acceptodds
Under review as a conference paper at ICLR 2027

ReSolve: Reusing Candidate Reasoning through Selective Generative Moderation

12 views

Abstract

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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Reasoning, Test-Time Computation, Generative Verification, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DexrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces ReSolve, an inference method that uses the visible candidate reasoning to generate a repaired answer for problems whose multi-sample answers do not agree with each other or have no parseable answer. Experiments on 130 math problems with Qwen3-4B-Thinking-2507 show that ReSolve can correct several answers but is not significantly better than SC@8

Critical strengths
  • The control experiment design is careful and clean. The key same-pool controls share the initial four candidates, and the answer-only ablation removes only the visible derivations.
Critical weaknesses
  1. The paper is hard to read. Something is extremely strange in the paper: (1) the symbols of F/H/D have no explanations. It is unusually to use "F" as a symbol for experimental setting (and "F" is started at "F2", not "F1"). (2) The only significant comparison illustrating ReSolve's clear effectiveness advantage is against SC@4 (but SC@4 has a lower budget cost). According to Appendix B, this comparison was added after the initial evaluation. I think the defensive writing in Appendix is more like an AI transparency disclosure. Thus, I think this paper is not ready for publish and needs more human inspections and major revisions before acceptance.
  2. ReSolve does not show a clear advantage over baselines in Table2, and all methods here are nearly saturated (close to Oracle's performance). I understand this might due to the dataset difficulty. However, such minor difference can be triggered by either randomness or cherrypick (since there are F2 and F3 settings but not F1, I am concerned more about this). Thus, I believe a repeated sampling evaluation or using more difficult datasets with clear discriminative capability are needed (I know that "repeated sampling seems like an AI generated comment, but I do believe it is needed in this paper).
  3. LLM-Blender and MoA have already provide the method paradigm like "generating a new answer from existing candidates". The paper cites these works in Section 4 but neither compare with any of them nor explain the reason why not to compare with them.
  4. The only new component in this paper method, i.e., the specialized route instructions, do not show clearly better than Uniform.
  5. This submission does not appear to provide an artifact package with reproducible code. Thus, I am not convinced by existing results.
Questions for authors
  1. What do F, H and D stand for, and is there an F1? Can you add a table in the main text summarizing F2/F3/H1/D1?
  2. Does ReSolve really show a stable performance over existing methods like MoA and LLM-Blender? If you think they are not fair baselines, please provide explanations. If not, it is better to provide an extra experiment. I do not expect ReSolve to be the most effective one, but a pareto optimization analysis is needed.
  3. Can you provide a repeated sampling experiment to illustrate the stability of ReSolve's advantage? (please use fixed seeds). Also, can you compare ReSolve with sampling more candidates under the same token budget on the triggered problems only? (Again, this is not AI-generated comment. I just need more convincing evidences.)
  4. If you choose to provide the experiments above, please also release the reproduction artifacts like code and configs. I am interested to reproduce these cost-efficiency-related works.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Dexr [1/2] — Fusion comparison, Pareto analysis, and contribution

We added the requested direct generative-fusion control and cost–accuracy analysis. The aggregator uses the same model and four candidate solutions as ReSolve, with an independent neutral synthesis instruction. We evaluate one aggregation layer both on every question and under the original invocation gate.

Table R4. Same-model fusion and moderation. Entries are Hybrid correct / total tokens M on 130 questions per pool.

ConfigurationPool A (F2)Pool B (F3)
ReSolve100 / 12.46499 / 12.349
MoA-style fusion, always-on101 / 13.34998 / 13.296
MoA-style fusion, same gate100 / 12.36899 / 12.403
Uniform loop97 / 12.41799 / 12.426
Uniform one call97 / 12.33098 / 12.373

Relative to always-on fusion, ReSolve uses 6.63%/7.12% fewer tokens, with solved counts 100/99 versus 101/98. The gated fusion and one-call points are explicitly included in the revised Table 6 (p. 7) and Figure 2 (p. 6). Readers can inspect accuracy and cost jointly rather than infer a frontier from one comparator.

Relation to MoA and LLM-Blender. MoA motivates aggregation from candidate responses; our controlled adaptation fixes the proposer, aggregator weights, candidate evidence and accounting to isolate the aggregation operation. It is labeled same-model, single-layer MoA-style fusion. LLM-Blender additionally uses separately trained PairRanker and GenFuser components. The revised related work and Appendix H.3 (p. 21) state these component differences and distinguish the controlled adaptation from a full heterogeneous MoA or LLM-Blender reproduction.

