Pith. sign in

REVIEW 2 major objections 5 minor 49 references

Tracking prompt difficulty with a Kalman filter lets RL finetuning match expensive oracles at a fraction of the rollout cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 01:28 UTC pith:LP4YQRV6

load-bearing objection Clean systems paper: logit-space Kalman with policy-coupled process noise beats stationary Beta trackers on rollout efficiency; the causal story for the coupling is only thinly ablated. the 2 major comments →

arxiv 2607.27610 v1 pith:LP4YQRV6 submitted 2026-07-30 cs.LG

Kalman Meets Curriculum: Efficient Dynamic Prompt Selection for Adaptive RL Finetuning

Haodong Zhu , Yangyang Ren , Yanjing Li , Sheng Xu , Haiguang Liu , Linlin Yang , Baochang Zhang This is my paper
classification cs.LG
keywords reinforcement learning finetuningprompt selectionKalman filtercurriculum learningnon-stationary banditsLLM reasoningrollout efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Reinforcement-learning finetuning of language models works best when training prompts sit at intermediate difficulty for the current policy, but that difficulty keeps shifting as the model improves. Evaluation-based selectors get accurate difficulty by rolling out many candidates and are therefore costly; prediction-based selectors are cheap but treat difficulty as roughly stationary and therefore drift out of date. This paper reframes the problem as online state estimation: each prompt’s success rate is a latent random walk in logit space whose process noise grows with the size of the latest policy update. A Kalman filter maintains a Gaussian posterior for every prompt from the same rollouts already used for training, and the next batch is chosen by maximizing the expected intermediate-difficulty utility under that posterior. The result is adaptive selection with no extra inference, higher final accuracy than strong prediction baselines, and accuracy that matches or slightly beats the expensive evaluation oracle while using far fewer rollouts across math, planning, and geometry tasks and several RL algorithms.

Core claim

Modeling each prompt’s logit success rate as a linear-Gaussian random walk with process noise proportional to the squared policy step, then selecting the batch that maximizes posterior-expected intermediate-difficulty utility, yields online prompt selection that is both more accurate and more rollout-efficient than stationary prediction methods and competitive with evaluation-based oracles that spend several times more rollouts.

What carries the argument

KGPS: a per-prompt Kalman filter in logit space whose process noise is Qt = γ∥θt − θt−1∥², whose observations are delta-method Gaussians of binomial success rates, and whose selection score is the posterior expectation of h(ψ) = σ(ψ)(1 − σ(ψ)).

Load-bearing premise

Prompt difficulty must evolve roughly like an independent scalar random walk whose uncertainty grows with the size of the policy parameter step, and a simple Gaussian approximation of a handful of binary rollouts must stay well-calibrated enough to rank prompts correctly.

What would settle it

Train the same models with KGPS versus MoPPS and Dynamic Sampling while logging predicted versus empirical success rates each step; if KGPS’s MAE and Spearman rank correlation do not stay clearly better, or if final accuracy under a fixed rollout budget no longer matches or beats the oracle, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • RL finetuning pipelines can drop expensive oversampling for selection and still reach oracle-level accuracy under a uniform rollout budget.
  • Any rollout-based RL algorithm (GRPO, PPO, Reinforce++) can plug in the same Kalman posterior and utility without changing its loss.
  • Uncertainty automatically revisits neglected prompts, so hand-tuned decay or forgetting schedules become unnecessary.
  • The same selection loop transfers across math, planning, and visual geometry without task-specific redesign.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Shared structure across similar prompts is left unused; a multi-prompt or embedding-coupled filter could shrink warmup cost and further cut early-training error.
  • Coupling process noise only to parameter-step size may under-react when a small weight change flips a whole skill; richer policy-change signals could tighten the filter.
  • The same state-estimation view could apply to other non-stationary curriculum problems beyond language-model RL, such as adaptive difficulty in interactive tutors.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes KGPS, an online prompt-selection method for RL finetuning of LLMs. It reframes each prompt’s latent success rate as a time-varying state in logit space, evolves it as a random walk whose process noise is Qt = γ∥θt − θt−1∥² (Eqs. 3–4), maintains a Gaussian posterior with a scalar Kalman filter (Eqs. 8–9), and selects the Top-B batch by the posterior-expected utility E[σ(ψ)(1−σ(ψ))] approximated by five-point Gauss–Hermite quadrature (Eqs. 10–12). The method uses only the rollouts already required for policy training. Experiments on math (MATH and six OOD suites), Countdown, and Geometry3k, across Qwen3 0.6B–8B, DeepSeek-R1-Distill-7B, and VL models, and under GRPO, PPO, and Reinforce++, report higher final accuracy than Uniform, GRESO, and MoPPS at the same rollout budget, and accuracy matching or slightly exceeding Dynamic Sampling (DS) while using 71–87% fewer rollouts (e.g., Table 1: DeepSeek-R1-Distill-7B, 55.56 vs DS 55.44 at 296k vs 1744k rollouts). Ablations address posterior expectation vs point estimate, warmup, and fixed vs dynamic Qt (Tables 3–4).

