Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Selective distillation on failed prompts rescues RL training from vanishing gradients.

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 ↗

RSTG selectively distills a teacher on negative zero-variance prompts, with confidence weighting, token-level selection, and auxiliary SFT, improving math and code RL post-training over naive GRPO+OPD.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful RLVR recipe with a plausible mechanism, but the headline gains sit on single runs and eval-tuned hyperparameters — worth refereeing, worth re-running before citing. the 4 major comments →

arxiv 2608.00782 v1 pith:ORQJ722U submitted 2026-08-01 cs.CL

Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance

Zhuowen Han , Jinwei Xiao , Zhengxi Lu , Renren Jin , Zhiyuan Yao , Yuxin Liu , Hongyan Hao , Yueqing Sun
show 4 more authors
Yu Yang Qi GU Xunliang Cai Deyi Xiong
This is my paper
classification cs.CL
keywords RLVRGRPOon-policy distillationzero-variance promptsnegative RL-groupsteacher confidence weightingtoken selectionLLM reasoning
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.

The reading

RSTG sets out to prove that the apparent incompatibility of group-relative policy optimization (GRPO) and on-policy distillation (OPD) is a design problem, not a fundamental one. The paper identifies three failure modes of naive GRPO+OPD — distillation applied where neither student nor teacher can help, premature collapse onto the teacher, and mostly negative token-level advantages — and counters each: distillation is confined to negative zero-variance prompts weighted by teacher confidence; only high-entropy or high-divergence tokens receive gradient updates; and an auxiliary SFT loss on teacher-correct trajectories injects positive signal. The claim is that this recovers usable learning signals exactly where GRPO's advantages are zero, and the reported gains over naive GRPO+OPD (+4.02% on math, +3.05% on code) are the evidence. A sympathetic reader would take the paper as establishing a recipe for combining RL and distillation without sacrificing exploration.

Core claim

On the paper's own terms, the central claim is that naive GRPO+OPD inherits the worst of both paradigms: sparse group-level advantages collapse to zero when all rollouts tie, dense teacher advantages are mostly negative because the teacher assigns low probability to student-generated tokens, and the student converges so quickly to the teacher that exploration and further improvement are capped. RSTG is the claim that all three problems are addressable by routing the distillation signal through the same structure that caused the silence. For prompts where every rollout fails — negative zero-variance prompts — the hybrid advantage is β·ω_i·A_OPD, with ω_i the teacher's mean@8 score as confiden

What carries the argument

The central object is the negative zero-variance prompt: a question for which all G student rollouts receive the same (typically zero or failed) reward, so GRPO's group-normalized advantage is zero for every token and the prompt contributes no gradient. RSTG uses this object as both a gate and a target. Three mechanisms carry the argument: (1) teacher confidence ω_i, the teacher's mean success over 8 rollouts, multiplies the OPD advantage so distillation is strong only where the teacher is reliable; (2) a token-masked OPD advantage keeps only the top-k% tokens ranked by the Soft-OR score s_t = ĥ_t + d̂_t − ĥ_t·d̂_t of normalized student entropy and teacher-student divergence; (3) an SFT term

Load-bearing premise

The load-bearing premise is that distillation is best spent on the narrow slice of prompts where the student fails every rollout and the teacher succeeds, plus the assumption that the teacher's average success over eight rollouts is a reliable per-prompt confidence weight; this was established in one small experiment on one model pair (Qwen3-1.7B to Qwen3-4B) on math, and the full method transfers that conclusion to other pairs, domains, and code.

What would settle it

Run the Figure 3 data-partition comparison on a different teacher-student pair or on code: if the student-wrong/teacher-right subset stops being the best training set, or if student-wrong alone beats it, the selection principle collapses. Alternatively, record per-prompt teacher mean@8 and the per-prompt improvement from distillation; if the correlation is not positive, the confidence weighting is not doing the claimed work.