What the additional evidence contributes. ReSolve defines a bounded policy linking answer availability, evidence-conditioned moderation and acceptance. Its mechanism is now tested against independent restart under the same gate, rather than only fixed SC@8. Table R1 in our response to Reviewer Ho3k [1/2] provides all budget-matched and full-length sampling points: ReSolve solves 100/99 versus 94/96 for same-gate +4, using 23.47%/25.09% fewer tokens. This directly addresses the possibility that savings arise only from selective allocation.

The evidence intervention is also replicated: deleting derivations changes 100→93 and 99→95 while increasing moderation cost. Response to Reviewer Whwx [1/2] gives the complete intervention and paired results. Table R3 in our response to Reviewer Whwx [2/2] adds 674 OlympiadBench questions and two backbones, where ReSolve again achieves lower token consumption than SC@8 with equal or higher solved counts. The revised Introduction organizes the contribution around these experimentally separated decisions.

Beyond selection. A fixed-pool oracle uses the reference answer to choose among existing candidates. ReSolve instead generates from candidate evidence without reference-answer access. The Olympiad extension includes two Oracle@4-unsolved recoveries (indices 109/395); the revised case study and appendix trace how unfinished intermediate work supports a new answer.

AuthorsOpenReview ↗

Response to Reviewer Dexr [2/2] — Setting names, F1, repeated sampling, and artifacts

Setting names are now explicit in the main text. Table 2 (p. 4) maps Competition Pool A/B to the original F2/F3 identifiers and lists the questions, model/budget and purpose of the historical studies. F means fresh candidate sampling; H1 is the historical shared pool; D1 is the diagnostic intervention study on that same pool. A/B each contain 130 competition questions; H1/D1 share 160 questions and a separate 16K moderator-output protocol. Main prose uses the descriptive names, retaining the original IDs for traceability.

F1 exists and is reported. It is the earlier completed candidate draw, superseded in the original primary source registry by F2/F3. We matched the archived F1 summary to its source manifest and added the completed-draw inventory in Table 25 (p. 25):

Completed drawSC4SC8ReSolveReSolve / SC8 tokens M
Earlier F190919412.257 / 23.140
Pool A (F2)919410012.464 / 23.193
Pool B (F3)92969912.349 / 23.378

All three completed draws show more solved questions than SC4/SC8 and approximately 46–47% fewer tokens than SC8. A/B remain the primary pools for the matched mechanism interventions. The appendix records fixed sampling identities and candidate ordering; repeated outcomes remain grouped by the same 130 problem identities rather than treated as new questions.

The post-hoc SC4 comparison. Its secondary status is retained. The new positive argument directly answers the requested budget question: same-gate continuation under comparable budget yields 91/92 versus ReSolve's 100/99, and the full-length curve yields 94/96 at greater cost. Table R1 in our response to Reviewer Ho3k [1/2] reports every intermediate point and the exact paired tests. This makes the continued-sampling controls, expanded evaluation and actual cost frontier central to the revision.

Reproduction materials. We prepared an anonymous source bundle containing frozen inference/controller code, sampling configurations, hash-checked score summaries, paired records and CPU table reconstruction. A portable inference entry point and controller-parity checks accompany the source; CPU table reconstruction requires no model weights or API credentials. The revision's reproduction attachment is the access point for this bundle. Appendix J (p. 25) explains the inventory and reconstruction procedure.

The shared results are centralized to reduce repetition: our response to Reviewer Dexr [1/2] contains fusion and Pareto comparisons, our response to Reviewer Ho3k [1/2] contains selective sampling, and our response to Reviewer Whwx [2/2] contains the full external benchmark and two new backbones. Each corresponds to an actual table or figure in the revised PDF.

Reviewer_WhwxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

ReSolve samples four solutions and selectively generates a moderated solution when answers disagree or cannot be extracted. Across two candidate pools on 130 mathematics problems, it solves 100 and 99 problems versus 91 and 92 for four-sample voting. Removing derivations reduces accuracy. Compared with eight-sample voting, it uses 46-47% fewer tokens without establishing an accuracy advantage.

