Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Beyond Solvability: Task Learnability as a Static Prior for LLM RL Post-Training

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A task's response to continued RL training is a stable, measurable property — learnability — distinct from current solvability, and a short probe run can estimate it well enough to improve data selection.

desk verdict Learnability is distinct from solvability and the within-bin analysis supports that, but the probe estimator's reliability for delta and the true compute cost are both under-reported. read the letter →

arxiv 2608.09217 v1 pith:KXFO2X6R submitted 2026-08-10 cs.LG cs.AI

Ting Zhou , Zhenqing Ling , Daoyuan Chen , Qianli Shen , Yilun Huang , Ying Shen , Yaliang Li This is my paper · ORCID
classification cs.LGcs.AI
keywords tasklearnabilityreinforcementlearningpost-trainingschedulingdataefficiencyLLMreasoningprobe-basedestimationrewardtrajectoriestrainingdynamics
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

Reinforcement-learning post-training for large language models usually samples tasks uniformly or weights them by how well the current policy already solves them. This paper argues that those signals miss a separate axis: how positively a task responds to further optimization under a fixed training regime, which it names learnability. It reports that per-task reward trajectories under GRPO, a group-relative policy-gradient RL method, are heterogeneous even among tasks with matched initial pass rates, that this heterogeneity reproduces across independently sampled training contexts, and that selecting or weighting tasks by it improves data efficiency on math and logic benchmarks. To make the signal usable before training, the paper introduces TrajVal, which learns a static per-task score from a short probe run on about 3% of the pool plus two inference passes. The practical claim is that a cheap pre-training estimate of learnability yields faster convergence and higher accuracy than uniform sampling, and complements online schedulers.

What carries the argument

The load-bearing object is the TrajVal score s(x) = (1 − \bar{r}_early(x)) · max(0, \bar{r}_late(x) − \bar{r}_early(x)), built from two endpoint reward estimates per task. The first factor encodes learning headroom; the second encodes directional alignment of the reward trajectory. The estimator rests on two reductions: dense per-task reward trajectories are summarized by mean early-window and late-window accuracies, and full-pool training is replaced by a 20-epoch GRPO probe run on a uniform 512-task subset whose resulting model serves as a stand-in for a later training state. The resulting static weights can be used alone or composed multiplicatively with an online scheduler's dynamic weights as w_final = w_TrajVal · w_online, which the paper argues supplies the stable offline prior that online methods lack early in training.

What would settle it

Train the same model and schedule using TrajVal weights derived from a probe run of the same length as the actual main run (1 epoch on math, 3 on logic) and compare against weights from the full 20-epoch probe; if the short-horizon weights no longer beat uniform sampling, the claimed predictive ordering is an artifact of probe horizon. Alternatively, measure the actual per-task reward gain during the main run's budget and check whether its rank correlation with probe-predicted scores is high; a low correlation would refute the proxy assumption.

Watch

Extended reading notes

Core claim

The paper's central claim is that learnability is a real and usable dimension of task value in LLM RL post-training: tasks with the same current solvability can differ sharply in how their rewards evolve under continued training, and this difference predicts downstream utility. In a controlled within-pre-bin analysis, after partitioning tasks into terciles by initial accuracy and matching average initial accuracy, the high-Δ subpools outperform low-Δ subpools on both Best Accuracy and AUC across all six domain-bin combinations. The paper further claims the signal is context-robust: agreement across independently sampled training sets reaches Cohen's κ = 0.879 for the binary learnable/non-learnable split, and a probe-trained model obtained from 20 epochs of RL on a 512-task subset preserves per-task endpoint ordering with Spearman ρ = 0.876 for late rewards. TrajVal operationalizes this by scoring each task as s(x) = (1 − \bar{r}_early(x)) · max(0, \bar{r}_late(x) − \bar{r}_early(x)), then converting scores to sampling weights with power compression and a probability floor. If correct, the result reframes task valuation for RL post-training: instead of asking only what the policy can already solve, practitioners should ask which tasks are poised to improve.

Load-bearing premise

The 20-epoch probe run on a 512-task subset is assumed to reproduce the training dynamics of the main run, which uses the full math or logic pool for only 1 or 3 epochs; if the probe's per-task ordering does not match the main budget's actual response, TrajVal's weights mis-rank tasks under the real training schedule.

