Stale Tool History Flips 32% of Correct Decisions

Wu and coauthors show stale tool history flips 32.1% of correct agent decisions on Qwen3-1.7B, turning a working trajectory into a wrong one.

A stale entry stays structurally valid and reads as plausible right up until it flips the model's next move. Anyone running long agent sessions with full history in context is already exposed, per Wu and coauthors.

Most of the flipped decisions reuse a stale entity or an outdated interface convention the agent saw earlier in the same trace, per the paper.

The team tested a fix: soft-supervision transfer from an Oracle-conditioned teacher model. That pushed Balanced Tool-Use Accuracy to 87.0%, against 66.3% for a standard Gold-SFT baseline on the same 1.7B model. Swapping in an 8B teacher lifted the same 1.7B student to 91.9%.

Longer context windows won't fix this. Stale tool history looks exactly as valid as fresh history until it flips a decision. The fix has to come from training or pruning stale history, not from giving models more context to hold onto old truths. Watch whether tool-use benchmarks start reporting a Gold-SFT baseline next to the headline number, the way this paper does.