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

If this is right

  • Combining RLVR with distillation no longer requires giving up exploration: the student can keep improving after approaching the teacher, because token gating and the annealed β slow the fit.
  • Only a small slice of the data — the student-fails/teacher-succeeds subset — needs to carry the distillation signal, cutting the added compute of OPD and letting the SFT data be pre-generated offline.
  • The recipe transfers across domains: math benchmarks gain +4.02% over naive GRPO+OPD on Pair 1 and code benchmarks gain +3.05% on Pair 2, with consistent gains on all three teacher-student pairs.
  • Response-length inflation, a known failure mode of OPD, is damped because the mask and teacher weighting prevent unconditional imitation of teacher-length trajectories.
  • The advantage-asymmetry problem is measurable: RSTG raises per-token advantages relative to naive GRPO+OPD during training, which the paper links to the performance gap.

Where Pith is reading between the lines

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

  • The data-selection finding suggests a dynamic curriculum the paper does not test: as the student improves, the set of negative zero-variance prompts shrinks and shifts, so deciding sample selection online rather than on a fixed pool could amplify the gain.
  • Teacher confidence as mean@8 is a coarse per-prompt scalar; a token-level or calibrated confidence signal could refine the weighting, and this extension is directly testable within RSTG's own equations.
  • The recipe is a general template for any dense auxiliary signal: gate it to regions where the RL advantage is identically zero and where the signal source is competent. That suggests testing RSTG with other dense signals, such as process rewards or critique models, beyond teacher logits.
  • Because the method only spends distillation compute on the initially small failure subset, it may combine naturally with rollout-selection strategies to further cut RLVR cost; the paper does not explore this.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 4 minor

Summary. The paper introduces RSTG, a method for combining GRPO with on-policy distillation (OPD) so that distillation is applied selectively to negative zero-variance prompts—those where all student rollouts fail. RSTG weights OPD advantages by the teacher's mean@8 accuracy on each prompt, restricts OPD updates to a selected subset of tokens based on student entropy and teacher–student divergence, and adds an auxiliary SFT loss on teacher-generated correct solutions for those same prompts. Experiments are reported across three teacher–student pairs from the Qwen2.5 and Qwen3 families on four math benchmarks and two code benchmarks, with comparisons against GRPO, OPD, GRPO+OPD, ReLIFT, and RL-ZVP. The paper reports consistent average gains over naive GRPO+OPD, plus analyses of advantage asymmetry, convergence speed, and response length.

Significance. The central derivation in Appendix A.1 is standard and correct, and the unification of GRPO, OPD, and SFT under a common RL framework is clean and useful. The paper identifies a real limitation of naive GRPO+OPD—no gradient on negative zero-variance prompts, and unselective distillation that can suppress learning—and proposes a principled, well-motivated recipe. The ablations show each component contributing on the reported run. If the empirical findings are robust, RSTG would be a practical advance for RLVR post-training. The main weaknesses are that the load-bearing partition finding (Figure 3) is a single small experiment, the headline numbers are single runs, and β is tuned on the evaluation benchmarks. The contribution is promising but not yet established with the rigor needed for acceptance.

