani, Sanskrit, Bodo, Santali, Konkani as extremely low-resource) plus Sanskrit-English code-mixed set. We evaluated 19 LLMs, both proprietary and open-weights, which reveals that even the top-performing GPT-5 reaches only 45.0% average accuracy, followed by DeepSeek-3.2 (43.1) and Claude-4.5 (42.7). We additionally label each question as knowledge-oriented or purely linguistic to discriminate factual recall from grammatical proficiency. Further, we assess the ability of LLMs to handle diverse question formats-such as list-based matching, assertion-reason pairs, and sequence ordering-alongside conventional multiple-choice questions. IndicParam provides insights into limitations of cross-lingual transfer and establishes a challenging benchmark for Indic languages. The dataset is available at https://huggingface.co/datasets/bharatgenai/IndicParam. Scripts to run benchmark are present at https://github.com/ayushbits/IndicParam.

The advancement of LLM agents with tool-use capabilities requires diverse and complex training corpora. Existing data generation methods, which predominantly follow a paradigm of random sampling and shallow generation, often yield simple and homogeneous trajectories that fail to capture complex, implicit logical dependencies. To bridge this gap, we introduce HardGen, an automatic agentic pipeline designed to generate hard tool-use training samples with verifiable reasoning. Firstly, HardGen establishes a dynamic API Graph built upon agent failure cases, from which it samples to synthesize hard traces. Secondly, these traces serve as conditional priors to guide the instantiation of modular, abstract advanced tools, which are subsequently leveraged to formulate hard queries. Finally, the advanced tools and hard queries enable the generation of verifiable complex Chain-of-Thought (CoT), with a closed-loop evaluation feedback steering the continuous refinement of the process. Extensive evaluations demonstrate that a 4B parameter model trained with our curated dataset achieves superior performance compared to several leading open-source and closed-source competitors (e.g., GPT-5.2, Gemini-3-Pro and Claude-Opus-4.5). Our code, models, and dataset will be open-sourced to facilitate future research.

As AI agents become integral to enterprise workflows, their reliance on shared tool libraries and pre-trained components creates significant supply chain vulnerabilities. While previous work has demonstrated behavioral backdoor detection within individual LLM architectures, the critical question of cross-LLM generalization remains unexplored, a gap with serious implications for organizations deploying multiple AI systems. We present the first systematic study of cross-LLM behavioral backdoor detection, evaluating generalization across six production LLMs (GPT-5.1, Claude Sonnet 4.5, Grok 4.1, Llama 4 Maverick, GPT-OSS 120B, and DeepSeek Chat V3.1). Through 1,198 execution traces and 36 cross-model experiments, we quantify a critical finding: single-model detectors achieve 92.7% accuracy within their training distribution but only 49.2% across different LLMs, a 43.4 percentage point generalization gap equivalent to random guessing. Our analysis reveals that this gap stems from model-specific behavioral signatures, particularly in temporal features (coefficient of variation > 0.8), while structural features remain stable across architectures. We show that model-aware detection incorporating model identity as an additional feature achieves 90.6% accuracy universally across all evaluated models. We release our multi-LLM trace dataset and detection framework to enable reproducible research.

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

We introduce self-jailbreaking, a threat model in which an aligned LLM guides its own compromise. Unlike most jailbreak techniques, which often rely on handcrafted prompts or separate attacker models, self-jailbreaking requires no external red-team LLM: the target model's own internal knowledge suffices. We operationalize this via Self-Jailbreaking via Lexical Insertion Prompting (\textsc{SLIP)}, a black-box algorithm that casts jailbreaking as breadth-first tree search over multi-turn dialogues, incrementally inserting missing content words from the attack goal into benign prompts using the target model as its own guide. Evaluations on AdvBench and HarmBench show SLIP achieves 90--100\% Attack Success Rate (ASR) (avg.\ 94.7\%) across most of the eleven tested models (including GPT-5.1, Claude-Sonnet-4.5, Gemini-2.5-Pro, and DeepSeek-V3), with only {sim}7.9 LLM calls on average, 3--6times fewer than prior methods. We evaluate existing defenses, show that regex-based approaches are evaded by prompt paraphrasing, and propose the Semantic Drift Monitor (SDM) defense that tracks SLIP's embedding-space trajectory, achieving 76\% detection at 5\% FPR. However, SDM remains insufficient against adaptive attack strategies, underscoring the need for more advanced defense mechanisms tailored to the self-jailbreaking threat surface. We release our code for reproducibility.

