Pith. sign in

REVIEW 3 major objections 5 minor 14 references

Rethinking Reward Models for Multi-Domain Test-Time Scaling

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read A generative outcome reward model, not a stepwise process verifier, is the most robust choice for multi-domain test-time scaling.

desk verdict Solid empirical comparison with a real finding—gORM beats PRMs across 14 domains—but the theoretical explanation analyzes aggregation schemes the experiments don't use, so the causal story is provisional. read the letter →

arxiv 2510.00492 v3 pith:MWPBOOML submitted 2025-10-01 cs.AI

classification cs.AI
keywords rewardmodelsprocessoutcomegenerativeverifierstest-timescalingbest-of-Nmulti-domainreasoningchain-of-thoughtverification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that, for test-time scaling across many domains, fine-grained step-level verification is not the right default. In a unified comparison of four verifier types on 14 domains, a generative outcome reward model—one that writes a verification rationale and a final Yes/No verdict—consistently outperforms discriminative outcome models, discriminative process models, and generative process models. Process reward models lose because they compound per-step errors as reasoning chains lengthen, and because auto-labeled step data carries noise that long chains amplify. If correct, the result overturns the common assumption that finer-grained supervision is always better, and points to outcome-level generative verification for multi-domain deployment.

What carries the argument

The comparison rests on four verifier variants—discriminative vs generative, and outcome vs process—and the central mechanism is the difference between scoring the whole chain once (outcome) and aggregating per-step scores (process: minimum aggregation for dPRM in experiments, product aggregation in the theory, and sampling with a final 'Yes' probability for gPRM). Theorems 4.1–4.3 give log-error bounds: ORM error is independent of chain length T, while PRM error lower bounds grow linearly in T. The length-distribution shift from consensus filtering, measured by Wasserstein distance, explains gPRM's extra degradation in the multi-domain setting.

What would settle it

Take a set of long CoTs with known outcome labels and compare dPRM scores computed by minimum aggregation versus product aggregation; if product aggregation does not reproduce the observed F1 decline with chain length, then the linear-in-length bound in the theory does not explain the empirical dPRM result. Alternatively, train gPRM without consensus filtering and match its training length distribution to the test distribution; the paper's account predicts gPRM would then match gORM.

Watch

Extended reading notes

Core claim

Across 14 domains of MMLU-Pro, with controlled backbones and four verifier variants, the paper finds the conventional math-domain ranking inverts: discriminative outcome and process models perform on par, a generative process reward model is not competitive, and a generative outcome reward model (gORM) is the most robust, improving best-of-N selection consistently across every domain. The cause is stepwise scoring: per-step errors and label noise compound as reasoning chains lengthen, so PRMs miss correct answers reached after an early mistake. The paper backs this with log-error bounds showing ORM error independent of chain length while PRM error grows at least linearly, and with empirical

Load-bearing premise

The theory's error-compounding result assumes that the way step scores are combined in the proof (multiplying them) matches how the experiments combine them (taking the minimum), and that per-step mistakes do not cancel as chains grow; if either fails, the linear-in-length bound would not explain the empirical results.

Editorial extensions

If this is right

  • For multi-domain deployment, the default verifier should be a generative outcome reward model; it yields consistent Best-of-N gains over dORM, dPRM, and gPRM without notable per-domain degradation.
  • The PRM-over-ORM advantage is not universal: it appears in math but disappears or reverses across broader domains such as law, engineering, and health.
  • Long reasoning chains, especially those with self-correction ('aha' moments), are systematically penalized by stepwise aggregators, so verifiers for such chains should avoid per-step scoring.
  • Multi-domain training benefits generative models more than discriminative ones: per-domain training sharply degrades dORM/dPRM, while gORM/gPRM stay close to their multi-domain performance.
  • When compute or latency is tight, dORM is a serviceable fallback; when labels are noisy, outcome-level verification is safer than process-level scoring.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same error-compounding argument implies that stepwise reward signals used in RL training will increasingly bias against long, self-correcting trajectories; outcome-level verification may be a safer reward signal for long-horizon tasks.
  • A testable extension: train gPRM without consensus filtering, matching its training length distribution to the test set; the paper's account predicts gPRM would then recover most of the gap against gORM, isolating length shift from stepwise scoring itself.
  • If gORM's robustness comes from judging the whole chain at once, a similar design could carry over to open-ended generation once a verifiable outcome is defined, though the paper does not test that setting.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a unified comparison of four reward-model variants for best-of-N test-time scaling: discriminative outcome (dORM), discriminative process (dPRM), generative outcome (gORM), and generative process (gPRM) reward models. The evaluation spans a math domain (PRM800K/ProcessBench) and a 14-domain multi-domain setting (MMLU-Pro), with multiple generator LLMs and shared reward backbones. The authors report that, consistent with prior work, dPRM beats dORM and generative variants beat discriminative ones in math; in the multi-domain setting, dORM is on par with dPRM, gPRM is not competitive, and gORM is overall the most robust. They propose two explanations: (i) stepwise scoring/aggregation compounds errors as reasoning chain length grows, supported by log-error lower bounds (Theorems 4.1–4.3); and (ii) label noise and consensus-filtering-induced length distribution shift hurt gPRM. Code, data, and checkpoints are released.