major comments (4)
  1. [§4.1, Figure 3] The central premise—that OPD on Dswtr (3.63% of the data) outperforms OPD on Dsw and on the full D—rests on a single experiment with one model pair (Qwen3-1.7B→Qwen3-4B-2507) on math only, with no seeds, error bars, or transfer evidence. Since Eqs. (9) and (14) encode this partition choice, the core mechanism is not yet shown to be robust. Please provide multi-seed results on at least one pair/domain, or explicitly reframe §4.1 as motivation rather than decisive evidence. As written, the method's main design choice is supported by a single small run.
  2. [§5.5, Table 2] β is selected by evaluating on AIME24, AIME25, MATH500, and OlympiadBench—the same benchmarks used for the main results. This is test-set tuning; it can overstate gains and makes the comparison to baselines, which do not receive comparable tuning, unfair. Use a held-out validation split or nested selection, and report the sensitivity of the main conclusions to β across all three model pairs rather than only Pair 1.
  3. [§5.1, Tables 1–3] All quantitative results are single runs with no variance estimates. Improvements of 0.89–1.97 points on some model pairs and benchmarks are within plausible run-to-run noise at these evaluation sizes. The central claim—that RSTG reliably improves over naive GRPO+OPD—requires at least 2–3 seeds for the key comparisons (RSTG vs. GRPO+OPD, and the component ablations in Table 3). Without variance information, the reported orderings cannot be distinguished from seed luck.
  4. [§4.2, Eq. (13)] The top-k percentage k, which fully determines the token mask, is never reported or ablated. This is a free hyperparameter with strong potential influence on the method's behavior. Please disclose k, explain how it was chosen, report sensitivity, and state whether the choice was made on validation or evaluation data. If k is not fixed, this should be acknowledged as an additional degree of freedom in the method.
minor comments (4)
  1. [§3.2] The sentence 'Eq. (17) takes the same form as Eq.(2)' should refer to Eq. (5) in this section; the appendix numbering is inconsistent.
  2. [Figure 3] The caption does not state the sample sizes of D, Dsw, and Dswtr. The paper mentions 57k, 9k, and 2k in §4.1, but the figure itself should report these values for immediate clarity.
  3. [§5.1, Appendix A.3] APPS is evaluated on a 500-instance subsample, but the MBPP+ subsample size is not stated. Please report both, and consider confidence intervals given the subsampling.
  4. [Limitations] The limitations section acknowledges computational scale but does not mention the single-run nature of the experiments or that β was selected on evaluation benchmarks. These are important limitations of the evidence and should be stated explicitly.

Circularity Check

0 steps flagged

No circularity found: RSTG's design is empirically motivated, not construction-equivalent to its inputs.

full rationale

The paper's derivation chain is empirical rather than formal. The core design decisions — restricting OPD to negative zero-variance prompts and weighting by teacher mean@8 confidence — are motivated by the partition experiment in Figure 3, but that experiment is a genuine exploratory comparison on training-data partitions, not a fitted parameter renamed as a prediction. The final objective in Eq. (14) is a designed training loss; it does not reduce by construction to the partition result, and the headline numbers are evaluated on held-out benchmarks (AIME24/25, MATH500, OLMPIAD, APPS, MBPP+). The β annealing schedule is tuned on the evaluation benchmarks (Table 2), which is an evaluative/leakage concern for soundness, not a circularity of the kind that makes a result equivalent to its input. The manuscript's Limitations section acknowledges that experiments were run at the largest feasible scale and extrapolation to larger scales is left to future work; this is a generalizability caveat, not a circular step. Self-citations to overlapping-author prior work (e.g., Jia et al. 2026, Lu et al. 2026) appear in motivational passages but are not load-bearing: the paper's own Figure 4 and ablation studies independently support the claims those citations accompany. No equation or fitted parameter is shown to be equivalent to a predicted result by construction. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

The method rests on standard policy-gradient machinery (GRPO, reverse-KL OPD, SFT-as-RL) plus several domain assumptions about teacher supervision. The most load-bearing assumptions are that negative zero-variance prompts are the right subset for distillation and that teacher mean@8 success is a reliable confidence weight; both are empirical and only lightly validated. The β schedule and token-selection top-k are free parameters; β is tuned on the evaluation set.

