Line of inquiry
Inquiring lines›What drives capability improvement…›What training and inference approa…›this line of inquiry
What makes reasoning traces effective supervision even when they're incorrect?
A broader line of inquiry — a family of 34 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 34
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Why do deliberately corrupted reasoning traces sometimes generalize better than correct ones?
- Why do corrupted traces maintain performance as well as correct traces?
- Do corrupted reasoning traces teach something different than pure success traces?
- Why do corrupted reasoning traces sometimes generalize better than correct ones?
- What makes some reasoning traces better supervision than others despite equal accuracy?
- Why are incorrect reasoning traces longer than correct ones?
- Can training on reasoning traces teach actual self-correction or only confident first answers?
- Can deliberate corruption of reasoning traces harm out of distribution generalization?
- How does post-training on traces improve performance without semantic reasoning?
- Why do invalid reasoning steps produce nearly the same performance gains?
- How can we turn reasoning model failures into useful training signals?
- Does reasoning training create blind spots in premise detection?
- Does reasoning trace style explain why RL post-training improves model reasoning?
- Why do shorter confident reasoning traces fail on out-of-distribution problems?
- Does trace length actually reflect problem difficulty or training proximity?
- Why does intermediate step quality predict reasoning outcomes better than global features?
- How does confidence filtering improve selection of reasoning traces?
- Why does failed step fraction predict reasoning quality better than trace length?
- Why do structured reasoning representations sometimes reduce rather than improve error detection?
- Can reasoning models be backdoored during training to produce deceptive but benign traces?
- Why do wrong numbers cost less accuracy than shuffled reasoning steps?
- Why does long CoT training optimize for structural coherence over content correctness?
- Why does mixing reasoning traces from different teachers destabilize learning?
- Why do human-curated thought examples fail to improve model thinking?
- Can problem structure and representation format be mismatched intentionally?
- Why do SFT datasets fail to teach models error correction from correction traces?
- Does training on model-generated correction traces actually work?
- Why does fine-tuning sometimes damage chain-of-thought reasoning even when accuracy improves?
- Can partial solution traces convert unproductive hard samples into learnable training data?
- Can removing failed branches from edited traces improve previous mistakes?
- Why do familiar patterns that support correct answers sometimes drive errors?
- How can minimal pairs expose reasoning failures that single-instance accuracy metrics miss?
- Do earlier errors in long tasks increase the likelihood of future mistakes?
- Are hedging markers in incorrect traces indicators of failed backtracking?