Significance. If the empirical ranking holds, the paper challenges the common assumption that fine-grained process supervision always improves verifier performance in multi-domain test-time scaling, and it provides a practical comparison that is more controlled than much prior work (shared backbones, four variants, multiple generator LLMs). The release of code, datasets, and checkpoints is a concrete strength that will facilitate follow-up work. The theoretical analysis, however, is not tightly connected to the experimental protocols: the dPRM theorem analyzes product aggregation while the experiments use minimum aggregation, and the gPRM theorem concerns single-sample product scores while the deployed estimator is an M-sample average of final-verdict probabilities. The headline abstract claim of consistent significant gains across every domain is also stronger than the per-domain results support. The empirical study itself is valuable and likely correct in its main qualitative trends, but the explanatory narrative needs substantial revision.

major comments (3)
  1. [§4.1 / Theorem 4.2 vs. §2.2, Appendix A.1] Theorem 4.2 analyzes a product-aggregated dPRM: Appendix A.1 defines f_hat_dPRM(x) := ∏_{t=1}^T û_t(x), and the lower bound E[Δ²_dPRM] ≥ (σ²−2γ)T is derived for this product. In contrast, §2.2 states that at test time dPRM 'aggregat[es] the step rewards with the minimum' (following Zeng et al., 2025). Min aggregation has different length behavior: it is dominated by the worst step rather than by the accumulation of all step errors, so the theorem does not bound the error of the estimator actually evaluated in Figures 3–7 and 12. The empirical support in Fig. 12 (Left) uses min-aggregated dPRM and therefore cannot rescue the mismatch. This is load-bearing: the paper's central explanation for why PRMs underperform in multi-domain TTS is precisely that 'stepwise aggregation compounds errors as reasoning length increases,' and that claim is not established for the dPRM variant used in the ex
  2. [§4.1 / Theorem 4.3 vs. Eqs. 8–9, Appendix A.1] Theorem 4.3 bounds the log-error of Δ_gPRM defined in Appendix A.1 as the log of a single-sample product of step-correctness probabilities, ∏_t F_t(x, v_{≤t}). The test-time estimator in Eq. 8 averages M samples of the normalized final-verdict probability p_gPRM(y=1 | x, v_{1:L+}), not the log of a single product. While p_gPRM(y=1|x,v) factors as a product when the final 'Yes' is deterministically equivalent to all steps being correct, the paper does not supply a rigorous bridge from the single-sample lower bound to the M-sample average used with M=10. Theorem A.4 addresses a mean predictor (E[e^L]) rather than the finite-M sample average, and the main text's pointer to 'cf. Eq. 5' is incorrect — Eq. 5 is the gORM estimator, not the gPRM estimator of Eq. 8. The theoretical support for gPRM's empirical degradation is therefore incomplete, and the stated linear-in-T lower bound is not dire
  3. [Abstract / §3.2 / Fig. 6] The abstract claims that gORM yields 'significant and consistent gains across every tested domain.' This is not supported by the paper's own per-domain results in Fig. 6. In Psychology, dORM (81.9) and dPRM (81.7) both exceed gORM (81.3); in Health, dPRM (79.2) exceeds gORM (78.2). The weaker statement in §3.2 that gORM outperforms the others 'without notable degradation in any domain relative to the others' is accurate, but the abstract overstates the finding. Furthermore, §D reports a single run for the multi-domain setting, so the word 'significant' is not backed by error bars or statistical tests. The manuscript should either soften the abstract to match the per-domain results or provide significance testing that justifies the stronger claim.
minor comments (5)
  1. [§2.2] Typo: 'Byeond Khalifa et al. (2025)' should be 'Beyond Khalifa et al. (2025)'.
  2. [Main text, after Theorem 4.3] The sentence referencing 'Theorem. A.4' for gPRM with Monte Carlo estimation cites 'cf. Eq. 5'; this should be Eq. 8, since Eq. 5 defines the gORM estimator.
  3. [§4.1 / Fig. 12] The correlation values in Fig. 12 (Right) are reported without confidence intervals or p-values. Given only 8 binned length categories (or 6 on the right), these correlations should be interpreted cautiously; adding a scatter plot with per-bin standard errors would strengthen the claim.
  4. [§4.2 / Fig. 14] The Wasserstein distance histograms in Fig. 14 aggregate over all domains; per-domain distances in Table 3 show large variation (e.g., Chemistry 6.922 vs. History 0.581 for gPRM). A per-domain scatter with domain labels would make the 'length shift hurts gPRM' mechanism more transparent than the single pooled number.
  5. [§C / Table 1] The table reports average CoTs per question for training and test splits, but the test-time generation uses N=128 for the main evaluation and N=16 for some additional generator LLMs. Clarify in the caption which N applies to which figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the empirical ranking is externally evaluated and the theoretical bounds are conditional on stated assumptions rather than fitted to the outcome.

