Search papers, labs, and topics across Lattice.
This study evaluates how misleading multi-turn histories can negatively impact tool-calling agents, revealing that 32.1% of decisions can be corrupted by outdated yet plausible historical traces. To address this issue, the authors introduce a novel benchmark called bench, which allows for the assessment of decision-making failures and the implementation of interventions that preserve gold-standard actions. Their proposed method, termed ours, leverages an Oracle-conditioned teacher policy to improve tool-use accuracy in a student model, achieving a significant increase in performance metrics compared to existing methods.
Misleading historical data can corrupt over 30% of tool-calling decisions, but a new method can restore accuracy by effectively transferring reliable policies from teacher to student models.
Tool-calling agents infer task state from accumulated dialogue and tool traces. In persistent interactions, however, historical traces may remain structurally valid and semantically plausible after they cease to be authoritative for the current request. We show that such history can hijack a policy the model already possesses: on Qwen3-1.7B, pollution flips 32.1% of decisions that are correct under the original trajectory and frequently induces reuse of corrupted entities or interface conventions. We introduce bench, a paired benchmark with synchronized Original, Polluted, and Oracle State views that preserve the system policy, current tools, latest request, and gold next action. Eleven gold-preserving interventions isolate failures in decision state, entity binding, and interface execution across complete calls and non-call decisions. We further propose ours, which transfers an Oracle-conditioned teacher policy to a student observing only polluted history through soft supervision on student-generated prefixes. On Qwen3-1.7B, ours achieves 87.0% Balanced Tool-Use Accuracy, outperforming Gold-SFT (66.3%), Oracle sequence distillation (82.3%), and off-policy token distillation (85.0%). The method scales consistently: an 8B teacher raises the same compact 1.7B student to 91.9%, while an 8B student reaches 93.0%. The resulting policies further transfer to clean histories, unseen functions, independently regenerated evaluation contexts, external tool-use benchmarks, and noisy multi-hop question answering. These results establish history reliability as a distinct tool-use bottleneck and demonstrate reliable-state policy transfer as an effective and scalable solution.
