For teams training models to reason, a simple dogma has long held sway: if a Chain-of-Thought (CoT) leads to the correct answer, it serves as a high-quality training signal. However, researchers from the University of Electronic Science and Technology of China and the Singapore University of Technology and Design (SUTD) have identified a critical flaw in this logic. Their paper, "Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces," proves that even "correct" chains can yield radically different results during fine-tuning. The problem lies in a phenomenon the authors call "post-conclusion continuation"—the reasoning tails that drag on by inertia long after the final answer has been logically established.
The Anatomy of Harmful Inertia
When a model continues to "reason" after its logic has reached saturation, it isn't just burning tokens. It is actively degrading the supervised fine-tuning (SFT) process. The research team, led by Chen He, Yuhao Wu, and Lei Wang, argues that these redundant fragments act as low-quality noise. Their analysis revealed a specific "gap between uncertainty and geometry": the model's predictions remain unstable while progress in hidden-state geometry toward the terminal goal virtually stalls. Put simply, the chain begins to wander in a void once the work is already done.
"We observe an improvement in SFT results after removing post-conclusion continuations identified by the editor. This directly indicates that such inertia is harmful to training."
To test this hypothesis, researchers applied a "delete-only" editor. This tool does not rewrite data—which might introduce unwanted variables—but instead surgically removes the suffixes following a justified answer while preserving the prefix and correctness. The result was predictable: removing these "garbage tails" increased training efficiency in subsequent stages. From a methodological standpoint, a long reasoning chain stops being useful the moment it enters a low-value phase. By training a model on such data, you are essentially forcing it to mimic unstable and unproductive thinking patterns.
Rethinking Data Quality Standards
This discovery challenges the industry standard of filtering SFT datasets based solely on the final answer. If a correct answer no longer guarantees a quality trajectory, developers need more sophisticated cleaning tools. As a solution, the team introduced HarmfulContinuationCut (HCC)—a lightweight proxy tool designed to identify the boundaries of utility. HCC locates the point where logic ends and "harmful inertia" begins, allowing for cleaner data selection than traditional methods or the rewriting of chains by external models.
For the industry, this signals a shift from chasing volume to prioritizing the hygiene of training data structure. The discrepancy between uncertainty and logical progress in excess tokens proves that by training on "extra" steps, you teach the model to value local chaos over logical results. Tech leads must face a pragmatic reality: your Long-CoT datasets are likely oversaturated with logical clutter. Filtering for the right answer is merely the baseline; the real work begins where you know when to stop and cut the fluff.