Hi, I've been using Qwopus3.6-27B with tool calling and found a persistent loop issue specifically with tool calls. The model repeatedly calls the same tool with the same arguments, even when the result is already provided. What happens:
LLM calls read_file("/path/to/file.py")
Result is returned
LLM calls read_file("/path/to/file.py") again with identical arguments
Result is returned again
Loop continues until max turns is reached What I tried:
System prompt with explicit "Do NOT call the same tool twice" instruction - no effect
Anti-repeat guard in code (skipping repeated calls) - model keeps trying anyway
Lowering temperature from 0.7 to 0.2 - helps slightly but doesn't fix
Action log injection so the model "remembers" what it did - still loops The fix: Adding --repeat-penalty 1.15 --penalty_last_n 256 to llama-server completely resolves the issue. The model stops repeating tool calls. Key insight: This is a model-level repetition tendency, not a prompt engineering problem. The model seems to have an inherent bias toward repeating tool calls in its training distribution. Other models (like Qwen2.5-Coder) don't exhibit this behavior with tool calling. Suggestion for future versions:
Apply stronger repetition penalties during SFT for tool calling
Consider training with examples where the model stops after getting tool results
The --repeat-penalty 1.15 --penalty_last_n 256 config works as a workaround but shouldn't be necessary The model is excellent for coding tasks - great code generation, good RAG performance. This is the only issue I've found. Happy to provide more details or test future versions!