Significance. If the results hold, KGPS is a practically useful contribution: it closes much of the accuracy gap between cheap prediction-based selectors and expensive evaluation-based oracles without extra LLM inference, and it is algorithm-agnostic (Table 2). The state-space formulation, delta-method observation model (Appendix C), and closed-form Kalman updates are cleanly derived and inexpensive. Strengths include multi-modality/multi-algorithm evaluation, explicit MAE/Spearman calibration diagnostics (Figs. 1b, 5, B), and an honest independence limitation (Appendix H). The work is significant for sample-efficient RLVR-style finetuning even if some modeling choices remain heuristic.

major comments (2)
  1. [§3.2 Eqs. 3–4; Table 4; Appendix B] The paper’s central narrative—that gains come from dynamic state estimation with policy-coupled process noise—rests on Eqs. 3–4 and is only thinly supported. Table 4 (Qwen3-0.6B Math only) shows fixed Qt=c vs dynamic Qt=γ∥Δθ∥² (+1.21 Avg). There is no γ or P0 sensitivity, no analogous ablation on the headline DeepSeek-R1-Distill-7B or larger models, and no isolation of “variance inflation + E[h]” versus the specific ∥Δθ∥² coupling. ∥θt−θt−1∥² is unnormalized and changes scale with model size, full FT vs LoRA, and optimizer, yet App. B fixes γ=0.1 and P0=1.0 across 0.6B–8B. If Qt is systematically ≪ or ≫ Rt, the filter reduces to generic uncertainty-aware ranking already supplied by Eq. 10. Please add (i) γ/P0 sweeps on at least one larger backbone, (ii) a control with constant or EMA-style inflation (no policy norm), and (iii) a short discussion of how practitioners should set γ when ∥Δθ
  2. [§3.2–3.3 Eqs. 5–7; Appendix C] The Gaussian observation model (Eqs. 5–7, Appendix C) uses the delta method with plug-in clipped variance and k=8 binomial rollouts. For k=8 the CLT/logit-linearization is coarse near the boundaries even with δ=1/(2k). MAE/Spearman (Figs. 1b, 5, B) show KGPS beats MoPPS but do not report calibration diagnostics that would validate the Gaussian Rt assumption (e.g., innovation whiteness, coverage of predictive intervals, or sensitivity to k). A brief check that ranking quality and final accuracy remain stable for k∈{4,8,16}, or a note that the method is used primarily as a ranking device rather than a calibrated density, would strengthen the load-bearing observation model.
minor comments (5)
  1. [Fig. 1; §4.2; Table 1] Fig. 1(a) caption and abstract cite “71% fewer” / “83% fewer” rollouts vs DS; Table 1 makes the budgets clear (296k vs 1024k/1744k/2310k), but a single sentence in §4.2 defining how “fewer rollouts” is computed (candidate oversampling cost only, or total environment steps) would avoid ambiguity.
  2. [§3.3; Algorithm 1] Algorithm 1 and §3.3: the update uses observations from step t−1 while selection is for step t; this one-step lag is correct but easy to misread. A short clarifying sentence would help.
  3. [§2] Related work could more explicitly contrast KGPS with other non-stationary bandit / restless-bandit and adaptive-curriculum RL selectors beyond MoPPS/GRESO/DS, even if only briefly.
  4. [Title block; Appendix C] Typos/notation: “Kalman-GuidedPromptSelection” missing spaces in places; duplicate “where ˙∼ denotes asymptotic distribution” in Appendix C; affiliation list shows “5Communication University of China” without a matching author mark in the author block.
  5. [Appendix H; §3.2] Appendix H correctly flags independent per-prompt filters. A one-sentence forward pointer in §3.2 or the conclusion would make the limitation visible in the main text.