free parameters (4)
  • β annealing schedule (β_init, δ, β_min) = β_init = 5e-3, δ = 5e-5, β_min = 1e-3
    Controls the magnitude of OPD and SFT losses. Selected by comparing performance on math evaluation benchmarks (Table 2) rather than a held-out set; fixing it changes the reported gains.
  • Top-k percentage for token selection = Not specified
    Section 4.2 selects the top k% of tokens by score s_t, but k is never given. It determines how many tokens receive OPD gradients and is a free design choice.
  • Difficulty threshold for math training data = 6
    Section 5.1: only DeepMath samples with difficulty >=6 are used for math RL. This changes the distribution of prompts and the fraction of negative zero-variance prompts.
  • Number of teacher samples for SFT references (n) = 8
    Section 5.1: n=8 teacher responses are sampled per prompt and the shortest correct one is kept as SFT reference. The choice of n affects SFT data quality.
axioms (7)
  • standard math Policy gradient with group-relative advantage (GRPO) is a valid RL update for LLMs
    Used throughout (Eq. 3). Standard result from Shao et al. (2024).
  • standard math The reverse-KL OPD gradient (Eq. 5/17) provides a dense token-level learning signal with advantage A_OPD = -(log π_θ - log π_*)
    Derived in Appendix A.1 using the log-derivative trick.
  • standard math SFT on teacher trajectories is equivalent to RL with constant advantage 1 (Eq. 8)
    Section 3.3; standard equivalence.
  • domain assumption Negative zero-variance prompts (all G rollouts fail) are the right subset for distillation
    Empirical finding from Section 4.1, Figure 3; based on one partition experiment on one model pair.
  • domain assumption Teacher mean@8 success rate is a reliable confidence weight for per-token OPD advantages
    Eq. (9) uses ω_i=mean@8 as a proxy for teacher proficiency; no analysis of when this proxy fails.
  • domain assumption SFT on teacher-generated correct trajectories injects purely positive gradients that offset OPD's negative advantage asymmetry
    Section 4.3 treats every teacher token as gold; assumes teacher responses on negative zero-variance prompts are correct and useful.
  • domain assumption High student entropy or large teacher-student divergence identifies the most informative tokens for distillation
    Section 4.2, adapted from TIP; no direct validation that this selection is optimal.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance." pith.science (2026). https://pith.science/paper/ORQJ722U

@misc{pith2026260800782,
  author       = {Pith},
  title        = {Pith review of: Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORQJ722U}},
  note         = {Machine review of arXiv:2608.00782}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models (LLMs). While Group Relative Policy Optimization (GRPO) is widely adopted, it suffers from sparse reward signals and loses gradients entirely when all responses within a group receive identical rewards. On-policy distillation (OPD) offers a natural remedy by providing dense, token-level supervision from a teacher model. However, naively combining GRPO with OPD leads to degraded performance, due to three underlying causes: not all samples benefit from distillation; fitting too quickly to the teacher undermines the exploratory capacity of RL; and OPD's advantages are asymmetric, suppressing most tokens. To address these challenges, we propose RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), which applies distillation selectively and precisely where it matters most. At the sample level, OPD is restricted to negative zero-variance prompts with each sample weighted by the teacher's confidence score. At the token level, distillation targets only tokens with high student entropy or large teacher-student divergence. We further augment training with SFT on correct trajectories generated by the teacher model, injecting positive gradient signals where RL yields none. Experiments demonstrate that RSTG substantially outperforms naive GRPO+OPD by +4.02% on math and +3.05% on code.

Figures

Figures reproduced from arXiv: 2608.00782 by Deyi Xiong, Hongyan Hao, Jinwei Xiao, Qi Gu, Renren Jin, Xunliang Cai, Yueqing Sun, Yuxin Liu, Yu Yang, Zhengxi Lu, Zhiyuan Yao, Zhuowen Han.

