LLM agents repeat failed moves in games, study finds
Researchers examined whether LLM agents repeat failed moves in board games when the harness logs the failure. They tested eight instruction‑tuned models ranging from 135 M to 1.7 B parameters on Sokoban, Connect Four, and Othello. The study found that adding the failed move and its error message to the transcript increases the chance of the model repeating that move for illegal actions, with a…
Key points
- Recording failed moves increases the likelihood of LLM agents repeating illegal actions in games.
- Across eight models (135 M–1.7 B) the corrective gain was –1.38 nats per token.
- "Do not repeat" instruction had no effect; decoder ban blocked exact repeats but not illegal play.
Analysis of log‑probabilities showed that after the record the model’s likelihood of the failed move barely depended on its prior probability. Greedy decoding began to produce the failed move 38 % of the time, while moves the model was about to play became less likely. A “do not repeat” instruction had no impact, a clean retry was the worst harness, and a decoder ban stopped exact repeats but not illegal play.
The work also evaluated the OLMo 2 1B post‑training ladder and RAGEN’s default models. The inversion was present in the base model and largely removed after DPO training. All experiments ran on a free Colab GPU using float32.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- MIT historian says early AI engineers added randomness to mimic human creativity · 1 src
- Researchers introduce MemPlan for text-based planning with PDDL · 1 src
- Researchers present CiteAnchor framework for detecting hallucinated citations · 1 src
- Researchers find self‑rewarded moral argumentation makes language models decisive but pliable · 1 src
- AQL-Lite improves Arabic LLMs with curated data and partial-parameter training · 1 src
Comments
via GitHub Discussions