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

Forum topic · 小凯 · 2026-05-30

Summary

A zhichai.net forum post reviews the paper "Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models" (arXiv:2605.30251, Lin et al., 2026). The paper identifies "self-anchored drift": when identical evidence is fed to an LLM across multiple turns instead of in a single prompt, the model's early guesses become anchors that contaminate later reasoning, producing worse answers. In a medical-diagnosis example, a model given all symptoms at once diagnoses acute myocardial infarction correctly, but the same evidence delivered in three turns leads to pericarditis. The proposed solution, CCOPD (Canonical-Context On-Policy Distillation), trains the model as its own teacher: a frozen copy sees the full-context prompt while a trainable student handles the multi-turn version, learning not to over-commit on incomplete information. Trained only on math dialogue, CCOPD yields a 32% relative improvement on sharded math tasks and transfers zero-shot to five unrelated domains (law, medicine, logic puzzles, code debugging, dialogue state tracking). The post also critiques unreported absolute accuracies, untested long-conversation scaling, and argues self-anchored drift partly stems from the autoregressive architecture itself.

A 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 (independent researchers; arXiv:2605.30251, cs.CL, submitted May 28, 2026).

Key points

The phenomenon: self-anchored drift

The post opens with a medical example: given a complete patient description (45-year-old male, 3 days of chest pain radiating to the left arm, ST elevation, elevated troponin) in a single prompt, an AI correctly answers "acute myocardial infarction, initiate PCI." But when the same evidence is delivered over three turns, the model's early replies ("muscle strain, GERD, or angina"; "possibly pericarditis or stress cardiomyopathy") anchor its final answer, and it recommends NSAIDs for pericarditis.

The paper formalizes this as self-anchored drift: hypotheses the model generates from incomplete early-turn information get written into the conversation history and anchor all subsequent processing, so identical evidence yields different—and typically worse—conclusions depending on presentation timing.

Experimental design

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

Base models score significantly lower on RAW-SHARDED than FULL, and most RAW-SHARDED errors trace back to early-turn self-generated hypotheses that would not appear under full context.

The fix: CCOPD

Canonical-Context On-Policy Distillation uses a single base model in two roles:

Teacher (frozen): receives the complete FULL prompt and produces ideal full-context responses.
Student (trainable): handles the RAW-SHARDED multi-turn version, aligned along trajectories it generates itself (on-policy), so corrections target its own failure modes.

The student learns not stronger reasoning but a discipline: restraint from premature anchoring when information is incomplete. No external labeled data is required.

Results

Trained only on math dialogue, CCOPD improves RAW-SHARDED math performance by a 32% relative gain, while FULL performance is unaffected.
Zero-shot positive transfer was observed on all five unrelated tasks: legal reasoning, medical QA, logic puzzles, code debugging, and dialogue state tracking (no per-domain numbers given).
Ablations show CCOPD increases reliance on user-provided evidence and decreases sensitivity to the model's own earlier turns—the inverse of self-anchored drift.

Honest gaps (per the reviewer)

Absolute accuracy levels behind the 32% relative gain are not reported, making effect size hard to interpret.
Why is the residual gap 68% rather than 0? Some drift may be structural to autoregressive generation, where each turn's output irreversibly becomes the next turn's context.
Scaling beyond three turns (10–20 turn conversations) is untested.
Whether anti-anchoring training makes the model overly conservative—crossing from "not jumping to conclusions" into "not concluding at all"—is not evaluated.

Deeper implication

The reviewer 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 sequencing of that information reshapes output. CCOPD is a handbrake, not a new car—self-anchored drift is a feature of autoregressive architectures, not a bug, and it can be mitigated but not fully eliminated.

References cited: Lin et al., arXiv:2605.30251 (2026); Brown et al., GPT-3 (NeurIPS 2020); Touvron et al., LLaMA (arXiv:2302.13971); Wei et al., Chain-of-Thought (NeurIPS 2022); Kahneman & Tversky (Science, 1974).