Figure 1
Figure 1. Figure 1: Training Dynamics on MATH. Naively GRPO + OPD proves ineffective, while ours RSTG achieves higher performance. and stability. GRPO normalizes outcome rewards across a group of rollouts to estimate a scalar ad￾vantage applied uniformly to every token, resulting in sparse rewards (Li et al., 2026c) and vanishing gradients when all rollouts are correct or incorrect; we refer to such prompts as positive and ne… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of RSTG framework. For negative zero-variance prompts, RSTG is applied; otherwise, standard GRPO is used. RSTG weights OPD advantages by the teacher’s confidence, restricts gradient updates to high-quality tokens, and performs SFT on teacher-generated reference answers. model performance: (1) Dsw (9k samples), where the student fails all 8 rollouts (mean@8 = 0), and (2) Dswtr (2k samples), a s… view at source ↗
Figure 3
Figure 3. Figure 3: Performance on the mathmatical benchmark [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: OPD advantage curves during training on positive and negative zero-variance prompts. To mitigate this advantage asymmetry of OPD, we augment the training objective for negative zero￾variance prompts with an auxiliary SFT loss com￾puted on correct trajectories pre-generated out by the teacher π ∗ . Formally, the final training objec￾tive is defined as: JFinal(θ) =    JGRPO  θ; β · ωi · AˆOPD i,t … view at source ↗
Figure 5
Figure 5. Figure 5: Performance on MATH500 for Qwen3-1.7B-Instruct [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of advantage values during train [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of response length during training [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance on mathematical benchmarks as components of RSTG are progressively added. MATH500 is excluded from the figure due to its larger score range, but is included in the AVG computation. line that naively combines the GRPO and OPD losses over all samples, and progressively incor￾porate each component of our RSTG. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance of Qwen3-1.7B-Instruct → Qwen3-4B-Instruct-2507 on mathematical benchmarks through￾out training, compared against standard GRPO. 0 100 200 300 400 500 Step 0.02 0.04 0.06 0.08 Accuracy AIME24 GRPO RSTG 0 100 200 300 400 500 Step 0.01 0.02 0.03 0.04 0.05 0.06 Accuracy AIME25 GRPO RSTG 0 100 200 300 400 500 Step 0.62 0.64 0.66 0.68 Accuracy MATH500 GRPO RSTG 0 100 200 300 400 500 Step 0.27 0.28 0… view at source ↗
Figure 10
Figure 10. Figure 10: Performance of Qwen2.5-3B-Instruct → Qwen2.5-14B-Instruct-2507 on mathematical benchmarks throughout training, compared against standard GRPO. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_10.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

39 extracted references · 10 canonical work pages

  1. [1]

    arXiv preprint arXiv:2604.14084 , year=

    Tip: Token importance in on-policy distillation , author=. arXiv preprint arXiv:2604.14084 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2603.25562 , year=

    Revisiting on-policy distillation: Empirical failure modes and simple fixes , author=. arXiv preprint arXiv:2603.25562 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2605.06387 , year=

    Asymmetric On-Policy Distillation: Bridging Exploitation and Imitation at the Token Level , author=. arXiv preprint arXiv:2605.06387 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2603.11137 , year=

    Scaling reasoning efficiently via relaxed on-policy distillation , author=. arXiv preprint arXiv:2603.11137 , year=

    arXiv

  5. [5]

    Lu, Zhengxi and Yao, Zhiyuan and Han, Zhuowen and Wang, Zi-Han and Wu, Jinyang and Gu, Qi and Cai, Xunliang and Lu, Weiming and Xiao, Jun and Zhuang, Yueting and others , journal=

    work page

  6. [6]

    arXiv preprint arXiv:2604.02268 , year=

    Skill0: In-context agentic reinforcement learning for skill internalization , author=. arXiv preprint arXiv:2604.02268 , year=

    Pith/arXiv arXiv

  7. [7]

    CoRR , volume =

    Zhiwei He and Tian Liang and Jiahao Xu and Qiuzhi Liu and Xingyu Chen and Yue Wang and Linfeng Song and Dian Yu and Zhenwen Liang and Wenxuan Wang and Zhuosheng Zhang and Rui Wang and Zhaopeng Tu and Haitao Mi and Dong Yu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2504.11456 , eprinttype =. 2504.11456 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.11456 2025

  8. [8]

    CoRR , volume =

    Ganqu Cui and Lifan Yuan and Zefan Wang and Hanbin Wang and Wendi Li and Bingxiang He and Yuchen Fan and Tianyu Yu and Qixin Xu and Weize Chen and Jiarui Yuan and Huayu Chen and Kaiyan Zhang and Xingtai Lv and Shuo Wang and Yuan Yao and Xu Han and Hao Peng and Yu Cheng and Zhiyuan Liu and Maosong Sun and Bowen Zhou and Ning Ding , title =. CoRR , volume =...