full rationale

The central empirical claim (gORM outperforms dORM/dPRM/gPRM across MMLU-Pro domains) is an externally measured ranking of four independently trained verifier variants on held-out benchmarks; nothing in the training objectives, aggregation rules, or evaluation protocol forces this ordering by construction. The theoretical results (Theorems 4.1-4.3, A.1-A.4) are conditional variance-of-sum bounds: under per-step variance floors and weak anti-correlation, the log-error of a product-aggregated PRM grows linearly in T. These are theorem statements with explicit assumptions, not fitted parameters renamed as predictions, and the paper does not claim to have estimated sigma^2 or gamma from the benchmark results. The Appendix's use of product aggregation while experiments use min aggregation for dPRM and final-verdict probability for gPRM is an internal-validity mismatch, but it is not a reduction of the conclusion to its inputs. No load-bearing self-citation chain exists: prior work (Zeng et al., Khalifa et al., Zhang et al.) is cited as external methodology, and the authors invoke no uniqueness theorem or prior result of their own to force the choice of gORM. Hence no circular step meets the evidentiary bar; the score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new theoretical entities; it imports existing reward model variants. Its explanatory constructs ('aha' CoTs, length shift, label noise) are categories of existing data, not new entities. The main load-bearing assumptions are the error-structure bounds in the appendix and the implicit equivalence between product and min aggregation.

free parameters (6)
  • M (number of verification CoTs sampled) = 16 (math), 10 (multi-domain)
    Number of MC samples for gORM/gPRM reward estimation; chosen following prior work and affects variance of the reward estimate.
  • Sampling temperature τ for verification CoT generation = 0.6
    Temperature for generating verification CoTs with QwQ-32B; follows Khalifa et al. (2025).
  • Consensus-filtering thresholds = token limits 4096/8192; discard unparsable, Chinese, label-inconsistent CoTs
    Ad hoc filtering rules that determine gORM/gPRM training set sizes and length distributions; central to the length-shift explanation.
  • LoRA rank and alpha = r=16/32, α=32/16
    Hyperparameters for dORM/dPRM vs gORM/gPRM, chosen from prior work; not fitted to the target.
  • Decision threshold for F1 = 0.5
    Standard threshold for outcome verification; a choice, though conventional.
  • Theoretical constants σ², τ², γ, κ = unspecified
    Assumed constants in Theorems A.1–A.4; not estimated from data, chosen so the lower bounds are positive.
assumptions (6)
  • domain assumption A correct final step implies all previous steps are correct (y=z_T=1 ⇒ z_1:T=1).
    Used in Appendix A.1 to factor f(x)=∏u_t(x). Explicitly violated by 'aha' CoTs, as acknowledged in §4.1.
  • ad hoc to paper Variance floors and weak anti-correlation: Var(ξ_t|x) ≥ σ², covariance sums ≥ −γT, and σ² > 2γ.
    Assumptions for Theorems A.1–A.4; not empirically verified in the paper.
  • ad hoc to paper ORM conditional variance is bounded independent of T: Var(ξ̄|x) ≤ τ²_orm.
    Assumption for Theorem 4.1/A.2 to obtain a T-independent error bound; not derived from ORM training.
  • ad hoc to paper Product aggregation of step scores behaves like min aggregation for dPRM.
    Unstated equivalence between the theory (§A.1) and the experimental method (§2.2); no proof is provided.
  • domain assumption Llama-3.1-70B auto-generated process labels on MMLU-Pro are sufficiently reliable for training dPRM/dORM.
    Standard practice following Zeng et al. (2025), but the paper itself argues label noise is substantial, making this a load-bearing assumption for the comparison.
  • domain assumption MMLU-Pro multiple-choice answers are valid verifiable outcome labels.
    Limits scope to verifiable outcomes, as acknowledged in the limitations section.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Reward Models for Multi-Domain Test-Time Scaling." pith.science (2026). https://pith.science/paper/MWPBOOML