Critical strengths
  1. The main comparisons share the same initial candidates and model weights, and token accounting includes reading candidate evidence and generating moderator responses. The analysis groups repeated outcomes by problem rather than treating the two candidate pools as independent test sets.
  2. The answer-only ablation retains answer keys, vote counts, controller settings, and the per-state output-cap rule while removing derivations, supporting their value beyond answer counts. The always-on Uniform control separately tests selective invocation. The authors acknowledge that superiority over SC@8 and Uniform is not established, distinguish improving the vote from preserving every correct candidate, and avoid equating fewer tokens with faster execution.
Critical weaknesses
  1. The methodological contribution is incremental. Generative fusion of candidate outputs is already studied by LLM-Blender, while Adaptive-Consistency allocates sampling effort using answer agreement. ReSolve combines candidate-conditioned generation with an answer-histogram gate and bounded acceptance rules. This training-free integration is useful, but the general principles of reusing candidate responses and selectively allocating inference computation are established. The distinctive contribution lies mainly in the particular controller and its empirical analysis.
  2. The most important missing baseline is adaptive additional sampling under the final evaluation protocol. SC@8 generates four extra candidates for every problem, whereas ReSolve allocates further computation only to selected problems. The comparison therefore changes both where computation is spent and whether it reuses derivations. The answer-only ablation supports the value of visible evidence, and the historical D1 fresh-solve control provides additional support under a different budget. Neither establishes the efficiency advantage over selective independent sampling in F2/F3. Evaluate Adaptive-Consistency, or a baseline that starts from the same four candidates and adds samples only on triggered problems, and compare accuracy across comparable total-token budgets. The principal F2/F3 evaluations use the same 130 problems and one 4B backbone, so the independent pools leave model dependence and sensitivity to the problem collection unresolved. Repeating the evidence-removal and always-on controls on F3 and testing another reasoning backbone would strengthen this evidence.
Questions for authors
  1. The F2 one-call Uniform control already matches closed-loop Uniform, and Uniform matches ReSolve on F3. Can you repeat the one-call comparison on F3 and report paired outcomes and token costs against full ReSolve? This would clarify whether the more elaborate routing and feedback procedure offers a practical benefit over a single generic moderation call.
  2. Agreement can bypass moderation even when only one candidate has a parseable answer. In F2, 14 skipped problems have only one parseable candidate, and six are correct. How would requiring at least two parseable agreeing candidates affect accuracy and cost? This would help assess whether answer availability should be treated separately from agreement.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Whwx [1/2] — Replicated interventions, one call, and answer availability

We completed the requested Pool B (F3) interventions while preserving its initial candidates, controller and scoring. The revised PDF places both-pool evidence removal in Table 5 (p. 6) and the gate study in Table 18 (p. 20).

Table R2. Replicated interventions. Counts are Hybrid correct /130; costs are millions of input plus output tokens.

InterventionPool A (F2)Pool B (F3)
Full evidence100; moderation 0.753M99; moderation 0.659M
Answer-only93; moderation 1.276M95; moderation 1.410M
Selective Uniform97; 50 calls99; 48 calls
Always-on Uniform97; 136 calls99; 134 calls
Original ReSolve gate100; total 12.464M99; total 12.349M
At least two agreeing parseable answers100; total 12.830M100; total 12.635M

Evidence removal on F3. Full versus Answer-only gives six fixes/two breaks, p=0.2891. Removing derivations increases calls from 47 to 70 and moderation tokens from 0.659M to 1.410M. Retaining evidence therefore saves 53.30% of moderation tokens, with four more solved questions. All 44 archived Full outputs were independently regraded and matched their original Hybrid and CV labels. This repeats the direction of the Pool A intervention; the PDF gives each pool's paired statistics separately.

Always-on on F3. The 86 additional checks give zero Hybrid fixes and zero breaks. Whole-pipeline tokens increase from 12.426M to 13.231M; selectivity saves 52.23% of moderation tokens. The corresponding Pool A saving is 53.61%.

Requested one-call outcome. On Pool B, ReSolve gives 99 correct versus 98 for one-call Uniform, with three fixes/two breaks (exact p=1). Total tokens are 12.349M versus 12.373M. Pool A gives 100 versus 97, with four fixes/one break. The complete one-call/loop operating points are in Table 6 (p. 7).

Answer availability versus agreement. We require at least two parseable agreeing answers at every controller step, including later states. We reran all affected trajectories (20 in A, 18 in B). This gives 100 correct in both pools: zero fixes/breaks in A and one fix/zero breaks in B, for 2.94%/2.32% extra total tokens. Thus the revision provides a directly tested parameterization of answer availability, rather than treating a lone parseable answer as the only possible bypass rule.

Selective independent sampling. To avoid duplicating a shared result, Table R1 in our response to Reviewer Ho3k [1/2] contains the full same-gate budget-matched and complete-length sampling curves. ReSolve uses 23.47%/25.09% fewer tokens than same-gate +4 complete candidates and solves 100/99 versus 94/96. All configurations are in the revised PDF, with the original candidate and gate identities retained.

