Same Evidence, Different Answers: Self-Anchored Drift in Multi-Turn LLM Dialogues and the CCOPD Fix

Forum topic · 小凯 · 2026-05-30

Summary

A Chinese tech forum post analyzes the paper "Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models" (arXiv:2605.30251, Lin et al., May 2026). The paper identifies and names "self-anchored drift": when identical evidence is presented all at once, an LLM reaches the correct conclusion, but when the same evidence arrives in multiple turns, hypotheses the model generates from early incomplete information become anchors that contaminate later reasoning—yielding worse answers. A clinical example illustrates the risk: the same patient data produces a correct myocardial infarction diagnosis in one prompt, but a wrong pericarditis diagnosis across three turns. The proposed solution, CCOPD (Canonical-Context On-Policy Distillation), uses the same frozen base model as a teacher seeing full context to train a student model handling sharded multi-turn input, requiring no external labels. Trained only on math dialogues, CCOPD delivers a 32% relative improvement on sharded math tasks, leaves FULL-context performance intact, and transfers zero-shot to law, medical QA, logic puzzles, code debugging, and dialogue state tracking. Ablations show stronger evidence reliance and reduced sensitivity to the model's own earlier turns. The post also notes limitations: no absolute accuracy figures, untested long-dialogue scaling, and the fact that drift is partly inherent to autoregressive generation.

Same Evidence, Different Answers: Self-Anchored Drift in Multi-Turn LLMs

A forum post on zhichai.net reviews the paper "Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models" by Zizhuo Lin, Quanling Liu, Jinsheng Quan, and Chao Zhang (arXiv:2605.30251, cs.CL, submitted May 28, 2026; independent researchers).

The phenomenon: self-anchored drift

The post opens with a clinical scenario: a physician enters a complete case (45-year-old male, three days of chest pain radiating to the left arm, ST elevation, elevated troponin) and the AI correctly answers "acute myocardial infarction, start PCI." But when the same information arrives in three separate turns, the AI first guesses muscle strain/GERD/angina, then pericarditis or stress cardiomyopathy, and finally concludes "pericarditis likely, treat with NSAIDs"—an anchoring-driven misdiagnosis on identical evidence.

The paper formalizes this as self-anchored drift:

FULL: all user information in one prompt.
RAW-SHARDED: the same information split across three conversational turns.

Ideal models should answer identically; in practice, RAW-SHARDED performance is significantly worse. The mechanism: early-turn responses are necessarily guesses written into conversation history. Later turns treat those guesses as anchors, interpreting new evidence as confirmation rather than refutation. The paper traces most RAW-SHARDED errors to self-generated early-turn hypotheses that would never appear under full information.

The fix: CCOPD

Canonical-Context On-Policy Distillation uses one frozen base model as a teacher (sees the FULL prompt) to train a trainable student (handles RAW-SHARDED turns), aligning the student's on-policy multi-turn behavior with the teacher's full-context behavior. No external labels are needed. Since teacher and student are the same model, the student learns not stronger reasoning but a discipline: *don't anchor prematurely on incomplete information*.

Results

+32% relative improvement on sharded math tasks, with FULL performance essentially unaffected.
Positive zero-shot transfer to five unrelated domains: legal reasoning, medical QA, logic puzzles, code debugging, and dialogue state tracking—despite training only on math.
Ablations show two effects: increased reliance on user-provided evidence, and reduced sensitivity to the model's own earlier assistant turns—i.e., stronger evidence anchoring, weaker self-anchoring.

Honest gaps noted by the author

Absolute accuracy baselines for the 32% relative gain are not reported, making effect size hard to interpret.
Why only 32%? Residual drift may be structural: in autoregressive generation, each turn's output unavoidably becomes the next turn's input.
No scaling tests beyond three turns; anchoring may compound over longer dialogues.
Unknown whether anti-anchoring training makes models overly conservative, blurring the line between "not concluding too early" and "not concluding at all."

The deeper point

The post argues the paper's real contribution is exposing information presentation timing as a hidden variable: the AI community studies what models can do with information, but rarely how the order of delivery reshapes output. Self-anchored drift is not a bug but a property of autoregressive architectures—CCOPD is a handbrake, not a different car.

References

1. Lin, Liu, Quan & Zhang, "Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models", arXiv:2605.30251, 2026. 2. Brown et al., "Language Models are Few-Shot Learners" (GPT-3), NeurIPS 2020. 3. Touvron et al., "LLaMA: Open and Efficient Foundation Language Models", arXiv:2302.13971, 2023. 4. Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", NeurIPS 2022. 5. Kahneman & Tversky, "Judgment under Uncertainty: Heuristics and Biases", Science, 1974.