@misc{pith2026251000492,
  author       = {Pith},
  title        = {Pith review of: Rethinking Reward Models for Multi-Domain Test-Time Scaling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MWPBOOML}},
  note         = {Machine review of arXiv:2510.00492}
}
read the original abstract

The reliability of large language models (LLMs) during test-time scaling is often assessed with \emph{external verifiers} or \emph{reward models} that distinguish correct reasoning from flawed logic. Prior work has studied both outcome reward models (ORMs), which assess only the final answer, and process reward models (PRMs), which score intermediate reasoning steps. Although PRMs are often viewed as advantageous due to their finer-grained supervision, much of the supporting evidence comes from math-adjacent settings, and their relative benefits across broader domains remain unclear. We present the first unified evaluation of four reward model variants, discriminative ORM and PRM (dORM, dPRM) and generative ORM and PRM (gORM, gPRM), across 14 diverse domains. Contrary to conventional wisdom, we find that (i) dORM performs on par with dPRM, (ii) gPRM is not competitive, and (iii) overall, gORM is the most robust, yielding significant and consistent gains across every tested domain. We attribute the worse performance of gPRM to the stepwise scoring process, which inherits label noise from LLM-based automatic labeling, leading to difficulties in evaluating long reasoning trajectories, including those involving self-correcting reasoning. Both our theoretical analysis and empirical observations indicate that stepwise aggregation compounds errors as reasoning length increases. These findings challenge the common assumption that fine-grained supervision is always better and support generative outcome verification for multi-domain deployment. Our \href{https://github.com/db-Lee/Multi-RM}{\underline{code}} is publicly available to facilitate future research in multi-domain settings.

Figures

Figures reproduced from arXiv: 2510.00492 by the authors.

