Rewind Before the Uncertainty Peak

TL;DR for operators

When a reasoning model makes a mistake partway through an answer, restarting from the beginning discards potentially correct work, while revising only the ending may preserve the mistaken premise. The operational challenge is choosing where to restart—not simply finding where the model appears most uncertain.

REWIND, a training-free inference controller, traces backward from the point of highest token uncertainty to an earlier sharp rise in uncertainty, aligns the restart with a preceding text boundary, and regenerates the remaining response.¹ On Qwen2.5-Math-7B/MATH500, this combined approach achieved 77.60% accuracy, compared with 74.20% for direct revision at the uncertainty peak, 75.00% for backward tracing without boundary alignment, and 72.60% for random-cut MCMC. The improvement demonstrates more productive intervention, not verified identification of the original reasoning error.

REWIND led the evaluated alternatives across all 15 model–benchmark configurations, but its gains come with additional inference work and occasional regressions from correct to incorrect answers. On Qwen2.5-Math-7B, it generated approximately 4.8% more tokens than MCMC, excluding uncertainty-measurement passes and other overhead. Operators should judge selective regeneration by validated repairs, regressions, unresolved errors, total computation, and measured latency—not assume that detecting uncertainty guarantees correction.

A reasoning error can survive a late correction

Consider a reasoning assistant that adopts an incorrect assumption halfway through an analytical calculation. Subsequent steps may follow that assumption consistently, producing a fluent but incorrect conclusion.

Rewriting the final paragraph is unlikely to fix the underlying calculation. Restarting from the beginning provides another opportunity to solve the problem, but discards earlier steps that may have been valid.

The alternative is to preserve a prefix and regenerate the remainder. Its effectiveness depends heavily on where the preserved prefix ends.

This is the intervention problem addressed by REWIND. The method extends Power Sampling, an inference-time approach that uses MCMC-style proposals to explore alternative reasoning trajectories. Where its random-cut baseline chooses revision sites without considering local uncertainty, REWIND uses changes in the model’s own token probabilities to guide an additional suffix-revision proposal.

The distinction matters because the point of maximum uncertainty need not be the point where changing the continuation will be most productive.

REWIND searches backward before changing the continuation

A language model assigns probabilities to possible next tokens. Shannon entropy measures how dispersed that probability distribution is: higher entropy means the model is less concentrated on a particular next-token prediction.

REWIND uses these probabilities without training a separate error detector or changing model parameters.

Its controller first locates the generated token with the highest entropy. It then searches backward within a bounded window for the largest positive change in entropy between adjacent positions. Rather than cutting exactly at that position, it looks for a preceding text delimiter and moves the revision point to immediately after that boundary.

The retained prefix remains intact. The model samples a replacement suffix, and an MH-style acceptance filter probabilistically decides whether to keep the proposal.

The default configuration uses a five-token gradient look-back window and a ten-token boundary search. It performs 16 refinement blocks, each containing ten random-cut MCMC proposals followed by one entropy-guided proposal.

The acceptance score uses powered model probabilities, with a target power of four in the main experiments, adjusted for the actual proposal probabilities. It is a model-based sampling criterion, not a verifier of factual or mathematical correctness.

Importantly, the implementation omits state-dependent revision-site probabilities required for an exact Metropolis–Hastings acceptance ratio. Its probabilistic filtering does not establish exact stationarity.

The earlier cut produces better answers

The most revealing evidence comes from the controlled intervention ablation on Qwen2.5-Math-7B/MATH500.

Source: Section 5.3, Table 3. Revision variants add a suffix proposal after the common random-MCMC steps; the MCMC baseline does not include that additional guided proposal.

The comparison between direct peak revision and the full controller isolates a meaningful design choice within the added-proposal variants. Tracing backward and aligning the cut produces a 3.40-percentage-point improvement over intervening directly at the uncertainty peak.

Boundary alignment also contributes. Removing it reduces accuracy from 77.60% to 75.00%, a 2.60-point difference under the tested configuration.

Additional signal comparisons support the usefulness of the tracing procedure: gradient-based intervention outperforms direct peak revision for all six uncertainty signals in the reported MATH500 comparison. Shannon entropy nevertheless produces the highest final accuracy among those tested.

These are component ablations, not evidence that the algorithm identifies the precise token where reasoning first became incorrect. They show that the selected intervention locations produce better final answers under a shared regeneration and acceptance procedure.