Circularity Check

0 steps flagged

No significant circularity: KGPS is a self-contained engineering method evaluated on held-out benchmarks; nothing reduces to its inputs by construction.

full rationale

The paper proposes a standard linear-Gaussian SSM + Kalman filter for tracking per-prompt logit success rates, with process noise Qt = γ∥Δθ∥² and selection by posterior-expected utility E[σ(ψ)(1−σ(ψ))]. These are modeling and design choices, not claimed first-principles derivations of external quantities. The utility peak at intermediate difficulty is imported from prior curriculum/RL literature (Bae et al., Chen et al., Qu et al.) as a scoring heuristic, not fitted to the reported test metrics. Filter updates use the same training rollouts already collected for the RL step; selection does not define or re-label the held-out accuracy numbers. Evaluation is on external benchmarks (MATH variants, Countdown, Geometry3k) under fixed protocols against Uniform/GRESO/MoPPS/DS. Warmup initializes means from one uniform epoch of empirical rates—data-driven init, not a circular fit-then-predict loop on the test targets. No load-bearing uniqueness theorem, no self-citation chain that forces the result, and no renaming of a known empirical law as a derived prediction. Weaknesses (scalar unnormalized Qt, independence of prompts) are assumption/correctness issues, not circularity. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard Kalman filtering math, a domain modeling choice that difficulty is a per-prompt logit random walk driven by policy step size, and a few hand-chosen scalars. No new physical entities are postulated. The training-utility peak at intermediate success rate is taken from prior RL-for-reasoning work rather than re-derived.

free parameters (4)
  • process noise scale γ = 0.1
    Multiplies ∥Δθ∥² to set Qt each step; controls how fast uncertainty inflates under policy drift. Chosen by hand, not derived.
  • initial posterior variance P0 = 1.0
    Sets starting uncertainty after the warmup logit means; affects early exploration via posterior-expected utility.
  • clipping δ for plug-in success rate = 1/(2k) with k=8
    Keeps observation noise Rt finite near 0/1; set to 1/(2k), which is a conventional but still design choice affecting Kalman gain on near-degenerate rollouts.
  • rollouts per prompt k and batch size B = k=8 (B implicit in protocol)
    Training protocol knobs that change binomial observation quality and selection granularity; fixed in experiments rather than swept exhaustively for the method claim.
axioms (5)
  • ad hoc to paper Each prompt’s latent success rate evolves as an independent random walk in logit space: ψt = ψt−1 + wt, wt ~ N(0, Qt).
    Core modeling choice in §3.2; not forced by theory of LLM policies, and Appendix H admits missing cross-prompt correlation.
  • ad hoc to paper Process noise variance is proportional to squared policy parameter update: Qt = γ∥θt−θt−1∥².
    Coupling mechanism that distinguishes KGPS from fixed-decay trackers; justified by intuition, supported by one ablation (Table 4), not derived from a policy-difficulty theorem.
  • standard math Binomial empirical success rates yield approximately Gaussian logit observations via the delta method with plug-in variance Rt.
    Appendix C; standard asymptotic approximation, strained at small k and extremes even with clipping.
  • domain assumption Training utility is maximized near intermediate success rates, scored by h(ψ)=σ(ψ)(1−σ(ψ)) (Bernoulli variance).
    Imported from cited curriculum/RL selection work (Bae et al., Chen et al., Qu et al.); selection objective in §3.4 depends on it.
  • domain assumption Linear-Gaussian Kalman filter posteriors remain sufficiently calibrated for Top-B ranking under non-stationary RL finetuning.
    Implicit throughout §3.3–3.4; empirical MAE/Spearman plots support it in the reported regimes but are not a proof of calibration.
invented entities (1)
  • KGPS (Kalman-Guided Prompt Selection) procedure no independent evidence
    purpose: Name for the end-to-end filter-plus-utility selection loop applied to RL prompt batches.
    Method branding rather than a new ontological object; built from standard SSM/Kalman pieces plus the paper’s noise-coupling and utility choices.