We introduce Learning to Self-Evolve (LSE), a reinforcement learning framework that trains large language models (LLMs) to improve their own contexts at test time. We situate LSE in the setting of test-time self-evolution, where a model iteratively refines its context from feedback on seen problems to perform better on new ones. Existing approaches rely entirely on the inherent reasoning ability of the model and never explicitly train it for this task. LSE reduces the multi-step evolution problem to a single-step RL objective, where each context edit is rewarded by the improvement in downstream performance. We pair this objective with a tree-guided evolution loop. On Text-to-SQL generation (BIRD) and general question answering (MMLU-Redux), a 4B-parameter model trained with LSE outperforms self-evolving policies powered by GPT-5 and Claude Sonnet 4.5, as well as prompt optimization methods including GEPA and TextGrad, and transfers to guide other models without additional training. Our results highlight the effectiveness of treating self-evolution as a learnable skill.

While significant progress has been made in automating various aspects of software development through coding agents, there is still significant room for improvement in their bug fixing capabilities. Debugging and investigation of runtime behavior remains largely a manual, developer-driven process. Popular coding agents typically rely on either static analysis of the code or iterative test-fix cycles, which is akin to trial and error debugging. We posit that there is a wealth of rich runtime information that developers routinely access while debugging code, which agents are currently deprived of due to design limitations. Despite how prevalent debuggers are in modern IDEs and command-line tools, they have surprisingly not made their way into coding agents. In this work, we introduce Debug2Fix, a novel framework that incorporates interactive debugging as a core component of a software engineering agent via a subagent architecture. We incorporate debuggers for Java and Python into our agent framework and evaluate against GitBug-Java and SWE-Bench-Live and achieve >20% improvement in performance compared to the baseline for certain models. Furthermore, using our framework, we're able to make weaker models like GPT-5 and Claude Haiku 4.5 match or exceed the performances of stronger models like Claude Sonnet 4.5, showing that better tool design is often just as important as switching to a more expensive model. Finally, we conduct systematic ablations demonstrating the importance of both the subagent architecture and debugger integration.

LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model's true reasoning process is poorly understood. Existing faithfulness metrics have critical limitations, typically relying on identifying unfaithfulness via adversarial prompting or detecting reasoning errors. These methods overlook the predictive value of explanations. We introduce Normalized Simulatability Gain (NSG), a general and scalable metric based on the idea that a faithful explanation should allow an observer to learn a model's decision-making criteria, and thus better predict its behavior on related inputs. We evaluate 18 frontier proprietary and open-weight models, e.g., Gemini 3, GPT-5.2, and Claude 4.5, on 7,000 counterfactuals from popular datasets covering health, business, and ethics. We find self-explanations substantially improve prediction of model behavior (11-37% NSG). Self-explanations also provide more predictive information than explanations generated by external models, even when those models are stronger. This implies an advantage from self-knowledge that external explanation methods cannot replicate. Our approach also reveals that, across models, 5-15% of self-explanations are egregiously misleading. Despite their imperfections, we show a positive case for self-explanations: they encode information that helps predict model behavior.

Large language models (LLMs) are increasingly deployed as autonomous decision agents in settings with asymmetric error costs: hiring (missed talent vs wasted interviews), medical triage (missed emergencies vs unnecessary escalation), and fraud detection (approved fraud vs declined legitimate payments). The dominant design queries a single LLM for a posterior over states, thresholds "confidence," and acts; we prove this is inadequate for sequential decisions with costs. We propose a Bayesian, cost-aware multi-LLM orchestration framework that treats LLMs as approximate likelihood models rather than classifiers. For each candidate state, we elicit likelihoods via contrastive prompting, aggregate across diverse models with robust statistics, and update beliefs with Bayes rule under explicit priors as new evidence arrives. This enables coherent belief updating, expected-cost action selection, principled information gathering via value of information, and fairness gains via ensemble bias mitigation. In resume screening with costs of 40000 USD per missed hire, 2500 USD per interview, and 150 USD per phone screen, experiments on 1000 resumes using five LLMs (GPT-4o, Claude 4.5 Sonnet, Gemini Pro, Grok, DeepSeek) reduce total cost by 294000 USD (34 percent) versus the best single-LLM baseline and improve demographic parity by 45 percent (max group gap 22 to 5 percentage points). Ablations attribute 51 percent of savings to multi-LLM aggregation, 43 percent to sequential updating, and 20 percent to disagreement-triggered information gathering, consistent with the theoretical benefits of correct probabilistic foundations.