Figure 1
Figure 1. Evaluating CoTs using gORM and gPRM. To this end, we present the first unified eval￾uation of four verifier variants, discriminative ORM and PRM (dORM, dPRM), and genera￾tive ORM and PRM (gORM, gPRM), across 14 diverse domains. We rigorously review these variants in §2 and, under controlled condi￾tions, evaluate them on math and multi-domain benchmarks (PRM800K, ProcessBench, and MMLU-Pro; Lightman et al., 2024; Zhe… view at source ↗
Figure 2
Figure 2. Conceptual illustration of reward models: r2 is the first incorrect step; the final answer is correct. 2.1 PROBLEM FORMULATION Notation. For a given question q with the corresponding ground-truth (GT) answer a, we leverage the reasoning ability of large language models (LLMs) to reliably predict a by generating a chain￾of-thought (CoT), i.e., r1:T := (r1, . . . ,rT ) ∼ pLLM(· | q). Following Zeng et al. (2025), we s… view at source ↗
Figure 3
Figure 3. Outcome verification results on ProcessBench in the math domain. 1 2 4 8 16 75 80 85 Taks accuracy (%) 1.5B - GSM8K 1 2 4 8 16 66 72 78 84 1.5B - Math 1 2 4 8 16 75 80 85 90 7B - GSM8K 1 2 4 8 16 66 72 78 84 7B - Math Number of CoTs (N) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (36 more)
Figure 5
Figure 5. Figure 5: Outcome verification results on MMLU-Pro in the multi-domain setting. Majority Voting dORM dPRM gORM gPRM 54 60 66 72 Overall 35 40 45 Law 60 65 70 Psychology 50 60 70 Chemistry 72 78 84 Biology 50 60 70 80 Physics 48 52 56 History 60 66 72 78 Economics 50 60 70 80 Mat…
Figure 6
Figure 6. Figure 6: Best-of-N results using Llama-3.1-8B-Instruct on MMLU-Pro in the multi-domain setting. formats in Figs. 20 and 21. We discard any verification CoT whose parsed labels are inconsistent with the targets (e.g., y or z1:T ), corresponding to the consensus filtering in §2.2…
Figure 7
Figure 7. Figure 7: Overall Best-of-N results using four different pLLM on MMLU-Pro in the multi-domain setting. 6 3 0 -6.6 -4.7 -0.3 -0.9 Overall 10 5 0 5 -8.6 -10.1 +2.5 +0.5 Law 9 6 3 0 -3.9 -6.6 -1.2 -8.4 Psychology 10 5 0 -10.9 -7.6 +0.1 -0.6 Chemistry 6 3 0 -3.8 -5.2 +0.2 +0.2 Biolo…
Figure 8
Figure 8. Figure 8: Best-of-N performance gap between all-domain and single-domain training on MMLU-pro [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Results on “aha” CoTs. “Aha” CoTs. As noted in §2.2, PRMs typically assume that once a reasoning step is incorrect, all subsequent steps are incorrect. However, recent rea￾soning models can recover from earlier mistakes and still arrive at the correct answer (an “aha” …
Figure 10
Figure 10. Figure 10: Results on randomly shuffled CoTs. Do ORMs overfit on “aha” CoTs? A natural con￾cern about ORM results on “aha” CoTs in [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Length dist. of “aha” CoTs. Risk increases with CoT length. “Aha” mo￾ments can also lengthen CoTs, an effect espe￾cially pronounced in the multi-domain setting ( [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: (Left): Outcome verification vs. CoT length; (Right): TTS improvement vs. average CoT length. Empirical support. To empirically support Theorems. 4.1 to 4.3, we plot the F1 score (%) for outcome-verification in the multi-domain setting as a function of the number of r…
Figure 13
Figure 13. Figure 13: Effect of labe noise on GSM8K. Label noise risk. Beyond CoT-length effects, la￾bel noise poses an additional risk, especially in multi-domain settings. Since human annotation of long CoTs is more costly in specialized do￾mains such as law and medicine than in math, pr…
Figure 14
Figure 14. Figure 14: Length distribution shift on MMLU-Pro (overall/per-domain) measured by Wasserstein distance. 0.5 1.0 4 6 8 10 : B o N - MJ dORM (r = 0.345) 0.5 1.0 4 6 8 10 12 dPRM (r = 0.048) 0 1 2 3 5 10 15 20 gORM (r = 0.082) 0 5 10 0 4 8 12 gPRM (r = -0.625) Wasserstein Distance …
Figure 15
Figure 15. Figure 15: Per-domain BoN improvement over majority voting vs. Wasserstein distance on MMLU-Pro. Length shift hurts gPRM. We further analyze why gPRM degrades in the multi-domain setting, despite its robustness to label noise in math. As CoTs lengthen, aligning stepwise verifica…
Figure 16
Figure 16. Figure 16: User prompt format for generating CoTs on GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021). [user] Given the following question and candidate answers, choose the best answer. [Question] {question #1} [/user] [assistant] {assistant response #1} [/assistant]…
Figure 17
Figure 17. Figure 17: User prompt format for generating CoTs on MMLU-Pro (Wang et al., 2024c) proposed by Zeng et al. (2025) 20 [PITH_FULL_IMAGE:figures/full_fig_p020_17.png]
Figure 18
Figure 18. Figure 18: System prompt format for auto-labeling process labels on MMLU-Pro (Wang et al., 2024c) proposed by Zeng et al. (2025) [user] The following is a multiple choice question and its ground truth answer. You are also given a student’s solution (split into steps, enclosed wi…
Figure 19
Figure 19. Figure 19: User prompt format for auto-labeling process labels on MMLU-Pro (Wang et al., 2024c) pro￾posed by Zeng et al. (2025) 21 [PITH_FULL_IMAGE:figures/full_fig_p021_19.png]
Figure 20
Figure 20. Figure 20: Prompt format of gORM (Zhang et al., 2025a). We use this format for both generating synthetic verification-CoTs and training/evaluation of gORM. [user] You are given a {category} problem and a proposed multiple-step solution (with a step on each line): [{Category} Pro…