pith-pipeline@v1.2.0-daily-grok45 · 26354 in / 3649 out tokens · 78243 ms · 2026-07-31T01:28:45.933054+00:00 · methodology

0 comments
read the original abstract

Reinforcement learning (RL) finetuning significantly enhances the reasoning capabilities of large language models (LLMs), yet its effectiveness critically depends on selecting prompts of appropriate difficulty for the current policy. This is challenging because prompt difficulty evolves throughout training. Existing online methods therefore face a trade-off: evaluation-based approaches are accurate but expensive, while prediction-based approaches are efficient but typically assume stationary difficulty, making them ill-suited to RL's non-stationary training dynamics. To address these issues, we propose a Kalman-Guided Prompt Selection method (KGPS), which reformulates prompt selection as a dynamic state estimation problem rather than static difficulty prediction. KGPS models each prompt's latent success rate in logit space using a linear-Gaussian state-space model, with process noise coupled to the magnitude of policy updates so that uncertainty increases when the policy changes more substantially. A Kalman filter then maintains a calibrated Gaussian posterior over prompt difficulty, and prompts are selected by maximizing a posterior-expected training utility that favors intermediate-difficulty prompts while naturally revisiting uncertain ones. The resulting procedure is adaptive to policy drift and requires no additional rollouts beyond standard policy training. Extensive experiments across mathematics, planning, and geometry reasoning benchmarks, as well as multiple RL algorithms, show that KGPS consistently improves both final accuracy and rollout efficiency over strong baselines, establishing state-of-the-art performance among online prompt selection methods. For example, on DeepSeek-R1-Distill-7B, KGPS uses 83% fewer rollouts than DS while even improving the average performance by 0.12 point across six math reasoning benchmarks.

Figures

Figures reproduced from arXiv: 2607.27610 by Baochang Zhang, Haiguang Liu, Haodong Zhu, Linlin Yang, Sheng Xu, Yangyang Ren, Yanjing Li.

Figure 1
Figure 1. Figure 1: On the Math dataset using the Qwen3-0.6B model, (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework Overview of KGPS. At each training step t, KGPS maintains a Gaussian posterior N (ψˆt τ , Pt τ ) over the latent logit-space difficulty of every candidate prompt via a Kalman filter. Prompts selected in the previous step receive a full Kalman update from their rollout observations; unselected prompts undergo only the prediction step, with variance inflated by Qt to reflect policy￾induced uncertai… view at source ↗
Figure 3
Figure 3. Figure 3: Test accuracy on the Math benchmark across three model scales under different data [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Test accuracy on Countdown and Geometry benchmarks, where KGPS (ours) demonstrates [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spearman correlation (p-values) between KGPS scores and sample difficulty across training steps. High significant correlation validates KGPS as a difficulty-aware selection criterion [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 26 linked inside Pith

  1. [11]

    Geometry3K : A large-scale multi-modal geometry reasoning dataset, 2025

    Hiyouga . Geometry3K : A large-scale multi-modal geometry reasoning dataset, 2025

    2025

  2. [16]

    Lewkowycz, A

    A. Lewkowycz, A. Andreassen, D. Dohan, E. Dyer, H. Michalewski, V. Ramasesh, A. Slone, C. Anil, I. Schlag, T. Gutman-Solo, et al. Solving quantitative reasoning problems with language models. Advances in Neural Information Processing Systems, 35: 0 3843--3857, 2022

    2022

  3. [20]

    P. Lu, R. Gong, S. Jiang, L. Qiu, S. Huang, X. Liang, and S.-C. Zhu. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. In The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021), 2021

    2021

  4. [21]

