REVIEW 4 major objections 5 minor 13 references
GenSelect: A Generative Approach to Best-of-N
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A reasoning LLM can pick the best of N candidate math solutions in one shot, beating pointwise scoring.
desk verdict GenSelect is a simple and useful empirical result—N-ary comparison with reasoning models beats pointwise scoring—but the unvalidated solution-summarization step and missing pairwise baseline keep it from being a clean win. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the GenSelect prompt, which converts best-of-N selection into a single N-ary comparison: the model must screen all candidate solutions, identify conflicting answers, eliminate unsound ones, and emit exactly one index as its judgment. Candidate solutions are first compressed from full reasoning traces into clean summaries by Qwen2.5-32B-Instruct, because the authors found in initial experiments that summaries work as well as full traces. For $N$ larger than the context window, the method uses an N-ary knockout tournament adapted from pairwise knockout design: with 64 candidates, a 16-way tournament needs two rounds and five total comparisons, whereas binary comparison needs six rounds and 63 comparisons.
What would settle it
On a held-out set of QwQ solutions with full traces, run GenSelect twice per problem: once on full traces and once on summarized versions, and compare the selected answer to the oracle best answer. If the two pipelines disagree on more than a small fraction of problems, or if full-trace GenSelect is meaningfully more accurate than summary GenSelect, the summary-faithfulness premise fails and the baselines were compared on different content.
Extended reading notes
Core claim
The authors establish that a long-reasoning LLM, given the problem and $N$ candidate solutions together, can reliably act as an N-ary judge: it groups solutions by final answer, explains contradictions, eliminates clearly wrong ones, evaluates the rest, and ends with a judgment that names the best index. They report that current open reasoning models do this zero-shot, and that the solution selected this way beats pointwise GenRM and majority voting when the same QwQ model both generates and selects. They further show that selection accuracy is stable as $N$ varies from 2 to 16, which justifies using large N-ary knockout tournaments to cut the number of sequential rounds and total comparisons. The paper's headline numbers are QwQ GenSelect@8 at 73.4% on Comp-Math-24-25, versus 69.1% for QwQ GenRM and 68.4% for majority voting, with an oracle Pass@64 ceiling of 85.2%.
Load-bearing premise
The compressed solution summaries preserve every step that determines correctness, so judging a summary is equivalent to judging the full reasoning trace; the paper says only that initial experiments showed no benefit from full traces, and gives no faithfulness measurement.
Editorial extensions
If this is right
- On competition math, GenSelect with no training beats majority voting, a 72B discriminative outcome reward model, and a pointwise GenRM built from the same QwQ model.
- GenSelect accuracy is stable across $N = 2, 4, 8, 16$, so widening the comparison width is a safe way to cut latency and compute in knockout tournaments.
- When inference compute is split between generation and selection, GenSelect is the better use of a fixed budget for QwQ at every tested $N$, and for DeepSeek-R1-0528 at $N \ge 8$.
- A model selecting among its own generations (Self-GenSelect) raises QwQ from 68.4% majority voting to 73.0%, and DeepSeek-R1-0528 from 84.0% to 87.1% on the full Comp-Math-24-25 set.
- The method applies immediately at test time to any current long-reasoning model without specialized reward-model training.
Reading between the lines
- An unstated consequence is that the comparison width $N$ itself becomes a tunable test-time knob: one could search for the optimal $N$ for a given model and budget, rather than just using the largest $N$ the context window allows.
- A testable extension beyond math: the same 'screen, eliminate, justify' prompt could be used for code repair or multi-document question answering, where candidates are alternate outputs and correctness is less cleanly verifiable.
- Because summaries are generated by a smaller instruct model, the faithfulness of those summaries is a bottleneck; if a better summarizer or full traces improve GenSelect results, the reported gains would shift accordingly.