AuthorsOpenReview ↗

Response to Reviewer Whwx [2/2] — Complete benchmark and two additional backbones

We expanded both benchmark coverage and model coverage. The revision adds the complete 674-question OlympiadBench evaluation and two additional backbones, GPT-OSS-120B and Qwen3-Next. The compact backbone comparison uses SC@4, SC@8, independent Direct solving and ReSolve; the full sources and controls are indexed in the appendix.

Table R3. Expanded evaluation. Counts use Hybrid. Tokens include every consumed proposal, moderation input and moderation output.

SettingNSC4SC8DirectReSolveReSolve / SC8 tokens M
OlympiadBench, Qwen3-4B67458258656858639.383 / 73.735
GPT-OSS-120B low130606652702.060 / 2.418
Qwen3-Next130767667813.662 / 6.013

External benchmark. On all 674 OlympiadBench questions, ReSolve matches SC@8's 586 solved questions with 46.59% fewer total tokens. Same-gate +4 also solves 586 but consumes 52.625M tokens, giving 25.16% savings for ReSolve. Against Direct32, the paired improvement is 28 fixes/10 breaks (exact p=0.0051; six-contrast Holm p=0.0306). This evaluation generates 5,392 candidates and 2,022 method outputs. Exact whitespace-normalized text overlap with the competition panel is zero.

Additional backbones. Each model generates a fresh eight-candidate pool on the original 130 competition questions. ReSolve accesses only the first four, while SC@8 receives all eight. On GPT-OSS, ReSolve solves four more questions than SC@8 with 14.79% fewer tokens. On Qwen-Next, it solves five more with 39.10% fewer tokens. Both preserve the original answer-distribution gate, maximum of three moderation calls, patience two and decoding temperatures.

Each added model yields 130 pools, 1,040 candidates and 390 method outputs. The revised Table 7 (p. 7) and Figure 3 (p. 8) show the transfer results; Appendix I (p. 22) records complete configurations, full controls, CV sensitivity and paired statistics.

Together with the same-gate sampling controls in our response to Reviewer Ho3k [1/2] and the replicated interventions in our response to Reviewer Whwx [1/2], these results test three distinct questions: how to spend continuation computation, which candidate information matters, and whether the measured efficiency behavior extends beyond the original competition evaluation.

Reviewer_WhwxOpenReview ↗

Response to authors rebuttal

I appreciate the authors’ efforts in providing additional clarifications in the rebuttal. However, my primary concern regarding the novelty of the proposed approach remains unresolved. Therefore, I have decided to maintain my original rating.

Reviewer_Ho3KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReSolve is a training-free inference procedure that generates four candidate solutions and selectively invokes the same model to inspect their reasoning when answers disagree or cannot be extracted. On 130 competition-math problems with two independently sampled candidate pools, it improves over four-sample voting and uses fewer tokens than eight-sample voting. Ablations examine the contributions of candidate derivations and selective invocation.

Critical strengths

The evidence-removal ablation is useful: retaining candidate derivations improves accuracy while reducing moderation tokens compared with providing only answers and vote counts. I also appreciate the shared-candidate comparisons and the careful distinction between observed gains and statistically supported conclusions.

Critical weaknesses

The contribution is limited relative to existing candidate fusion, self-refinement, and adaptive computation methods. The distinctive route-specific instructions do not show a reliable advantage over generic Uniform moderation: the two methods solve 100 versus 97 problems in F2 and 99 versus 99 in F3. The results support the usefulness of candidate-conditioned verification, but provide little evidence that the proposed routing and feedback machinery adds substantial value.

The evaluation is too narrow to establish a compelling advance. It uses one 4B model and 130 unique problems; resampling candidates does not broaden that coverage. More importantly, the main efficiency comparison uses fixed SC@8, which spends additional computation on every problem. A baseline that selectively generates fresh solutions under the same gate and comparable budget is needed to separate the benefit of reasoning reuse from that of selective allocation. The historical fresh-solve control is helpful but does not resolve this under the final protocol.

Questions for authors

Under the final protocol, how does ReSolve compare with selectively generating fresh candidates using the same invocation gate and a comparable token budget?

Is there evidence that the multi-round controller improves over a single generic moderation call? The F2 Uniform results suggest that one call already captures most of the benefit.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Ho3k [1/2] — Same-gate, comparable-budget sampling

The new experiments directly separate where further computation is allocated from whether it reuses existing reasoning. They share the original first four candidates, the invocation gate, and the 44 triggered questions in each competition pool. The revised PDF reports both budget-matched and full-length controls in Table 4 (p. 5) and the cost–accuracy plot Figure 2 (p. 6).

