AI models often commit to a wrong answer partway through their thinking — checking earlier steps in the same run can be more accurate.

This explores two different ways to squeeze better answers out of a reasoning model: mining the intermediate points inside a single trace versus running many separate traces and aggregating their outputs. The two turn out to be less rival strategies than complementary ones — and the corpus suggests the interesting action is often *inside* the trace, not across a fleet of them.

The case for mining intermediate points is surprisingly strong. Segmenting a trace into subthoughts and prompting a fresh completion from each intermediate cut, then taking the mode, beats reading off the final answer by up to 13% — because the model tends to commit early, and the earlier points still hold alternative paths the ending has already discarded Can intermediate reasoning points yield better answers than final ones?. Not all intermediate points are equal, though: planning and backtracking sentences act as sparse 'thought anchors' that causally steer everything downstream, so mining is really about finding those pivots rather than sampling uniformly Which sentences actually steer a reasoning trace?. And confidence read at the step level catches reasoning breakdowns that a single global score smooths over, letting you stop a bad trace early Does step-level confidence outperform global averaging for trace filtering?.

Aggregating separate traces is the more familiar move — majority voting over N samples — but the corpus reframes what makes it work. It isn't quantity: step-level filtering hits the same accuracy as naive majority voting with far fewer traces Does step-level confidence outperform global averaging for trace filtering?. What a multi-trace aggregator actually needs is *complementarity*, not agreement. Set-level RL rewards a batch of traces for being collectively useful rather than individually correct, and finds that diverse mediocre traces beat redundant strong ones — because the aggregator needs raw material to arbitrate between, not the same confident answer echoed back Can diverse mediocre traces outperform redundant expert traces?.

Here's the thing you didn't know you wanted to know: both approaches are downstream of the same blind spot in how models are trained. Test-time compute has three primitives — sequential (one long trace), parallel (many independent traces), and aggregative (combining them) — but post-training only ever optimizes the sequential one. So models never actually learn to produce jointly informative traces or to aggregate them well; parallelism gets wasted and compute is misallocated at inference Why does post-training ignore parallel and aggregative reasoning?. Mining intermediate points is, in effect, a way to recover parallel/aggregative value from a single sequential trace the model was trained to produce.

Why does either beat just trusting the final answer? Because reasoning models fail structurally, not for lack of compute — they wander into dead ends and switch away from good paths prematurely Why do reasoning models abandon promising solution paths?, and correct-looking traces can even keep 'reasoning' past the point the answer was settled, which actively hurts Does every correct chain-of-thought trace improve fine-tuning?. Both mining and aggregating are bets that the good answer already existed somewhere in the model's exploration and just needs to be recovered before early commitment buries it. The deeper lesson from process-verification work is that reliability comes from checking intermediate states, not final outputs — intermediate verification lifted task success from 32% to 87% Where do reasoning agents actually fail during long traces? — which is the same wager both strategies are making, one within a trace and one across many.