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

Researchers identify harmful continuation in long chain-of-thought training data where LLMs continue reasoning after the answer is sufficiently supported, degrading fine-tuning performance. Using a delete-only editor, they remove post-conclusion continuations and demonstrate improved SFT outcomes, introducing Harmful Continuation Cut (HCC) as a lightweight solution to detect and eliminate this problematic pattern.

This research addresses a critical inefficiency in how large language models are trained on reasoning tasks. Chain-of-thought supervision has become standard practice for teaching LLMs to reason step-by-step, yet the study reveals that answer-correct traces can contain superfluous reasoning that actively harms training outcomes. The phenomenon, termed harmful continuation, occurs when the model's reasoning extends beyond the point where the answer is sufficiently justified, creating confusion in the learning signal.

The problem emerges from a fundamental mismatch between what supervisory data communicates and what the model should learn. When human annotators or synthesis systems produce long reasoning traces, they often continue elaborating even after reaching correct answers. The research demonstrates this isn't merely redundant—it introduces measurable performance degradation. By analyzing hidden states and uncertainty metrics, the researchers discovered that post-conclusion continuations exhibit persistent local uncertainty alongside weakened directional progress toward conclusions, suggesting the model receives contradictory training signals.

This finding has significant implications for the broader AI training community. As reasoning-focused LLMs become increasingly important for coding, mathematics, and scientific applications, the quality of training supervision directly impacts performance. Organizations building or fine-tuning reasoning models should evaluate whether their training data contains similar harmful patterns. The lightweight Harmful Continuation Cut proxy offers a practical tool for filtering problematic traces without computational overhead.

Looking forward, this research opens questions about other subtle pathologies in supervised fine-tuning data. Similar mismatches between answer correctness and training value likely exist across other domains, suggesting that trace-level auditing could become standard practice in responsible AI development.

• →Post-conclusion continuation in chain-of-thought traces harms model fine-tuning despite answers being correct
• →Researchers identified uncertainty-geometry mismatch as the underlying cause of harmful continuation effects
• →Removing editor-identified post-conclusion continuations demonstrably improves SFT training outcomes
• →Harmful Continuation Cut provides a lightweight, scalable boundary detection proxy for production use
• →The finding suggests training data quality requires trace-level auditing beyond answer correctness verification