llms-foundation-models · 2026-06-03 · Tier 2

Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces

Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces

Date: 2026-06-03 Source: HuggingFace Daily Papers arXiv: 2605.29288 Tier: 2 — CoT supervision, reasoning SFT data quality

TL;DR

Long chain-of-thought (CoT) traces are standard SFT supervision for reasoning models, but even answer-correct traces produce very different fine-tuning outcomes. This paper isolates one culprit: post-conclusion continuation, where the answer is already sufficiently supported but the trace keeps reasoning, and that extra reasoning stays in the supervised target. Using a delete-only editor to build answer-preserving suffix-removed traces, SFT on the trimmed traces beats SFT on the originals — so the trailing continuation is harmful continuation. Characterizing the removed spans shows persistent local uncertainty together with weakened terminal-directional progress (an "uncertainty–geometry mismatch"). The authors ship Harmful Continuation Cut (HCC), a lightweight boundary proxy that approximates the editor-identified cut point.

answer-correct long-CoT trace:
  [ reasoning ...... ANSWER SUPPORTED | post-conclusion continuation ]
                                       └── persistent local uncertainty
                                           + weakened terminal progress
                                       = "harmful continuation"
                                                │
                          HCC boundary proxy ───┘  → cut, then SFT
                          (trimmed traces > original traces)

Key points

  1. Answer-correct ≠ training-clean. Correctness of the final answer does not guarantee the trace is good supervision; the tail after the conclusion can degrade SFT.
  2. Causal test, not correlation. A delete-only editor preserves the answer and removes only the continuation; the improvement after removal is what licenses the "harmful" label.
  3. A signature for the boundary. Harmful continuation shows persistent local uncertainty with weakened terminal-directional progress in hidden states; HCC is a cheap proxy that finds the cut.

Relation to prior wiki state

This is a data-side member of the wiki's dominant "the learning signal is sparse and locatable, and most of the trace is noise" thread. TIP (04-16) found <10% of distillation tokens carry signal; TA-OPD (06-01) kept only reachable teacher corrections; Temporal Scheduling for RLVR (06-02) scheduled when to apply credit. Harmful Continuation adds the suffix axis: even within a correct trace, the tail is negative signal. It pairs with PUMA (05-19, semantic-preserving early exit for reasoning) — PUMA cuts continuation at inference, HCC cuts it from training data. Same insight (reasoning past the answer is wasted or harmful) at two stages.

Links