When Agent Feedback Makes Small Models Repeat a Failed Action

The paper opens with a reported tool-calling finding: when a harness records a failed action and its failure message before asking again, small models may be more likely to repeat that action. Its own experiments test games, not production tool calls. For teams building agents with small language models, the result is a reason to test how retry transcripts affect behavior, not evidence that the same effect occurs with real tools. ¹

What the paper found about agents repeating failed actions

The experiments covered Sokoban, Connect Four, and Othello. A move could fail by being illegal or by being legal but losing; the outcomes were verified exactly. ¹ On fixed probe items built by perturbing a correct move, recording the failed move made it more likely to be repeated across eight instruction-tuned models ranging from 135 million to 1.7 billion parameters. The pooled corrective gain was −1.38 nats per token. In this setting, the negative value indicates that the record increased the failed move's likelihood rather than correcting away from it. ¹

That is evidence of a retry effect in these game experiments. It does not establish that an agent will repeat a failed database query, API call, or file write in the same way. ¹

Why a failure transcript might change the next move

In an exploratory analysis, the failed move's log-probability after the failure record was only weakly related to its probability beforehand: the reported slope was 0.18. Moves that had been unlikely became more likely, while moves the model had been close to choosing became less likely. Greedy decoding still repeated 38% of the latter moves. ¹

That describes a shift in candidate-move probabilities, not a settled mechanism. One engineering interpretation is that error context may reweight options rather than simply suppressing the failed move. That is a hypothesis to test, not a result the paper establishes.

For runtime builders, the practical implication is to measure whether a retry format helps in the system being built. The game results show that a failure record can have counterintuitive effects in these experiments; they do not establish how it will affect tool use. ¹ This is the kind of gap that behavioral diagnostics, not just aggregate scores are meant to expose, and a model that behaves well in chat may still behave differently once it has tools, as the piece Refusing in Chat Doesn't Mean Refusing with Tools argues.

Where the repeat effect breaks down in the game experiments

The repeat effect operated through the move's surface form for illegal moves, but not for legal moves that lost. It also did not appear in every model. ¹

For tool-using language models, that suggests testing distinct failure classes, such as rejected actions and actions that execute but return an unhelpful result. This is a test-design recommendation, not evidence that tool failures divide in the same way as illegal and losing game moves. The paper reports different behavior across its two game-failure types; teams should test whether an equivalent distinction matters for their tools. ¹

Because the effect did not appear in every model, run the same retry transcripts across several candidate models rather than assuming one result transfers. A model catalog behind a single API makes that kind of side-by-side comparison easier to set up.

Which retry designs deserve comparison in agent harness design

The paper compares several ways of presenting a failure. A description that omitted the failed move's string kept the model's preference, while a clean retry was the worst-performing harness in the comparison. A “do not repeat” instruction had no effect, and a decoder ban stopped exact repeats but not illegal play generally. ¹

These results do not establish a general fix for real tool use. Omitting the action string changed the response in the game experiments, but the paper does not show that it improves task completion with real tools. A clean retry may discard diagnostic context; whether that helps or hurts recovery is something to measure in the target system. The decoder-ban result likewise shows that preventing an exact repeat is not the same as preventing invalid actions in the game setting. ¹

For a related look at evaluating agents through interactive behavior rather than static scores, see CogArena Tests Multimodal Agents Through Interactive Behavioral Experiments.