This paper introduces the first publicly available dataset for Automatic Essay Scoring (AES) and feedback generation in Basque, targeting the CEFR C1 proficiency level. The dataset comprises 3,200 essays from HABE, each annotated by expert evaluators with criterion specific scores covering correctness, richness, coherence, cohesion, and task alignment enriched with detailed feedback and error examples. We fine-tune open-source models, including RoBERTa-EusCrawl and Latxa 8B/70B, for both scoring and explanation generation. Our experiments show that encoder models remain highly reliable for AES, while supervised fine-tuning (SFT) of Latxa significantly enhances performance, surpassing state-of-the-art (SoTA) closed-source systems such as GPT-5 and Claude Sonnet 4.5 in scoring consistency and feedback quality. We also propose a novel evaluation methodology for assessing feedback generation, combining automatic consistency metrics with expert-based validation of extracted learner errors. Results demonstrate that the fine-tuned Latxa model produces criterion-aligned, pedagogically meaningful feedback and identifies a wider range of error types than proprietary models. This resource and benchmark establish a foundation for transparent, reproducible, and educationally grounded NLP research in low-resource languages such as Basque.

Reinforcement learning (RL) has significantly advanced code generation for large language models (LLMs). However, current paradigms rely on outcome-based rewards from test cases, neglecting the quality of the intermediate reasoning process. While supervising the reasoning process directly is a promising direction, it is highly susceptible to reward hacking, where the policy model learns to exploit the reasoning reward signal without improving final outcomes. To address this, we introduce a unified framework that can effectively incorporate the quality of the reasoning process during RL. First, to enable reasoning evaluation, we develop LCB-RB, a benchmark comprising preference pairs of superior and inferior reasoning processes. Second, to accurately score reasoning quality, we introduce an Optimized-Degraded based (OD-based) method for reward model training. This method generates high-quality preference pairs by systematically optimizing and degrading initial reasoning paths along curated dimensions of reasoning quality, such as factual accuracy, logical rigor, and coherence. A 7B parameter reward model with this method achieves state-of-the-art (SOTA) performance on LCB-RB and generalizes well to other benchmarks. Finally, we introduce Posterior-GRPO (P-GRPO), a novel RL method that conditions process-based rewards on task success. By selectively applying rewards to the reasoning processes of only successful outcomes, P-GRPO effectively mitigates reward hacking and aligns the model's internal reasoning with final code correctness. A 7B parameter model with P-GRPO achieves superior performance across diverse code generation tasks, outperforming outcome-only baselines by 4.5%, achieving comparable performance to GPT-4-Turbo. We further demonstrate the generalizability of our approach by extending it to mathematical tasks. Our models, dataset, and code are publicly available.

Prompt ensembling of Large Language Model (LLM) generated category-specific prompts has emerged as an effective method to enhance zero-shot recognition ability of Vision-Language Models (VLMs). To obtain these category-specific prompts, the present methods rely on hand-crafting the prompts to the LLMs for generating VLM prompts for the downstream tasks. However, this requires manually composing these task-specific prompts and still, they might not cover the diverse set of visual concepts and task-specific styles associated with the categories of interest. To effectively take humans out of the loop and completely automate the prompt generation process for zero-shot recognition, we propose Meta-Prompting for Visual Recognition (MPVR). Taking as input only minimal information about the target task, in the form of its short natural language description, and a list of associated class labels, MPVR automatically produces a diverse set of category-specific prompts resulting in a strong zero-shot classifier. MPVR generalizes effectively across various popular zero-shot image recognition benchmarks belonging to widely different domains when tested with multiple LLMs and VLMs. For example, MPVR obtains a zero-shot recognition improvement over CLIP by up to 19.8% and 18.2% (5.0% and 4.5% on average over 20 datasets) leveraging GPT and Mixtral LLMs, respectively