Table R1. Same-gate continuation under the final protocol. Entries are Hybrid correct / total tokens in millions; each pool contains the same 130 questions.

Strategy after the original gatePool A (F2)Pool B (F3)
ReSolve100 / 12.46499 / 12.349
Budget-capped +1 fresh candidate91 / 12.44592 / 12.336
Budget-capped +2 fresh candidates91 / 12.46392 / 12.348
Budget-capped +4 fresh candidates91 / 12.46392 / 12.349
Full-length +1 candidate92 / 12.85695 / 12.864
Full-length +2 candidates94 / 13.98996 / 14.049
Full-length +3 candidates94 / 15.16899 / 15.280
Full-length +4 candidates94 / 16.28696 / 16.484
Sequential agreement, up to +494 / 15.63896 / 15.717

Comparable budget. Each capped configuration receives ReSolve's continuation-budget ceiling, including fresh-generation input and output, then votes over the original and additional candidates. ReSolve gains 9/7 questions with zero breaks; exact paired p=0.0039/0.0156, and Holm p=0.0117/0.0469 over the three sampling contrasts. The PDF records the resulting generation caps alongside the complete-length control.

Complete independent reasoning. The full-length curve replays complete fifth-through-eighth trajectories in their original order, without shortening candidates to match a budget. Against same-gate +4, ReSolve solves 6/3 more questions with 23.47%/25.09% fewer total tokens. Pool B's +3 point also reaches 99, but costs 15.280M rather than 12.349M. The sequential rule stops when at least two answers are parseable and top support reaches 75%, or at four additions; it is included explicitly in Table R1.

Thus the efficiency result persists when the gate is shared: the benefit comes from what the continuation computes with the available evidence, beyond simply skipping easy questions. All intermediate points are retained in the revised tables, rather than selecting a single additional-sampling budget.

The broader evaluation is collected in Table R3 in our response to Reviewer Whwx [2/2]: complete OlympiadBench plus GPT-OSS-120B and Qwen3-Next. It independently extends the same accuracy–token argument beyond the original backbone and competition panel.

AuthorsOpenReview ↗

Response to Reviewer Ho3k [2/2] — Single-call comparison and contribution

Single-call comparison. The revised Table 6 (p. 7) reports ReSolve, generic Uniform, one-call Uniform, and generative fusion under shared candidates and complete token accounting. ReSolve solves 100/99 questions in Pools A/B, versus 97/98 for one-call Uniform. In Pool B, ReSolve costs 12.349M tokens versus 12.373M for one call. Our response to Reviewer Whwx [1/2] gives its requested paired outcomes and the replicated evidence/invocation interventions; Table R4 in our response to Reviewer Dexr [1/2] gives the full fusion comparison.

What is being tested. ReSolve's contribution is an explicit inference policy for converting sampled reasoning into further computation: answer availability and agreement determine invocation; candidate derivations provide the evidence; generated responses update the pool under bounded acceptance. The revision tests these decisions separately rather than treating the method as one undifferentiated prompt:

  • Evidence: removing derivations changes 100→93 and 99→95 correct and increases moderation tokens in both pools.
  • Invocation: adding verification to all skipped agreement states adds 86 calls per pool with zero Hybrid fixes or breaks.
  • Continuation: the same-gate sampling controls in our response to Reviewer Ho3k [1/2] show what independent restart costs relative to evidence reuse.
  • Instruction and acceptance: one-call, Uniform-loop and fusion operating points are all retained in the cost–accuracy analysis.

Generic verification already performs candidate-conditioned reasoning, so its performance is useful evidence about the shared reuse mechanism. We distinguish this mechanism from the particular route wording. The controller supplies an inspectable, bounded implementation; its practical value is assessed through the measured operating points and interventions above.

Broader coverage. The revised Table 7 (p. 7) adds all 674 OlympiadBench questions: ReSolve and SC@8 each solve 586, with 46.59% fewer tokens for ReSolve; against same-gate +4 the saving is 25.16% at the same total solved count. The two additional backbones in our response to Reviewer Whwx [2/2] each show more solved problems than SC@8 with lower token use.

The revised Introduction, Section 3.3 (p. 5), Section 3.5 (p. 7), and related-work discussion connect these experiments to adaptive sampling, candidate fusion and refinement. This makes the positive claim concrete: existing derivations can replace a substantial amount of additional independent reasoning, and the shared-gate controls identify that effect directly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.