Figure 21
Figure 21. Figure 21: Prompt format for generating verification-CoTs for gPRM following Khalifa et al. (2025). [user] You are given a {category} problem and a proposed step-by-step solution: [{category} Problem] {problem} [Solution] {solution} Review and critique each step in the proposed …
Figure 22
Figure 22. Figure 22: Prompt format of gPRM for training and evaluation following Khalifa et al. (2025). C DATASET In this section, we provide more details on the datasets used in this paper. Math Datasets. For the math domain, we use the widely adopted PRM800K (Lightman et al., 2024) for …
Figure 23
Figure 23. Figure 23: Training example of gORM on the law domain of MMLU-pro (Wang et al., 2024c). 25 [PITH_FULL_IMAGE:figures/full_fig_p025_23.png]
Figure 24
Figure 24. Figure 24: Training example of gPRM on the law domain of MMLU-pro (Wang et al., 2024c). F ADDITIONAL RESULTS In this section, we present additional results as follows: • [PITH_FULL_IMAGE:figures/full_fig_p026_24.png]
Figure 25
Figure 25. Figure 25: Weighted majority voting results using Llama-3.1-8B-Instruct on MMLU-Pro with R1-distill￾Llama-8B backbone for reward models. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_25.png]
Figure 26
Figure 26. Figure 26: Best-of-N performance using Llama-3.1-8B-Instruct when trained and evaluated on each domain of MMLU-Pro with R1-distilled-Llama-8B backbone for reward models. 54 60 66 72 Overall 36 42 48 Law 60 64 68 72 Psychology 50 60 70 Chemistry 72 78 84 Biology 50 60 70 80 Physi…
Figure 27
Figure 27. Figure 27: Weighted majority voting performance using Llama-3.1-8B-Instruct when trained and evalu￾ated on each domain of MMLU-Pro with R1-distilled-Llama-8B backbone for reward models. 50 55 60 65 Overall 35 40 45 Law 60 64 68 Psychology 50 60 70 Chemistry 70 75 80 85 Biology 5…
Figure 28
Figure 28. Figure 28: Best-of-N results using Llama-3.1-8B-Instruct on MMLU-Pro (Wang et al., 2024c) with R1- distill-Llama-8B backbone for reward models. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_28.png]
Figure 29
Figure 29. Figure 29: Weighted majority voting results using Llama-3.1-8B-Instruct on MMLU-Pro (Wang et al., 2024c) with R1-distill-Llama-8B backbone for reward models. 40 48 56 64 Overall 20 25 30 35 Law 48 56 64 Psychology 50 60 70 Chemistry 64 72 80 88 Biology 50 60 70 Physics 36 42 48 …
Figure 30
Figure 30. Figure 30: Best-of-N results using SmolLM3-3B on MMLU-Pro with R1-distill-Llama-14B backbone for reward models. 40 48 56 64 Overall 20 25 30 35 Law 48 56 64 Psychology 50 60 70 Chemistry 64 72 80 88 Biology 50 60 70 Physics 36 42 48 54 History 48 56 64 72 Economics 60 70 80 Math…
Figure 31
Figure 31. Figure 31: Weighted majority voting results using SmolLM3-3B on MMLU-Pro with R1-distill-Llama-14B backbone for reward models. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_31.png]
Figure 32
Figure 32. Figure 32: Best-of-N results using Qwen2.5-7B-Instruct on MMLU-Pro with R1-distilled-Llama-14B back￾bone for reward models. 56 60 64 68 Overall 30 36 42 48 Law 63 66 69 72 Psychology 60 66 72 Chemistry 76 80 84 88 Biology 64 72 80 Physics 48 51 54 57 History 64 68 72 76 Economic…
Figure 33
Figure 33. Figure 33: Weighted majority voting results using Qwen2.5-7B-Instruct on MMLU-Pro with R1-distill￾Llama-14B backbone for reward models. 52 56 60 64 Overall 36 40 44 48 Law 57 60 63 66 Psychology 48 54 60 66 Chemistry 81 84 87 90 Biology 48 56 64 72 Physics 51 54 57 History 64 68…
Figure 34
Figure 34. Figure 34: Best-of-N results using gemma-2-9b-it on MMLU-Pro with R1-distill-Llama-14B backbone for reward models. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_34.png]
Figure 35
Figure 35. Figure 35: Weighted majority voting results using gemma-2-9b-it on MMLU-Pro with R1-distill-Llama￾14B backbone for reward models. 64 68 72 76 Overall 48 52 56 60 Law 72 75 78 81 Psychology 64 72 80 Chemistry 84 88 92 Biology 72 80 88 Physics 63 66 69 History 75 78 81 84 Economic…
Figure 36
Figure 36. Figure 36: Best-of-N results using Llama-3.1-70B-Instruct on MMLU-Pro with R1-distill-Llama-14B back￾bone for reward models. 64 68 72 76 Overall 48 52 56 60 Law 72 75 78 81 Psychology 64 72 80 Chemistry 81 84 87 90 Biology 66 72 78 84 Physics 66 69 72 History 76 80 84 Economics …
Figure 37
Figure 37. Figure 37: Weighted majority voting results using Llama-3.1-70B-Instruct on MMLU-Pro with R1-distill￾Llama-14B backbone for reward models. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_37.png]
Figure 38
Figure 38. Figure 38: Majority voting results of Llama-3.1-8B-Instruct vs. CoT length on MMLU-pro. 45 60 75 90 67.4 50.2 81.1 47.1 1.5B - Overall (15.3% = 521/3400) 80 85 90 95 92.3 83.3 86.3 92.3 1.5B - GSM8K (1.8% = 7/400) 40 60 80 57.4 38.9 85.1 34.6 1.5B - Math (9.4% = 94/1000) 40 60 8…
Figure 39
Figure 39. Figure 39: Performance on “aha” CoTs in ProcessBench using R1-distill-Qwen-1.5B and -7B. 62.5 65.0 67.5 70.0 72.5 75.0 77.5 80.0 65 70 75 80 85 10% 20% 30% 40% 50% 10% 20% 30% 40% 50% Process-noise ratio (%) 1.5B dPRM (Range: 67.23-81.20) 10% 20% 30% 40% 50% 1.5B gPRM (Range: 77…
Figure 40
Figure 40. Figure 40: Outcome-verification results of PRMs vs. label noise on GSM8K. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_40.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 1 linked inside Pith

