Stale-but-valid tool history hijacks agents: polluted traces flip 32.1% of previously correct decisions

arXiv·medium signalXBlueskyLinkedInCopy link

Wu et al. (arXiv 2608.06057, submitted August 6) identify history reliability as a distinct tool-use failure mode — trace entries that stay structurally valid and semantically plausible after they stop being authoritative for the current request. On Qwen3-1.7B, polluted history flips 32.1% of decisions that were correct under the original trajectory, frequently causing reuse of stale entities or interface conventions. Their paired benchmark synchronizes Original, Polluted and Oracle State views across eleven gold-preserving interventions, and their soft-supervision transfer from an Oracle-conditioned teacher reaches 87.0% Balanced Tool-Use Accuracy versus 66.3% for Gold-SFT and 85.0% for off-policy token distillation; an 8B teacher lifts the same 1.7B student to 91.9%.