    M. Luo, S. Tan, J. Wong, X. Shi, W. Tang, M. Roongta, C. Cai, J. Luo, T. Zhang, E. Li, R. A. Popa, and I. Stoica. Deepscaler: Surpassing o1-preview with a 1.5b model by scaling rl, 2025. Notion Blog

    2025

  5. [22]

    American mathematics competitions, 2023

    Mathematical Association of America . American mathematics competitions, 2023

    2023

  6. [23]

    American invitational mathematics examination, 2024

    Mathematical Association of America . American invitational mathematics examination, 2024

    2024

  7. [24]

    American invitational mathematics examination, 2025

    Mathematical Association of America . American invitational mathematics examination, 2025

    2025

  8. [25]

    F. Meng, L. Du, Z. Liu, Z. Zhou, Q. Lu, D. Fu, B. Shi, W. Wang, J. He, K. Zhang, et al. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning. CoRR, 2025

    2025

  9. [26]

    J. Pan, J. Zhang, X. Wang, L. Yuan, H. Peng, and A. Suhr. Tinyzero. https://github.com/Jiayi-Pan/TinyZero, 2025. Accessed: 2025-01-24

    2025

  10. [33]

    Sheng, C

    G. Sheng, C. Zhang, Z. Ye, X. Wu, W. Zhang, R. Zhang, Y. Peng, H. Lin, and C. Wu. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256, 2024

    Pith/arXiv arXiv 2024

  11. [43]

    arXiv preprint arXiv:2405.07863 , year=

    Rlhf workflow: From reward modeling to online rlhf , author=. arXiv preprint arXiv:2405.07863 , year=

    Pith/arXiv arXiv

  12. [44]

    arXiv preprint arXiv:2310.12773 , year=

    Safe rlhf: Safe reinforcement learning from human feedback , author=. arXiv preprint arXiv:2310.12773 , year=

    Pith/arXiv arXiv

  13. [45]

    arXiv preprint arXiv:2307.04964 , year=

    Secrets of rlhf in large language models part i: Ppo , author=. arXiv preprint arXiv:2307.04964 , year=

    Pith/arXiv arXiv

  14. [46]

    arXiv preprint arXiv:2412.16720 , year=

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    Pith/arXiv arXiv

  15. [48]

    5: Scaling reinforcement learning with llms , author=

    Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

    Pith/arXiv arXiv

  16. [49]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  17. [50]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  18. [51]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    Pith/arXiv arXiv

  19. [52]

    arXiv preprint arXiv:2501.03262 , year=

    Reinforce++: A simple and efficient approach for aligning large language models , author=. arXiv preprint arXiv:2501.03262 , year=

    Pith/arXiv arXiv

  20. [53]

    arXiv preprint arXiv:2504.05118 , year=

    Vapo: Efficient and reliable reinforcement learning for advanced reasoning tasks , author=. arXiv preprint arXiv:2504.05118 , year=

    Pith/arXiv arXiv

  21. [54]

    arXiv preprint arXiv:2410.01679 , year=

    Vineppo: Unlocking rl potential for llm reasoning through refined credit assignment , author=. arXiv preprint arXiv:2410.01679 , year=

    Pith/arXiv arXiv

  22. [55]

    2024 , journal =

    HybridFlow: A Flexible and Efficient RLHF Framework , author =. 2024 , journal =

    2024

  23. [56]

    arXiv preprint arXiv:2502.03387 , year=

    LIMO: Less is More for Reasoning , author=. arXiv preprint arXiv:2502.03387 , year=

    Pith/arXiv arXiv

  24. [57]

    arXiv preprint arXiv:2502.11886 , year=

    Limr: Less is more for rl scaling , author=. arXiv preprint arXiv:2502.11886 , year=

    Pith/arXiv arXiv

  25. [58]

    arXiv preprint arXiv:2503.24290 , year=

    Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model , author=. arXiv preprint arXiv:2503.24290 , year=

    Pith/arXiv arXiv

  26. [59]

    5-math technical report: Toward mathematical expert model via self-improvement , author=

    Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement , author=. arXiv preprint arXiv:2409.12122 , year=