Editorial extensions

If this is right

  • As a standalone replacement for uniform sampling, TrajVal improves Best Accuracy and AUC in all four domain–scale configurations tested, with steps-to-baseline as low as 40% on Logic at 4B scale.
  • Multiplying TrajVal weights into the BOTS and GRESO online schedulers improves both accuracy metrics and steps-to-baseline in every reported configuration, indicating the learnability prior is complementary to online solvability estimates.
  • Training only on tasks labeled learnable, about 41% of the pool, matches the full pool's converged accuracy in 3.6× fewer GRPO steps.
  • Within each initial-solvability tercile, high-Δ tasks outperform low-Δ tasks with matched average initial accuracy, so the predictive signal is not explained by current pass rate alone.
  • The learnability score's ranking stabilizes by probe epoch 14 of 20, so the probe horizon can be shortened without materially changing the task ordering.

Reading between the lines

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

  • A natural testable extension is to shorten the probe to roughly 70% of the nominal horizon and verify that the convergence gains persist, since the reported rank correlation reaches 0.85 by epoch 14.
  • The cross-regime reuse experiments suggest the probe cost can be amortized: a prior estimated on a smaller same-family model transferred to a larger one retained most of the in-regime gain, which if robust would make learnability priors a shareable asset across a model family.
  • Because learnability is defined as regime-conditional, recomputing the same score under different reward functions or base models would test how much of the signal is task-intrinsic versus regimen-dependent.
  • If the endpoint-pair estimator transfers to non-binary or model-graded rewards, the same lightweight recipe could apply to open-ended generation, although the paper leaves that case unverified.
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

4 major / 5 minor

Summary. The paper proposes task learnability as a regime-conditional property of how positively a task's reward responds to continued RL post-training, distinct from snapshot solvability. It reports empirical evidence from Qwen3-1.7B/4B and Llama-3.2-3B on math and logic that per-task reward trajectories are heterogeneous, reproducible across independently sampled probe contexts (Cohen's kappa = 0.776 for the six-profile taxonomy and 0.879 for the binary learnable grouping), and predictive of downstream accuracy. It then introduces TRAJVAL, a static prior estimated from a short probe RL run on a subset of the pool, using early/late endpoint rewards to score tasks as s(x) = (1 - r_early(x)) * max(0, Delta(x)). Experiments show TRAJVAL improves over uniform sampling and online schedulers, and a within-pre-bin analysis indicates that Delta carries information beyond initial solvability.

Significance. If the central claims hold, the paper adds a useful dimension to LLM RL data allocation: rather than filtering only by current pass rate, one can prioritize tasks with high expected positive response to training. The oracle-based within-pre-bin analysis and the cross-context reproducibility measurements are genuinely informative, and the code release is a plus. However, the practical and conceptual case rests on an unvalidated link: the probe-based Delta and score are never checked against oracle values, and the probe cost is not compared with the main-run compute budget. These gaps need to be closed before the headline efficiency claims are fully supported.

major comments (4)
  1. [§4.1, Eq. (1)–(2); Fig. 9] The paper reports Spearman correlations of 0.940 for r_early and 0.876 for r_late between probe-derived and oracle endpoint estimates, but never reports the analogous correlation for Delta(x) or the TRAJVAL score s(x). Because Delta is a difference of two noisy estimates, its reliability can be far lower than that of either endpoint, and s(x) is a nonlinear function of both. Figure 9 only shows agreement of intermediate probe scores with the same probe's final score, which is self-consistency and does not validate the estimator against the oracle. Please report the probe-versus-oracle rank correlation for Delta and for s, and ideally show how sampling weights change when they are recomputed from probe-estimated rather than oracle Delta.
  2. [§4.1, Reduction II; Table 4] The substitution of a 20-epoch probe on a 512-task subset for the main run is not validated against the main run's actual training budget. The main runs use 1 epoch over 17,398 tasks for Math and 3 epochs over 4,998 tasks for Logic, so the 20-epoch probe response may rank tasks differently from the 1-3 epoch response that the main run actually exploits. The reported endpoint correlations are computed against a 20-epoch oracle, not against epoch-1 or epoch-3 responses. Please provide evidence that the probe-derived ordering transfers to the short main-run horizon, or state more cautiously what the probe is a proxy for.
  3. [§4.1; Appendix B.5; Table 4] The 'lightweight' characterization of TRAJVAL is not supported by the cost accounting. The probe runs T=20 epochs on |P|=512, while the main runs consume 1 epoch over 17,398 tasks (Math) or 3 epochs over 4,998 tasks (Logic). In task-epoch units the probe is roughly 10,240/17,398 ~= 59% of the Math main-run compute and 10,240/14,994 ~= 68% of the Logic main-run compute. Appendix B.5 measures only the per-step selector overhead (0.01% of pipeline time) and omits this dominant probe training cost. Please report end-to-end compute including the probe and show that total-compute efficiency, not just steps-to-baseline, improves.
  4. [§5.3, Table 2; Table 3] The within-pre-bin analysis validates oracle learnability, not the TRAJVAL estimator. The high-Delta and low-Delta subpools in Table 2 are formed from full 20-epoch oracle trajectories, so the result that Delta predicts utility beyond matched pre is about the oracle quantity. It does not show that the probe-derived TRAJVAL score preserves the same signal. Given that Table 3 shows pre-only reaches 0.4084 Best Acc on Math versus 0.4105 for full TRAJVAL, the marginal contribution of the probe-derived Delta term to the headline gains remains unclear. Please repeat the within-pre-bin comparison using probe-estimated Delta, and add an analysis of how much of the Table 1 gain survives when the headroom term is removed.
minor comments (5)
  1. [Figure 6] The pie chart in Figure 6(c) omits the Forgetting category in its legend, although this category appears in the profile taxonomy and in the other panels; please make the legends consistent.
  2. [Appendix A.2.2 and §4.1] The symbol T is overloaded: in the S2B definition it denotes the baseline's Best Acc, while in Section 4 it denotes the probe horizon. Please rename one of them.
  3. [Eq. (1) and surrounding text] Equation (1) uses acc(x, theta_t), but the text frequently refers to mean reward; please clarify whether the endpoint statistics are per-prompt binary accuracy, group-mean reward, or a normalized variant.
  4. [§5.4, 'Diagnostic analyses'] The statement that 'the score ranking stabilizes early' relies on Figure 9, which measures agreement between intermediate probe scores and the same probe's final score; this is self-consistency and should be phrased as such rather than as validation against the oracle.
  5. [Abstract and §3.2] The abstract and Section 3.2 state that learnability is 'predictive of downstream utility' before the probe estimator is introduced; consider separating the oracle-level claim from the claim that TRAJVAL approximates oracle learnability, to avoid conflating the two.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TRAJVAL is validated on external held-out benchmarks, and the self-citations are baselines or infrastructure, not load-bearing premises.