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.01456 2025

  9. [9]

    CoRR , volume =

    Lu Ma and Hao Liang and Meiyi Qiang and Lexiang Tang and Xiaochen Ma and Zhen Hao Wong and Junbo Niu and Chengyu Shen and Runming He and Bin Cui and Wentao Zhang , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.07527 , eprinttype =. 2506.07527 , timestamp =

    work page doi:10.48550/arxiv.2506.07527 2025

  10. [10]

    CoRR , volume =

    Thanh. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2509.21880 , eprinttype =. 2509.21880 , timestamp =

    work page doi:10.48550/arxiv.2509.21880 2025

  11. [11]

    CoRR , volume =

    Hongling Xu and Qi Zhu and Heyuan Deng and Jinpeng Li and Lu Hou and Yasheng Wang and Lifeng Shang and Ruifeng Xu and Fei Mi , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.02208 , eprinttype =. 2506.02208 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.02208 2025

  12. [12]

    Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual , year =

    Dan Hendrycks and Steven Basart and Saurav Kadavath and Mantas Mazeika and Akul Arora and Ethan Guo and Collin Burns and Samir Puranik and Horace He and Dawn Song and Jacob Steinhardt , editor =. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual , year =

    work page 2021

  13. [13]

    CoRR , volume =

    Yaxuan Li and Yuxin Zuo and Bingxiang He and Jinqian Zhang and Chaojun Xiao and Cheng Qian and Tianyu Yu and Huan. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.13016 , eprinttype =. 2604.13016 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.13016 2026

  14. [14]

    CoRR , volume =

    Qwen Team , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2505.09388 , eprinttype =. 2505.09388 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  15. [15]

    CoRR , volume =

    Kimi Team , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.02276 , eprinttype =. 2602.02276 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.02276 2026

  16. [16]

    CoRR , volume =

    DeepSeek. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2501.12948 , eprinttype =. 2501.12948 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.12948 2025

  17. [17]

    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 =. 2024 , url =. doi:10.48550/ARXIV.2402.03300 , eprinttype =. 2402.03300 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300 2024

  18. [18]

    CoRR , volume =

    Gemini Team , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2507.06261 , eprinttype =. 2507.06261 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.06261 2025

  19. [19]

    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 =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.02177 , eprinttype =. 2506.02177 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.02177 2025

  20. [20]

    CoRR , volume =

    Yunzhen Feng and Parag Jain and Anthony Hartshorn and Yaqi Duan and Julia Kempe , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2510.08696 , eprinttype =. 2510.08696 , timestamp =

    work page doi:10.48550/arxiv.2510.08696 2025

  21. [21]

    The Twelfth International Conference on Learning Representations,

    Rishabh Agarwal and Nino Vieillard and Yongchao Zhou and Piotr Stanczyk and Sabela Ramos Garea and Matthieu Geist and Olivier Bachem , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  22. [22]

    The Twelfth International Conference on Learning Representations,

    Yuxian Gu and Li Dong and Furu Wei and Minlie Huang , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  23. [23]

    CoRR , volume =

    Wenkai Yang and Weijie Liu and Ruobing Xie and Kai Yang and Saiyong Yang and Yankai Lin , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.12125 , eprinttype =. 2602.12125 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.12125 2026

  24. [24]

    Hinton and Oriol Vinyals and Jeffrey Dean , title =

    Geoffrey E. Hinton and Oriol Vinyals and Jeffrey Dean , title =. CoRR , volume =. 2015 , url =. 1503.02531 , timestamp =

