Search papers, labs, and topics across Lattice.
The paper identifies "self-anchored drift" as a key reason why LLMs perform worse when information is revealed gradually across turns compared to a single prompt. To mitigate this, they introduce Canonical-Context On-Policy Distillation (CCOPD), which aligns a student model's multi-turn responses with a teacher model's full-context behavior using on-policy distillation. CCOPD improves performance on multi-turn tasks by 32% while preserving full-context performance, suggesting it strengthens grounding in user evidence.
LLMs struggle with multi-turn conversations not because they lack information, but because they get anchored on unsupported assumptions made early in the conversation.
Large language models (LLMs) often solve a task when all instructions are given in a single prompt, but fail when the same information is revealed gradually across turns. When a clean FULL prompt and a RAW-SHARDED conversation contain the same complete user evidence, the model should still arrive at the same answer. We argue that a key reason for this gap is self-anchored drift: responses produced under partial information introduce unsupported assumptions, and those assumptions later distort the final answer. To reduce this effect, we propose Canonical-Context On-Policy Distillation (CCOPD). During training, the same base model is used in two roles: a frozen teacher conditioned on the clean FULL prompt and a trainable student that receives the same evidence incrementally through a multi-turn conversation; CCOPD aligns the student's behavior on its own trajectories with the teacher's canonical full-context behavior. Trained only on math problem conversations, CCOPD yields a 32\% average relative improvement in RAW-SHARDED performance over the original base model across math and five zero-shot out-of-domain task families, while largely preserving full-context performance. Further analyses suggest that CCOPD strengthens grounding in user evidence and reduces sensitivity to contamination from earlier assistant turns.