full rationale

The paper's derivation chain is not circular. Learnability is explicitly defined as a regime-conditional response to continued training, and TRAJVAL estimates it from a short probe run and endpoint evaluations (Eqs. 1-3). The main claims—cross-context reproducibility, predictive value for downstream utility, and data-efficiency gains—are each tested against evidence that is not used to construct the estimator: Cohen's kappa on shared tasks across independently sampled probe sets, held-out benchmarks (AIME24/25, MATH500, Minerva, OlympiadBench, ARC-AGI, Ordering Puzzles), and matched pre-bin comparisons. The within-pre-bin analysis in Section 5.3 deliberately uses oracle delta values, but its downstream outcome is held-out benchmark accuracy, so the evaluation loop is broken rather than self-referential. The omission of a reported correlation for the probe-derived delta is a real statistical validity concern, but it is a reliability gap, not a circular reduction: the paper does not define the prediction in terms of the fitted input. Self-citations to BOTS and Trinity-RFT are used as a baseline and an implementation framework, respectively, and no load-bearing argument reduces to those citations. No uniqueness theorem is imported, no fitted parameter is relabeled as a prediction, and the score formula is presented as a heuristic rather than as a consequence of the definition. The limitations section also explicitly acknowledges the scope of the claims. Overall, the central derivation is self-contained and externally falsifiable.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The ledger shows that TRAJVAL rests on several empirical assumptions rather than fitted numeric parameters. The main free parameters (alpha, epsilon, |P|, T, K) are manually chosen and lightly ablated. The costliest implicit choice is the probe horizon T=20, which is not matched to the main-run epochs and drives the compute mismatch reported here.

