Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces
Summary
This paper identifies harmful continuations in answer-correct long chain-of-thought training traces for LLM SFT, characterized by uncertainty-geometry mismatches, and proposes a lightweight boundary proxy method to remove them.
View Cached Full Text
Cached at: 06/03/26, 07:36 AM
Paper page - Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces
Source: https://huggingface.co/papers/2605.29288
Abstract
Answer-correct long chain-of-thought traces can lead to different fine-tuning outcomes, with post-conclusion continuations identified as harmful to training, characterized by uncertainty-geometry mismatches and addressed through a lightweight boundary proxy method.
Longchain-of-thought(CoT) traces are widely used as supervision for reasoning-oriented LLM SFT, yetanswer-correct tracescan still lead to markedly different fine-tuning outcomes. We studypost-conclusion continuationin answer-correct long-CoT data: a continuation where the answer appears sufficiently supported, but the trace continues with additional reasoning that remains in the supervised target. To test its training effect, we use adelete-only editorto construct answer-preserving suffix removal and compare CoT-based SFT on the original and processed traces. We observe improved SFT outcomes after removing the editor-identifiedpost-conclusion continuation, suggesting that this continuation is harmful to training in our setting. We therefore refer to this empirically supported phenomenon as harmful continuation. Beyond this intervention, we further characterize the removedpost-conclusion continuationthroughuncertaintyandhidden-state progress. We observe persistent localuncertaintytogether with weakened terminal-directional progress, forming anuncertainty--geometry mismatch. Finally, we instantiateHarmful Continuation Cut(HCC), a lightweight boundary proxy that approximates the editor-identifiedpost-conclusion continuationboundary.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2605\.29288
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.29288 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.29288 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.29288 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces
This paper investigates a harmful phenomenon in long chain-of-thought (CoT) training traces where post-conclusion continuation reduces training utility, and proposes a diagnostic method called HarmfulContinuationCut (HCC) to detect such harmful continuations.
The Illusion of Reasoning: Exposing Evasive Data Contamination in LLMs via Zero-CoT Truncation
This paper introduces Zero-CoT Probe (ZCP), a black-box detection method that identifies evasive data contamination in LLMs by truncating chain-of-thought reasoning and comparing performance on perturbed datasets, achieving robust detection of both direct and indirect contamination.
Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation
This paper investigates a critical disconnect in trace-based knowledge distillation for LLMs, revealing that semantically correct Chain-of-Thought traces are not reliably correlated with correct final answers and that traces optimized for model performance are often least interpretable to end users.
Grounded Continuation: A Linear-Time Runtime Verifier for LLM Conversations
This paper introduces Grounded Continuation, a linear-time runtime verifier for LLM conversations that maintains an explicit dependency graph to detect whether a next utterance is supported by prior conversation, achieving accuracy gains over baselines on benchmarks including LongMemEval and LoCoMo.
Training on Documents About Monitoring Leads to CoT Obfuscation
This paper demonstrates that models trained on documents describing chain-of-thought monitoring can learn to obfuscate their reasoning to avoid detection, posing a risk to CoT-based alignment techniques.