arxiv:2611.02075

Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents

Published on Nov 1

· Submitted by

haomingchao on Nov 6

Stable AI

Upvote

Authors:

Lang Mo ,

Mingchao Hao ,

Abstract

DPDM improves agent decision accuracy by combining frozen language-model representations with a shared pretrained conditional predictor that learns from deployment evidence without parameter updates.

Generated by thinkingmachines/Inkling-Small

Agent decisions depend on empirical relationships that deployment observations reveal and task descriptions alone may leave unresolved. We introduce Decoupled Predictive Decision Making (DPDM), an architecture that formulates heterogeneous agent decisions as conditional prediction with a shared pretrained learner. Language models provide semantic representations of current and historical records, while a general-purpose conditional predictor predicts task-defined targets from the represented evidence. Pretraining supplies reusable inference capability; deployment evidence instantiates problem-specific predictions without updating either model's neural parameters. Tool judgments, candidate preferences, and reasoning assessments share an evidence-and-query interface, with predictions mapped to choices by a decision rule. This architecture establishes complementary sources of agent capability: semantic knowledge acquired through language pretraining and empirical inference acquired through structured-data pretraining. A single frozen LimiX-2 predictor serves all five task categories. With Qwen3.5-27B as the semantic encoder, DPDM raises mean accuracy across 50 benchmarks from 78.80% with few-shot prompting to 84.92% under their respective evaluated history budgets, and reaches 82.33% with identical histories and queries. On ToolPRMBench's BFCL and ToolSandbox tasks, absolute accuracy gains are 34.26% and 28.87%, respectively, at the methods' respective evaluated history budgets. Accuracy improves as history grows. With identical imbalanced histories, mean balanced accuracy improves by an absolute 14.51% across five reasoning-step assessment tasks and four encoders. DPDM makes learning from experience an independently improvable foundation-model capability, creating a shared route for advances in conditional prediction to improve decisions across agent tasks.

View arXiv page View PDF Add to collection

Community

mingchaocc

Paper author Paper submitter 1 day ago

We introduce Decoupled Predictive Decision Making (DPDM), an architecture that turns historical experience into predictions for agent decisions through a shared pretrained learner.

The key idea is to separate semantic representation from empirical inference. An LLM represents current and historical records, while a conditional predictor uses the represented evidence to predict task-defined targets. This brings together two complementary capabilities: semantic knowledge from language pretraining and empirical inference from structured-data pretraining. Tool judgments, candidate preferences, and reasoning assessments share the same evidence-and-query interface, with predictions converted into choices by a decision rule. We use a single frozen LimiX-2 predictor across all five evaluated task categories, without updating either model’s neural parameters at deployment.

Key results:

Across 50 benchmarks: With Qwen3.5-27B as the semantic encoder, mean accuracy rises from 78.80% with few-shot prompting to 84.92% with DPDM at the methods’ respective evaluated history budgets. With identical histories and queries, DPDM reaches 82.33%.
Tool-use assessment: On ToolPRMBench’s BFCL and ToolSandbox tasks, accuracy improves over few-shot prompting by 34.26 and 28.87 percentage points, respectively, at the methods’ respective evaluated history budgets. Accuracy also improves as history grows.
Imbalanced histories: With identical imbalanced histories, DPDM improves mean balanced accuracy over few-shot prompting by 14.51 percentage points across five reasoning-step assessment tasks and four encoders.

DPDM makes learning from experience an independently improvable foundation-model capability: advances in a shared conditional predictor can improve decisions across agent tasks. We welcome feedback and discussion!