free parameters (5)
  • Power compression exponent alpha = 0.3
    Default chosen and tested over {0.1, 0.3, 0.5}; a hand-tuned hyperparameter in Equation 3, not fitted to a validation set.
  • Probability floor epsilon = 0.05
    Smoothing floor in Equation 3; chosen by hand to keep full-pool coverage.
  • Probe set size |P| = 512
    Roughly 3 percent of the 17,398-task pool; tested at 256 and 1024, with 512 giving the best trade-off.
  • Probe horizon T = 20 epochs
    Used in the diagnostic study and TRAJVAL runs (Section B.4, Figure 9). It is not aligned with the main-run budget of 1 epoch (math) or 3 epochs (logic), which is the source of the compute mismatch flagged in this report.
  • Endpoint window size K = 1
    Evaluated in the economical K=1 setting; larger K reduces rollout variance at higher inference cost.
assumptions (5)
  • domain assumption A uniformly sampled probe set P of size 512 preserves the learnability composition of the full pool D.
    Invoked in Reduction II (Section 4.1) and supported by profile-distribution comparisons in Appendix B.2 for one 2,048-task diagnostic pool; not tested across pools or regimes.
  • ad hoc to paper The multiplicative score s(x) = (1 - pre(x)) * max(0, Delta(x)) captures the learnability ordering relevant for data allocation.
    Equation 2 is a hand-designed operationalization, not derived from first principles. Ablations support it empirically, but the functional form is a free modeling choice.
  • domain assumption Per-task reward trajectories over 20 epochs in a small pool generalize to the main run with a different pool size and training budget.
    The probe horizon (20 epochs) is longer than the main-run budget (1 epoch math, 3 epochs logic); the paper verifies endpoint correlations against a 20-epoch full-pool oracle but not against a 1-epoch full-pool oracle.
  • domain assumption Binary exact-match rewards in math and logic provide a suitable signal for measuring learnability.
    Stated in the Limitations: transfer to non-binary or model-graded rewards is open.
  • standard math Standard GRPO setup and the chosen hyperparameters (n=16 rollouts, batch 64, etc.) define the regime under which learnability is measured.
    The method is regime-conditional by definition; results could shift under different RL algorithms or rollout settings.
invented entities (1)
  • Learnability score s(x) independent evidence
    purpose: Task-level sampling prior for RL post-training, capturing expected positive response to continued training.
    Operationalized as (1 - pre(x)) * max(0, Delta(x)); validated through within-pre-bin predictive gaps and held-out benchmark improvements, providing an out-of-sample handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Solvability: Task Learnability as a Static Prior for LLM RL Post-Training." pith.science (2026). https://pith.science/paper/KXFO2X6R

@misc{pith2026260809217,
  author       = {Pith},
  title        = {Pith review of: Beyond Solvability: Task Learnability as a Static Prior for LLM RL Post-Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXFO2X6R}},
  note         = {Machine review of arXiv:2608.09217}
}
read the original abstract

Reinforcement learning (RL) has become a central post-training paradigm for eliciting reasoning capabilities in large language models, yet uniform task sampling allocates compute without regard to differences in how tasks respond to optimization. Existing task-valuation methods mostly rely on snapshot-based signals such as current pass rate or reward, which estimate how solvable a task is under the current policy. However, tasks with similar current solvability can still differ substantially in how positively they respond to further training. We study this residual axis as task learnability: a regime-conditional measure of expected positive response to continued training under a fixed RL post-training regime. By analyzing per-task reward trajectories, we find that learnability is reproducible across independently sampled training contexts and predictive of downstream utility. To make this signal practical before training begins, we propose TrajVal, a lightweight probe-based estimator that approximates per-task learnability from a short probe run and two endpoint evaluations. TrajVal can be used either as a standalone static prior for task sampling or as a multiplicative prior for existing online schedulers. Experiments on mathematical and logical reasoning benchmarks across multiple model scales show that TrajVal improves data efficiency over uniform sampling and provides complementary gains when combined with online scheduling methods.

Figures

Figures reproduced from arXiv: 2608.09217 by the authors.

Figure 1
Figure 1. Representative per-task reward trajectories from the diagnostic pool [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of TRAJVAL. Top: per-task reward trajectories reveal heterogeneous training response beyond initial solvability. Bottom: TRAJVAL approximates this response using two reductions: from dense trajectories to an endpoint-based summary, and from full-pool training to a lightweight probe run. The resulting score is used as a static prior for prior-weighted sampling in RL training, either alone or in combination w… view at source ↗
Figure 3
Figure 3. Effect of learnability-based data allocation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: All tasks in D projected onto the slope–R2 space of their training trajectories. Colors indicate pro￾file labels derived from the endpoint pair (¯rearly, r¯late). a much cheaper probe training run on a small uni￾formly sampled subset. Together, these reductions yield a…
Figure 5
Figure 5. Figure 5: Within-pre-bin evaluation accuracy curves on [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qwen3-1.7B: Stability of learnability profiles [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 8
Figure 8. Figure 8: Llama-3.2-3B: All tasks in D projected onto the slope–R2 space of their reward trajectories, col￾ored by profile labels derived from the endpoint pair (¯rearly, r¯late). The same broad structural separation ob￾served in [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Spearman rank correlation between sk(x) and sT (x) as a function of probe epoch k, evaluated on D. Two independently sampled probe sets are shown. The ordering stabilizes (ρ ≥ 0.85) by k = 14, well before the full probe horizon T = 20. |D| = 17,398 tasks, corresponding…
Figure 10
Figure 10. Figure 10: Training-time rollout accuracy under differ [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Evaluation accuracy curves for all main configurations on Math and Logic. The x-axis shows training [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 6 canonical work pages

