ReSolve: Reusing Candidate Reasoning through Selective Generative Moderation
The paper tackles the waste of intermediate reasoning in multi-sample LLM math inference, where self-consistency discards unfinished arguments and only votes on final answers. ReSolve proposes a training-free selective moderation loop: after sampling four candidates, a controller invokes the same 4B model to read existing derivations when answers disagree or are missing, then generates a repaired or completed solution within a bounded budget. The idea matters because it reframes already-generated reasoning as reusable evidence, potentially replacing independent resampling with cheaper evidence-conditioned refinement.
ReSolve is a carefully scoped inference procedure whose strongest result is a clean accuracy gain over voting on the exact same four candidates, with no observed regressions. However, the paper is commendably honest that it does not establish superiority over simply sampling eight independent solutions, and its specialized routing instructions fail to beat a generic uniform verifier. The evaluation is limited to a single 4B model on 130 repeated competition-math problems, so the broader claim that candidate reasoning is a reusable resource remains suggestive rather than proven.
The controlled ablations are the most convincing part of the paper. Removing visible derivations while keeping answer keys, counts, and output caps causes accuracy to drop from 100 to 93 correct on F2, demonstrating that the visible reasoning is doing real work. The selective-invocation control is also solid: always-on Uniform moderation attains the same 97 correct answers as selective moderation while burning roughly twice the moderation tokens, supporting the gate as a token-saving device without cost to accuracy.
The headline comparison against eight-sample self-consistency is statistically inconclusive. In F2 the Holm-corrected $p$-value is 0.14, and in F3 it is 1.0 after seven wins and four losses, so the paper cannot claim a reliable accuracy advantage over additional sampling.
The specialized route instructions do not outperform a single generic Uniform instruction, yet the paper's framing centers on this routing design. The two pools are resamplings of the same 130 problems, testing sampling variance but not generalization. Finally, the 46-47% token savings are undercut by latency: triggered problems add a median 108 seconds, and the paper explicitly warns that these ratios are not a controlled estimate of production speedup.
The evidence firmly supports the weaker claim that reading existing derivations improves over discarding them in favor of a simple vote: same-pool SC@4 comparisons are significant ($p=0.0039$ and $p=0.0156$) and the answer-only ablation isolates the contribution of visible reasoning. However, comparisons to related methods are limited; there are no direct experiments against Tree of Thoughts, LLM-Blender, or process-verification baselines, and the Uniform control shows that much of the gain comes from candidate-conditioned generation rather than from the proposed state-specific instructions. The use of Hybrid scoring with a rule grader and CompassVerifier-3B sensitivity analysis is responsible, and the McNemar paired testing is appropriate.
The appendices provide unusually detailed protocol documentation—answer extraction regexes, rendering limits (36,000-character candidate cap), acceptance rules, and token accounting—but the paper does not mention releasing code, prompts, or candidate logs. All experiments use the public Qwen3-4B-Thinking-2507 model, yet reproduction would require replicating exact vLLM serving settings, prompt renderers, and output-cap computations that differed between F2 and F3 (e.g., cache utilization 0.88 versus 0.42). The F3 batched implementation was validated for equivalent controller behavior, but since artifacts are not shared, independent reproduction would hinge on painstakingly recreating these inference engineering details.
Sampling multiple solutions spends computation on intermediate deductions and unfinished arguments as well as final answers. We introduce ReSolve, a training-free inference procedure that reuses this candidate reasoning through selective generative moderation. An answer-distribution controller invokes a model to examine existing derivations when candidates disagree or lack a parseable answer, then incorporates the generated solution into a bounded loop. Under Hybrid scoring on 130 competition-mathematics problems evaluated with two independently sampled candidate pools, ReSolve obtains 100 and 99 correct answers, compared with 91 and 92 for voting over the same four candidates, with no correct-to-incorrect changes relative to that vote in either pool. Eight-sample self-consistency obtains 94 and 96 correct answers while consuming substantially more tokens; ReSolve uses 46.3% and 47.2% fewer tokens in the two evaluations. A controlled ablation removes visible derivations while retaining answer keys, vote counts, and the per-state output-cap rule, reducing accuracy from 100 to 93 correct despite increasing computation. Selective and always-on Uniform moderation both solve 97 problems, while selectivity reduces moderation tokens by approximately 54% and total pipeline tokens by 6.2%. These results support candidate reasoning as reusable inference computation. They do not establish an accuracy advantage over additional sampling or a distinct benefit from specialized route instructions.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.