When History Lies: Evaluating and Improving Tool Use under Misleading Multi-Turn Histories

Tool-using language models often operate in persistent, multi-turn environments where previous interactions accumulate. The authors investigate whether these historical traces—even when they are no longer relevant or authoritative—can 'hijack' a model's decision-making process, leading it to reuse outdated entities or incorrect interface conventions despite the model possessing the correct policy for the current request.

To diagnose this, the authors developed ContextPollute-Bench, a benchmark that provides three synchronized views of a decision point: an Original trajectory, a Polluted trajectory (containing misleading historical traces), and an Oracle State (containing only reliable, decision-relevant information). They then proposed Oracle-Guided On-Policy Distillation (Oracle-OPD). This method uses a teacher model conditioned on the Oracle State to provide soft supervision to a student model that only observes the polluted history. By training on student-generated prefixes, the student learns to ignore misleading context and align its behavior with the authoritative policy.

The study reveals that pollution flips over 32% of correct decisions in the Qwen3-1.7B model, confirming that history reliability is a distinct bottleneck in tool use. Oracle-OPD significantly outperforms standard supervised fine-tuning (SFT) and other distillation baselines, achieving 87.0% Balanced Tool-Use Accuracy on the 1.7B model. The method scales effectively, with larger teacher models further boosting the performance of compact students, and the resulting policies demonstrate strong transferability to unseen functions and external benchmarks.

As agents are increasingly deployed in long-running, stateful interactions, the ability to distinguish between relevant historical context and 'noise' becomes critical. This work provides a scalable, robust solution for training agents that remain reliable even when their interaction history contains misleading or obsolete information.