    Pith/arXiv arXiv

  27. [60]

    arXiv preprint arXiv:2504.05185 , year=

    Concise reasoning via reinforcement learning , author=. arXiv preprint arXiv:2504.05185 , year=

    arXiv

  28. [61]

    arXiv preprint arXiv:2505.24864 , year=

    ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models , author=. arXiv preprint arXiv:2505.24864 , year=

    Pith/arXiv arXiv

  29. [62]

    arXiv preprint arXiv:2502.01456 , year=

    Process reinforcement through implicit rewards , author=. arXiv preprint arXiv:2502.01456 , year=

    Pith/arXiv arXiv

  30. [63]

    CoRR , year=

    Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning , author=. CoRR , year=

  31. [64]

    arXiv preprint arXiv:2504.03380 , year=

    Online difficulty filtering for reasoning oriented reinforcement learning , author=. arXiv preprint arXiv:2504.03380 , year=

    arXiv

  32. [65]

    arXiv preprint arXiv:2505.14970 , year=

    Self-Evolving Curriculum for LLM Reasoning , author=. arXiv preprint arXiv:2505.14970 , year=

    arXiv

  33. [66]

    Jiayi Pan and Junjie Zhang and Xingyao Wang and Lifan Yuan and Hao Peng and Alane Suhr , title =

  34. [67]

    arXiv preprint arXiv:2501.12948 , year=

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  35. [68]

    arXiv preprint arXiv:2503.18892 , year=

    SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild , author=. arXiv preprint arXiv:2503.18892 , year=

    Pith/arXiv arXiv

  36. [69]

    2025 , note=

    DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL , author=. 2025 , note=

    2025

  37. [70]

    arXiv preprint arXiv:2506.06632 , year=

    Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning , author=. arXiv preprint arXiv:2506.06632 , year=

    arXiv

  38. [71]

    arXiv preprint arXiv:2507.04632 , year=

    Can Prompt Difficulty be Online Predicted for Accelerating RL Finetuning of Reasoning Models? , author=. arXiv preprint arXiv:2507.04632 , year=

    arXiv

  39. [72]

    arXiv preprint arXiv:2510.26374 , year=

    BOTS: A Unified Framework for Bayesian Online Task Selection in LLM Reinforcement Finetuning , author=. arXiv preprint arXiv:2510.26374 , year=

    arXiv

  40. [73]

    arXiv preprint arXiv:2507.06167 , year=

    Skywork-R1V3 Technical Report , author=. arXiv preprint arXiv:2507.06167 , year=

    Pith/arXiv arXiv

  41. [74]

    arXiv preprint arXiv:2504.15900 , year=

    SARI: Structured Audio Reasoning via Curriculum-Guided Reinforcement Learning , author=. arXiv preprint arXiv:2504.15900 , year=

    Pith/arXiv arXiv

  42. [75]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring Mathematical Problem Solving With the MATH Dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    Pith/arXiv arXiv

  43. [76]

    arXiv preprint arXiv:2305.20050 , year=

    Let's Verify Step by Step , author=. arXiv preprint arXiv:2305.20050 , year=

    Pith/arXiv arXiv

  44. [77]

    American Invitational Mathematics Examination , year =

  45. [78]

    American Mathematics Competitions , year =

  46. [79]

    Advances in Neural Information Processing Systems , volume=

    Solving quantitative reasoning problems with language models , author=. Advances in Neural Information Processing Systems , volume=

  47. [80]

    arXiv preprint arXiv:2402.14008 , year=

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems , author=. arXiv preprint arXiv:2402.14008 , year=

    Pith/arXiv arXiv

  48. [81]

    Inter-GPS: Interpretable Geometry Problem Solving with Formal Language and Symbolic Reasoning , author =. The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021) , year =

    2021

  49. [82]

    Bartoldson and Bhavya Kailkhura and Fan Lai and Jiawei Zhao and Beidi Chen , title =

    Haizhong Zheng and Yang Zhou and Brian R. Bartoldson and Bhavya Kailkhura and Fan Lai and Jiawei Zhao and Beidi Chen , title =. arXiv preprint arXiv:2506.02177 , year =

    Pith/arXiv arXiv