- The GenSelect formulation invites an RL-trained verifier that distills N-ary comparison into a smaller model, which the authors mention only as future work but which would turn the method from a prompting trick into a scalable component.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes GenSelect, a best-of-N selection method in which a reasoning LLM is prompted to compare N candidate solutions simultaneously and return the index of the best one; for candidate pools larger than the context window, the method uses an N-ary knockout tournament. The experiments on a self-curated Comp-Math-24-25 benchmark (256 AIME/HMMT problems) show that QwQ-32B and DeepSeek-R1-0528 with GenSelect outperform majority voting and a pointwise GenRM baseline, and that GenSelect is stable across tournament widths and compute-allocation choices.
Significance. If the result holds, GenSelect is practically significant: it offers a training-free, prompt-only method for improving test-time scaling in math reasoning, and its N-ary tournament has lower latency than pairwise knockout tournaments. The paper makes concrete falsifiable claims, reports prompts in the appendix, and includes a useful compute-allocation analysis for two open-weight reasoning models. However, several load-bearing pieces of evidence are missing: summary faithfulness is not audited, no pairwise baseline is included, error bars or significance tests are absent, and the HMMT evaluation relies on an LLM judge from the authors' own prior work. The central claim is therefore defensible but not yet established at the required level.
major comments (4)
- [§3, Appendix A.1] The solution summaries are load-bearing for every method, but their faithfulness is not established. The summary prompt tells Qwen2.5-32B-Instruct to write "a clean version of the final correct solution" and presupposes that "eventually that solution gets to the right approach and solves the problem." For incorrect or abandoned reasoning traces, this instruction may systematically produce clean, plausible-looking summaries that omit the actual errors or even change the boxed final answer. Since both GenRM and GenSelect score these summaries rather than the raw QwQ traces, a systematic distortion would mean the comparison is between artifacts rather than candidate solutions. The paper's statement in Section 3 that "initial experiments revealed no significant benefit from using complete reasoning traces" is not supported by any reported data. I ask for a faithfulness audit (e.g., checking whether the summary preserves the original final answer and key steps), an error analysis of summaries, or an ablation running GenSelect and GenRM on full traces.
- [§4.3, Table 3] The empirical comparison omits a pairwise baseline, even though the paper's motivation is framed around the limitations of pairwise GenRMs. Table 3 includes majority voting, a discriminative RM, and a pointwise QwQ GenRM, but no pairwise GenRM knockout tournament such as PairJudge-RM or a pairwise prompt converted to a tournament. Without this baseline, the headline claim that GenSelect "outperforms existing scoring approaches" is only established against pointwise scoring, not against the pairwise paradigm that the paper explicitly discusses as the main alternative. A pairwise baseline using the same QwQ model and the same tournament structure is needed to isolate the benefit of N-ary comparison.
- [§4.3, Tables 3 and 5] No error bars, confidence intervals, or significance tests are reported. The headline difference in Table 3 is 73.4% vs. 69.1% on 256 problems; under an independent-proportion approximation, the standard error of each proportion is roughly 2.8 points and the standard error of the difference is roughly 4.0 points, so the observed 4.3-point advantage is on the order of one standard error. Since the methods are evaluated on the same candidate pools, a paired test would be more powerful, but no paired analysis is reported. Please report variance across problem subsets, confidence intervals, and a paired significance test for the main comparisons in Tables 3 and 5.
- [§4.3, Table 5 footnote] The HMMT evaluation relies on the "LLM-judge setup of (Toshniwal et al., 2025)", a work that shares authors with this paper. The GenSelect gains are largest on the HMMT-24-25 split (e.g., QwQ-32B goes from 56.1% maj@64 to 70.4% with Self GenSelect@32, compared with 86.7% to 90.0% on AIME24), so the accuracy and neutrality of the judge are load-bearing. No validation of this judge on the HMMT subset is reported. I ask for a validation against human labels on a sample, or a sensitivity analysis using a second, independent judge, or a clear statement of how the judge is scored and why it is not a source of bias.
minor comments (5)
- [§4.3, Table 5] The term "Self-GenSelect" is used without a definition; please state explicitly that the same model generates the candidate solutions and performs the GenSelect comparisons.