Accuracy gains extend beyond one mathematical benchmark

The main evaluation covers three model families and five benchmarks: GSM8K, MATH500, GPQA Diamond, AIME2025, and BBH formal fallacies. Results are averaged over three independent runs.

REWIND leads the evaluated alternatives in all 15 model–benchmark combinations, with improvements over random-cut MCMC ranging from 0.51 to 10.00 percentage points.

Selected comparisons illustrate the variation:

Source: Section 5.2, Table 2. Main benchmark comparisons.

The 10-point AIME2025 improvement is substantial in relative terms, but that benchmark contains only 30 questions. Three-run averages do not eliminate uncertainty arising from its small evaluation set.

Supplementary transfer evaluations provide some evidence beyond the main reasoning settings. On HumanEval, REWIND improves over MCMC from 56.1% to 58.5% for Qwen2.5-Math-7B and from 62.8% to 66.5% for Tülu-8B. Standard eight-shot chain-of-thought prompting also preserves gains over MCMC on the three reported mathematical benchmarks.

These are transfer and prompting-robustness checks, rather than proof of general effectiveness across arbitrary business workflows.

Entropy is an intervention signal, not an error label

REWIND’s uncertainty mechanism has a useful empirical foundation.

In a preliminary study of 10,000 generated responses to 500 MATH500 questions, 9,951 responses were nonempty. Among the 463 questions containing both correct and incorrect generations, incorrect responses had approximately 112% higher mean token entropy in the average within-question comparison.

Mean response entropy achieved an AUROC of 0.746 for distinguishing incorrect responses, versus 0.724 for the maximum entropy gradient and 0.677 for peak entropy.

These measurements answer a different question from the intervention experiment. Mean entropy is the strongest reported response-level diagnostic among these three statistics, while REWIND uses token-position entropy changes to select where a revision should begin.

The paper offers a conditional theoretical explanation: predictive entropy can be decomposed into uncertainty within latent semantic interpretations and disagreement between interpretations. If disagreement rises sharply enough relative to background variation, the largest local entropy gradient identifies that fragmentation event.

But the assumptions are not validated for actual first reasoning errors. The benchmarks lack token-level first-error annotations, so the study cannot establish that REWIND consistently cuts at the first mistaken step.

Repair quality and inference cost determine deployment value

For a team deploying a fixed reasoning model in mathematical analysis, structured calculations, or code generation, REWIND suggests a possible reliability improvement without fine-tuning.

Its benefit must be measured at the response level.

On Qwen2.5-Math-7B/MATH500, relative to MCMC, REWIND averages 46 incorrect-to-correct repairs and 21 correct-to-incorrect regressions per 500-question run. That produces 25 net additional correct answers. The distinction matters: a repair mechanism can improve aggregate accuracy while degrading individual answers that were already correct.

Unresolved failures remain significant. The paper’s manual analysis classifies 7.6% of all MATH500 problems as confident errors without a salient uncertainty signal and 14.8% as detected-but-unrecovered cases. Successful detection therefore does not imply successful correction.

Additional computation is another constraint. On Qwen2.5-Math-7B, reported generated-token consumption is 9.26 times single-pass temperature sampling for REWIND, compared with 8.84 times for MCMC—approximately 4.8% more generated tokens than MCMC.

Those figures exclude entropy-only forward passes and other overhead. They are not normalized FLOP measurements or production-latency benchmarks.

A production evaluation should therefore record final-answer accuracy, repaired and regressed responses, unresolved confident errors, failed regenerations, total forward-pass computation, and wall-clock latency. Access to full-vocabulary token distributions and economical suffix rescoring must also be confirmed in the serving infrastructure.

The decision is whether the additional validated corrections justify the complete cost of the repair loop under the application’s actual workload.

The productive restart matters more than the uncertainty peak

REWIND’s contribution is a better-supported method for choosing where to revise a reasoning trajectory. Its benchmark results and ablations indicate that uncertainty dynamics can guide more productive suffix regeneration than direct peak intervention or random cutting.

The remaining boundary is equally concrete: uncertainty does not establish correctness, a promising cut does not guarantee recovery, and higher benchmark accuracy does not establish cheaper inference.

For analytical AI systems, targeted revision deserves evaluation as a distinct inference-control option. Its value will depend on how reliably it replaces wrong continuations without damaging correct ones—and how much computation that improvement requires.

Cognaptus: Automate the Present, Incubate the Future.