  1. [1]

    (Variance floors)Var(ξ t |x)≥σ 2,Var(ξ (g) t |x)≥σ 2 +τ 2

  2. [2]

    (Weak anti-correlation) P 1≤s<t≤T Cov(ξs,ξt |x)≥ −γT, P 1≤s<t≤T Cov(ξ(g) s ,ξ(g) t |x)≥ −γT

  3. [3]

    ForgPRMwithsampledverification CoTs, sampling contributes per-step noise:Var(ξ (g) t |x)≥ σ2 +τ 2 for someτ 2 >0

    (Positive slope)σ 2 >2γ. ForgPRMwithsampledverification CoTs, sampling contributes per-step noise:Var(ξ (g) t |x)≥ σ2 +τ 2 for someτ 2 >0. For ORMs, assumeVar( ¯ξ|x)≤τ 2 orm <∞(noT-dependence). Theorem A.1(Log-error lower bound ofdPRM).Let∆ dPRM := log ˆfdPRM(x)−ζ(x). Under the assumptions above, E ∆2 dPRM ≥(σ 2 −2γ)T. Theorem A.2(Log-error bound ofdORMor...

  4. [4]

    Finally, Jensen’s inequality gives(max{0,E[∆ mean]})2 ≤E[∆ 2 mean], so the MSE bound follows

    Expectations and MSE.Taking expectations overXand applying Cauchy–Schwarz to E[B(g)(X)]yields E[∆mean]≥ κ 2 (σ2 +τ 2 −2γ)T+ (σ 2 A −2γ A) − q E[B(g)(X) 2]. Finally, Jensen’s inequality gives(max{0,E[∆ mean]})2 ≤E[∆ 2 mean], so the MSE bound follows. In the calibrated caseB (g) ≡0, the stated simplified bounds hold. B PROMPTS In this section, we present pr...