- [§4.2, Tables 3 and 5] Table 5 reports top-p 0.95, but Table 3 and the verification sampling in Section 3.1 do not state the top-p used; please report the exact sampling parameters for all experiments.
- [§3.1, Figure 1] The GenRM baseline says the model is sampled multiple times, but it is not specified how multiple verifications are aggregated (e.g., majority vote over Yes/No judgments, average token probability, or another method); please clarify, since this affects the fairness of the comparison with GenSelect@8.
- [§3.2, Table 5] For N=8 with 64 candidates, the tournament uses 8 parallel GenSelect calls in the first round and one call in the second round, so the text "two rounds of scoring eight solutions each" is ambiguous; please state the total number of GenSelect calls per problem.
- [§4.1, Table 1] The benchmark is self-curated and not released; for reproducibility, please provide the exact problem list or release the benchmark, and state the filtering criteria for excluding proof-based or partial-credit questions.
Circularity Check
No significant circularity: GenSelect is an empirical prompt-based method evaluated against baselines on the same input representation; the only self-citation is a shared answer-verification tool that does not encode the method's output.
full rationale
GenSelect's derivation chain is self-contained: the method directly prompts a reasoning LLM to select the best of N candidate solutions, and its accuracy is measured by whether the selected candidate's final answer matches the ground truth. No parameter is fitted to the evaluation data, no prediction is constructed from its own inputs by definition, and no uniqueness theorem or ansatz is imported from the authors' prior work. The one citation to the authors' own prior work is the HMMT answer-verification LLM-judge setup of Toshniwal et al. (2025). That judge is applied identically to all compared methods and verifies final answers rather than selecting solutions, so it does not reduce GenSelect's reported advantage to a self-citation. The summary-faithfulness assumption (Section 3: 'Our initial experiments revealed no significant benefit from using complete reasoning traces compared to solution summaries') is a potentially important unvalidated preprocessing choice, but it applies equally to the GenRM and GenSelect pipelines and therefore does not make the comparison circular. The paper's central claim remains an empirical, falsifiable result: a zero-shot N-ary comparison prompt outperforms pointwise scoring on the same summarized candidates.
Assumptions & free parameters
free parameters (4)
- Tournament width N =
8
- GenSelect repetitions =
32
- Solution summary model =
Qwen2.5-32B-Instruct
- Sampling temperature/top-p =
0.6 / 0.95
assumptions (4)
- domain assumption Comp-Math-24-25 is a representative benchmark for competition-level math reasoning.
- domain assumption The HMMT answer verification via the LLM-judge setup of Toshniwal et al. (2025) is unbiased.
- domain assumption Solution summaries are faithful to the original solutions.
- ad hoc to paper The reasoning models' long-reasoning capability can be transferred to N-ary comparison via a simple prompt.
Cite this review
Pith. "Pith review of GenSelect: A Generative Approach to Best-of-N." pith.science (2026). https://pith.science/paper/R5GAB3QU
@misc{pith2026250717797,
author = {Pith},
title = {Pith review of: GenSelect: A Generative Approach to Best-of-N},
year = {2026},
howpublished = {\url{https://pith.science/paper/R5GAB3QU}},
note = {Machine review of arXiv:2507.17797}
}
read the original abstract
Generative reward models with parallel sampling have enabled effective test-time scaling for reasoning tasks. Current approaches employ pointwise scoring of individual solutions or pairwise comparisons. However, pointwise methods underutilize LLMs' comparative abilities, while pairwise methods scale inefficiently with larger sampling budgets. We introduce GenSelect, where the LLM uses long reasoning to select the best solution among N candidates. This leverages LLMs' comparative strengths while scaling efficiently across parallel sampling budgets. For math reasoning, we demonstrate that reasoning models, such as QwQ and DeepSeek-R1-0528, excel at GenSelect, outperforming existing scoring approaches with simple prompting.