  1. [1]

    arXiv preprint arXiv:2503.14476 , year=

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

    arXiv

  2. [2]

    arXiv preprint arXiv:2506.14965 , year=

    Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective , author=. arXiv preprint arXiv:2506.14965 , year=

    arXiv

  3. [3]

    arXiv preprint arXiv:2412.04604 , year=

    Arc prize 2024: Technical report , author=. arXiv preprint arXiv:2412.04604 , year=

    arXiv 2024

  4. [4]

    arXiv preprint arXiv:2505.11831 , year=

    Arc-agi-2: A new challenge for frontier ai reasoning systems , author=. arXiv preprint arXiv:2505.11831 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2411.02272 , year=

    Combining induction and transduction for abstract reasoning , author=. arXiv preprint arXiv:2411.02272 , year=

    arXiv

  6. [6]

    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=

    arXiv

  7. [7]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    arXiv

  8. [8]

    Advances in neural information processing systems , volume=

    Solving quantitative reasoning problems with language models , author=. Advances in neural information processing systems , volume=

Show all 36 references
  1. [9]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  2. [10]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  3. [11]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Dataset cartography: Mapping and diagnosing datasets with training dynamics , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  4. [12]

    Advances in neural information processing systems , volume=

    Deep learning on a data diet: Finding important examples early in training , author=. Advances in neural information processing systems , volume=

  5. [13]

    arXiv preprint arXiv:2507.17512 , year=

    Can one domain help others? a data-centric study on multi-domain reasoning via reinforcement learning , author=. arXiv preprint arXiv:2507.17512 , year=

    arXiv

  6. [14]

    arXiv preprint arXiv:2505.17826 , year=

    Trinity-rft: A general-purpose and unified framework for reinforcement fine-tuning of large language models , author=. arXiv preprint arXiv:2505.17826 , year=

  7. [15]

    arXiv preprint arXiv:2304.11277 , year=

    Pytorch fsdp: experiences on scaling fully sharded data parallel , author=. arXiv preprint arXiv:2304.11277 , year=

    arXiv

  8. [16]

    Proceedings of the 29th symposium on operating systems principles , pages=

    Efficient memory management for large language model serving with pagedattention , author=. Proceedings of the 29th symposium on operating systems principles , pages=

  9. [17]

    Advances in neural information processing systems , volume=

    Flashattention: Fast and memory-efficient exact attention with io-awareness , author=. Advances in neural information processing systems , volume=

  10. [18]

    Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo , title =. CoRR , volume =

  11. [19]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  12. [20]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    arXiv

  13. [21]

    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=

  14. [22]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  15. [23]

    The Thirteenth International Conference on Learning Representations , year=

    DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search , author=. The Thirteenth International Conference on Learning Representations , year=

  16. [24]

    Advances in Neural Information Processing Systems , volume=

    Lima: Less is more for alignment , author=. Advances in Neural Information Processing Systems , volume=

  17. [25]

    Second Conference on Language Modeling , year=

    LIMO: Less is More for Reasoning , author=. Second Conference on Language Modeling , year=

  18. [26]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    s1: Simple test-time scaling , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  19. [27]

    arXiv preprint arXiv:2502.11886 , year=

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

    arXiv

  20. [28]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2412.16720 , year=

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

    arXiv

  22. [30]

    arXiv preprint arXiv:2510.01135 , year=

    Prompt curriculum learning for efficient llm post-training , author=. arXiv preprint arXiv:2510.01135 , year=

  23. [31]

    arXiv preprint arXiv:2603.25184 , year=

    Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model , author=. arXiv preprint arXiv:2603.25184 , year=

    arXiv

  24. [32]

    Training , volume=

    Scale down to speed up: Dynamic data selection for reinforcement learning , author=. Training , volume=

  25. [33]

    arXiv preprint arXiv:2506.06632 , year=

    Curriculum reinforcement learning from easy to hard tasks improves LLM reasoning , author=. arXiv preprint arXiv:2506.06632 , year=

  26. [34]

    Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Can prompt difficulty be online predicted for accelerating rl finetuning of reasoning models? , author=. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

  27. [35]

    arXiv preprint arXiv:2505.14970 , year=

    Self-evolving curriculum for llm reasoning , author=. arXiv preprint arXiv:2505.14970 , year=

  28. [36]

    arXiv preprint arXiv:2602.20532 , year=

    Actor-Curator: Co-adaptive Curriculum Learning via Policy-Improvement Bandits for RL Post-Training , author=. arXiv preprint arXiv:2602.20532 , year=

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.