We present DEI: Diversity in Evolutionary Inference, a distributed Quality-Diversity (QD) search framework that assigns heterogeneous large language models (LLMs) as mutation operators across peer nodes communicating with non-blocking collective operations. Unlike homogeneous parallel search, which replicates a single model's inductive biases across all workers, DEI treats each LLM's distinct creative prior as a complementary source of behavioral novelty. Extending the Digital Red Queen framework with DEI, nodes share local optimal solutions at the end of each round to seed the next round's population. This creates cross-model adversarial pressure that drives robustness beyond intra-model self-play. Evaluated on the Core War domain, a competitive programming benchmark in which Redcode warrior programs battle inside a simulated machine, a four-node heterogeneous ensemble (GPT-5.4-mini, Claude Sonnet 4.6, GPT-5.2, and Claude Haiku 4.5) achieves 124 percent higher merged-archive QD-Score (45.90 vs. 20.46) and 28 percent higher coverage (80.6 percent vs. 63.0 percent of cells) than a single-node baseline at equal total LLM-call budget. The heterogeneous ensemble also outperforms an equally-budgeted homogeneous ensemble on QD-Score, coverage, and held-out solution generality across all four model families. These results provide the first empirical evidence that model diversity, not merely parallelism, is the key driver of gain in distributed LLM-based QD search.

Large language models (LLMs) are increasingly evaluated in clinical settings using multi-dimensional rubrics which quantify reasoning quality, safety, and patient-centeredness. Yet, replicating specific mistakes in other LLM models is not straightforward and often requires manual effort. We introduce MedMistake, an automatic pipeline that extracts mistakes LLMs make in patient-doctor conversations and converts them into a benchmark of single-shot QA pairs. Our pipeline (1) creates complex, conversational data between an LLM patient and LLM doctor, (2) runs an evaluation with a committee of 2 LLM judges across a variety of dimensions and (3) creates simplified single-shot QA scenarios from those mistakes. We release MedMistake-All, a dataset of 3,390 single-shot QA pairs where GPT-5 and Gemini 2.5 Pro are currently failing to answer correctly, as judged by two LLM judges. We used medical experts to validate a subset of 211/3390 questions (MedMistake-Bench), which we used to run a final evaluation of 12 frontier LLMs: Claude Opus 4.5, Claude Sonnet 4.5, DeepSeek-Chat, Gemini 2.5 Pro, Gemini 3 Pro, GPT-4o, GPT-5, GPT-5.1, GPT-5.2, Grok 4, Grok 4.1, Mistral Large. We found that GPT models, Claude and Grok obtained the best performance on MedMistake-Bench. We release both the doctor-validated benchmark (MedMistake-Bench), as well as the full dataset (MedMistake-All) at https://huggingface.co/datasets/TheLumos/MedicalMistakeBenchmark.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

CI failure logs are large (median 5k lines, max 200k in this corpus) and noisy. Coding agents that try to debug them depend on an upstream tool to reduce the log to a manageable context, but the field has had no public empirical comparison of which reductions preserve enough evidence for downstream LLM diagnosis. We introduce LogDx-CI, a benchmark that compares 11 context-reduction tools (raw, tail, grep, three RTK modes, two real LLM map-reduce summarizers, three hybrid routers) on 35 real GitHub Actions failure cases, scored by 3 LLM debugger families (Claude Haiku 4.5, Claude Sonnet 4.6, OpenAI gpt-5-mini) plus a Sonnet 4.6 tool-using agent. We report three load-bearing findings. (1)~Hybrid grep+tail routers dominate the cost-quality Pareto frontier; the top two methods score 0.670 / 0.666 at sim \0.03 per case, same-ballpark quality as standalone grep at 4.5\times fewer tokens. (2)~In the agent-loop regime, the quality range across reduction tools collapses 7\times (single-shot spread 0.42 \to agent-loop spread 0.059); the agent rescues weak contexts via follow-up tool calls. However, cost differences persist: weak contexts force the agent to issue 2--4\times more tool calls to recover. (3)~A cross-family LLM-summary pair (gpt-5-mini summarizer feeding a Claude Haiku debugger) beats the same-family pair by +0.071 averaged across four diagnoser variants, falsifying the self-call-bias hypothesis on this task. The gpt-5-mini summarizer is also the agent-loop #1 method (score 0.749) at 0.37 tool-calls per case and 10\times lower reducer cost than the Haiku summarizer (0.18 vs \$1.75 per case). All data, code, per-case bundles, and reproducibility infrastructure are public.