  5. [5]

    URLhttps://aclanthology.org/2025

    doi: 10.18653/v1/2025.findings-acl.547. URLhttps://aclanthology.org/2025. findings-acl.547/. Jian Zhao, Runze Liu, Kaiyan Zhang, Zhimu Zhou, Junqi Gao, Dong Li, Jiafei Lyu, Zhouyi Qian, Biqing Qi, Xiu Li, et al. GenPRM: Scaling test-time compute of process reward models via generative reasoning.arXiv preprint arXiv:2504.00891, 2025. Chujie Zheng, Zhenru Z...

  6. [6]

    SinceK X (0) = 0andK ′ X (0) =E[L|X], we obtain logµ(X) =E[L|X] + Z 1 0 (1−θ) Var θ(L|X)dθ

    Jensen-gap identity.Taylor with integral remainder atθ= 0gives KX (1) =K X (0) +K ′ X (0) + Z 1 0 (1−θ)K ′′ X (θ)dθ. SinceK X (0) = 0andK ′ X (0) =E[L|X], we obtain logµ(X) =E[L|X] + Z 1 0 (1−θ) Var θ(L|X)dθ. By definition of the mean predictor, ∆mean(X) = logµ(X)−ζ A(X),whereµ(X) =E[e L |X]. Plugginglogµ(X) = ∆ mean(X) +ζ A(X)withB (g)(X) :=E[L|X]−ζ A(X)...

  7. [7]

    Moreover, sinceL=ζ A + ∆g-prm =ζ A +B (g) +N (g) withE[N (g) |X] = 0, and sinceζ A and B(g)(X)are constants when conditioning onX, we have Var(L|X) = Var(N (g) |X)

    Lower bound onδ J and variance linkage.By tilt-stability, δJ (X)≥ κ 2 Var(L|X). Moreover, sinceL=ζ A + ∆g-prm =ζ A +B (g) +N (g) withE[N (g) |X] = 0, and sinceζ A and B(g)(X)are constants when conditioning onX, we have Var(L|X) = Var(N (g) |X). Expanding and using the variance floors and weak anti-correlation conditions (as in Theorem A.3), Var(N(g) |X)≥σ...

  8. [8]

    •Verifiable: The step can be verified using common knowledge, simple calculations, or a quick reference (e.g., recalling a basic theorem)

    GOOD StepA step is classified as GOOD if it meets all of these criteria: •Correct: Everything stated is accurate and aligns with known principles or the given problem. •Verifiable: The step can be verified using common knowledge, simple calculations, or a quick reference (e.g., recalling a basic theorem). If verifying requires extensive effort (e.g., deta...

Show all 14 references
  1. [9]

    Good job!

    OK StepA step is classified as OK if it is: •Correct and Verifiable: Contains no errors and can be verified. •Unnecessary or Redundant: Adds little value, such as restating prior information or provid- ing basic encouragement (e.g., “Good job!”). •Partially Progressing: Makes ...

  2. [10]

    •Is Hard to Verify: Requires significant effort to confirm due to poor explanation

    BAD StepA step is classified as BAD if it: •Is Incorrect: Contains factual errors, misapplies concepts, derives an incorrect result, or con- tradicts the ground truth answer. •Is Hard to Verify: Requires significant effort to confirm due to poor explanation. •Is Off-Topic: Inc...

  3. [11]

    A Multiple Choice Question

  4. [12]

    A Ground Truth Answer

  5. [13]

    Verification: Is the answer correct (Yes/No)? X

    A Student’s Step-by-Step Solution, where each step is enclosed with tags and indexed from 0. Once you identify a BAD step, return the index of the earliest BAD step. Otherwise, return the index of -1 (which denotes all steps are GOOD or OK). Please put your final answer (i.e.,...

  6. [2024]

    Verification: Is the answer correct (Yes/No)? X

    for training, where the process labelsz 1:T are human-annotated. For training ORMs, we set the outcome labely=1 z1:T =1 T (rather thany=1(ˆa(r T ) =a)), since PRM800K provides high-quality ground-truth process labels. As a testbed, we useProcessBench(Zheng 22 Preprint et al., ...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.