Five mediocre but different answers can outperform five copies of the same right one — a decision-maker needs variety, not confirmation.

This explores why a batch of reasoning traces can be worth more than the sum of its individually-scored parts — and it turns out the corpus has a surprisingly direct answer: complementarity, not correctness. The clearest case is SPIRAL's 'set reinforcement learning,' which stops rewarding each trace for being right and instead rewards a sampled *set* for being collectively useful Can diverse mediocre traces outperform redundant expert traces?. The counterintuitive result: a spread of diverse, mediocre traces beats a stack of redundant expert ones, because whatever aggregates them (a voter, a judge, a downstream model) needs raw material to arbitrate between — not five copies of the same confident answer confirming each other. Collective usefulness is about coverage of the possibility space, so the aggregator has something to choose from.

That reframes what you even want out of any single trace. If sets win on diversity, then the value of a trace is partly conditional on the company it keeps. Confidence-aware filtering pushes the same idea from the other end: instead of averaging a trace's global confidence, you look at step-level confidence to catch the exact point where reasoning breaks — which lets you keep the useful parts and discard traces early, hitting majority-voting accuracy with far fewer samples Does step-level confidence outperform global averaging for trace filtering?. The lesson across both: what you're curating is a portfolio, and portfolio value comes from where traces *differ* and where they *pivot*, not from each one's average score.

And not all parts of a trace carry equal weight in that portfolio. 'Thought anchors' research finds that a sparse set of planning and backtracking sentences disproportionately steer everything downstream — these are the functional pivots Which sentences actually steer a reasoning trace?. So a set is collectively useful partly because of the variety it holds *at those decision points*: different anchors mean genuinely different branches explored. Meanwhile, the tail of a trace can actively hurt — reasoning that continues after the answer is already settled degrades fine-tuning even when it stays correct Does every correct chain-of-thought trace improve fine-tuning?. Individual correctness is clearly not the right unit.

Here's the part you might not expect to want to know: if traces work as computational scaffolding rather than literal reasoning, then the *content* of a trace matters less than its structural role in the set. Models trained on deliberately corrupted or irrelevant traces perform comparably to those trained on correct ones, sometimes generalizing better out-of-distribution Do reasoning traces need to be semantically correct?. Combined with the finding that RLVR makes traces locally coherent without making them globally valid Does RLVR actually improve mathematical reasoning or just coherence?, the picture is that a trace's per-step 'quality' is a weak proxy for what a set contributes. This is also why evaluation is moving toward scoring final solutions, not traces — trace-level scoring inflates results by rewarding stylistic mimicry that a diverse set would expose anyway Should reasoning benchmarks score final answers or reasoning traces?. The through-line: a set's worth lives in its diversity, its pivot points, and its arbitrability — quality is what each trace has alone; usefulness is what the set can do together.