Figures
Reference graph
Works this paper leans on
-
[3]
Guo, J., Chi, Z., Dong, L., Dong, Q., Wu, X., Huang, S., and Wei, F
URL https://arxiv.org/abs/2501.12948. Guo, J., Chi, Z., Dong, L., Dong, Q., Wu, X., Huang, S., and Wei, F. Reward Reasoning Model,
-
[4]
Hassid, M., Synnaeve, G., Adi, Y ., and Schwartz, R
URL https://arxiv.org/abs/2505.14674. Hassid, M., Synnaeve, G., Adi, Y ., and Schwartz, R. Don’t Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning,
-
[5]
Jiang, D., Ren, X., and Lin, B
URL https: //arxiv.org/abs/2505.17813. Jiang, D., Ren, X., and Lin, B. Y . LLM-Blender: Ensem- bling Large Language Models with Pairwise Ranking and Generative Fusion. In Rogers, A., Boyd-Graber, J., and Okazaki, N. (eds.), ACL,
-
[8]
URL https: //arxiv.org/abs/2410.12832. OpenAI. OpenAI o1 System Card,
-
[9]
URL https: //arxiv.org/abs/2412.16720. Shi, W. and Jin, X. Heimdall: test-time scaling on the generative verification,
-
[10]
URL https://arxiv. org/abs/2504.10337. Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. In ICLR,
-
[11]
URL https://arxiv.org/abs/ 2408.00724. Yang, A., Zhang, B., Hui, B., Gao, B., Yu, B., Li, C., Liu, D., Tu, J., Zhou, J., Lin, J., Lu, K., Xue, M., Lin, R., Liu, T., Ren, X., and Zhang, Z. Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self- Improvement,
-
[12]
Zhang, L., Hosseini, A., Bansal, H., Kazemi, M., Kumar, A., and Agarwal, R
URL https://arxiv.org/ abs/2409.12122. Zhang, L., Hosseini, A., Bansal, H., Kazemi, M., Kumar, A., and Agarwal, R. Generative Verifiers: Reward Mod- eling as Next-Token Prediction. In ICLR, 2025a. Zhang, Z., Zheng, C., Wu, Y ., Zhang, B., Lin, R., Yu, B., Liu, D., Zhou, J., and Lin, J. The Lessons of De- veloping Process Reward Models in Mathematical Rea-...
Show all 13 references
-
[13]
org/abs/2502.01839
URL https://arxiv. org/abs/2502.01839. 8 GenSelect: A Generative Approach to Best-of-N A. Prompts A.1. Solution Summary Prompt Summary Prompt I w i l l g i v e you a math problem and a l o n g s o l u t i o n t o t h a t problem e x p l o r i n g d i f f e r e n t a p p r o a ...
-
[2021]
DeepSeek-AI
URL https:// arxiv.org/abs/2110.14168. DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Ca- pability in LLMs via Reinforcement Learning,
-
[2023]
Liu, Y ., Zhou, H., Guo, Z., Shareghi, E., Vulic, I., Korhonen, A., and Collier, N
URL https://arxiv.org/abs/2305.20050. Liu, Y ., Zhou, H., Guo, Z., Shareghi, E., Vulic, I., Korhonen, A., and Collier, N. Aligning with Hu- man Judgement: The Role of Pairwise Preference in Large Language Model Evaluators. arXiv preprint arXiv:2403.16950,
-
[2024]
Pair- Judge RM: Perform Best-of-N Sampling with Knockout 7 GenSelect: A Generative Approach to Best-of-N Tournament, 2025a
Liu, Y ., Yao, Z., Min, R., Cao, Y ., Hou, L., and Li, J. Pair- Judge RM: Perform Best-of-N Sampling with Knockout 7 GenSelect: A Generative Approach to Best-of-N Tournament, 2025a. URL https://arxiv.org/ abs/2501.13007. Liu, Z., Chen, Y ., Shoeybi, M., Catanzaro, B., and Ping...
-
[2025]
Cobbe, K., Kosaraju, V ., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J
URL https://arxiv.org/abs/ 2505.02387. Cobbe, K., Kosaraju, V ., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. Training Verifiers to Solve Math Word Problems,
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.