    Pith/arXiv arXiv 2015

  25. [25]

    CoRR , volume =

    Chenxu Yang and Chuanyu Qin and Qingyi Si and Minghui Chen and Naibin Gu and Dingyu Yao and Zheng Lin and Weiping Wang and Jiaqi Wang and Nan Duan , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.03128 , eprinttype =. 2604.03128 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.03128 2026

  26. [26]

    CoRR , volume =

    Gengsheng Li and Tianyu Yang and Junfeng Fang and Mingyang Song and Mao Zheng and Haiyun Guo and Dan Zhang and Jinqiao Wang and Tat. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.02288 , eprinttype =. 2604.02288 , timestamp =

    work page doi:10.48550/arxiv.2604.02288 2026

  27. [27]

    CoRR , volume =

    Jongwoo Ko and Sara Abdali and Young Jin Kim and Tianyi Chen and Pashmina Cameron , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2603.11137 , eprinttype =. 2603.11137 , timestamp =

    work page doi:10.48550/arxiv.2603.11137 2026

  28. [28]

    CoRR , volume =

    Feng Luo and Yu. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.08527 , eprinttype =. 2604.08527 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.08527 2026

  29. [29]

    CoRR , volume =

    An Yang and Baosong Yang and Beichen Zhang and Binyuan Hui and Bo Zheng and Bowen Yu and Chengyuan Li and Dayiheng Liu and Fei Huang and Haoran Wei and Huan Lin and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Yang and Jiaxi Yang and Jingren Zhou and Junyang Lin and Kai Dang and Keming Lu and Keqin Bao and Kexin Yang and Le Yu and Mei Li and Mi...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2024

  30. [30]

    The Thirteenth International Conference on Learning Representations,

    Bofei Gao and Feifan Song and Zhe Yang and Zefan Cai and Yibo Miao and Qingxiu Dong and Lei Li and Chenghao Ma and Liang Chen and Runxin Xu and Zhengyang Tang and Benyou Wang and Daoguang Zan and Shanghaoran Quan and Ge Zhang and Lei Sha and Yichang Zhang and Xuancheng Ren and Tianyu Liu and Baobao Chang , title =. The Thirteenth International Conference ...

    2025

  31. [31]

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

    Chaoqun He and Renjie Luo and Yuzhuo Bai and Shengding Hu and Zhen Leng Thai and Junhao Shen and Jinyi Hu and Xu Han and Yujie Huang and Yuxiang Zhang and Jie Liu and Lei Qi and Zhiyuan Liu and Maosong Sun , editor =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2024 , url =. doi:10.1865...

    doi:10.18653/v1/2024.acl-long.211 2024

  32. [32]

    CoRR , volume =

    Jiarui Yao and Yifan Hao and Hanning Zhang and Hanze Dong and Wei Xiong and Nan Jiang and Tong Zhang , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2505.02391 , eprinttype =. 2505.02391 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.02391 2025

  33. [33]

    , author=

    American Invitational Mathematics Examination - AIME 2024. , author=

    work page 2024

  34. [34]

    , author=

    American Invitational Mathematics Examination - AIME 2025. , author=

    work page 2025

  35. [35]

    Jiawei Liu and Chunqiu Steven Xia and Yuyao Wang and Lingming Zhang , editor =. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

    work page 2023

  36. [36]

    CoRR , volume =

    Wenhong Zhu and Ruobing Xie and Rui Wang and Pengfei Liu , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.20244 , eprinttype =. 2604.20244 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.20244 2026

  37. [37]

    CoRR , volume =

    Yongliang Wu and Yizhou Zhou and Zhou Ziheng and Yingzhe Peng and Xinyu Ye and Xinting Hu and Wenbo Zhu and Lu Qi and Ming. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2508.05629 , eprinttype =. 2508.05629 , timestamp =

    doi:10.48550/arxiv.2508.05629 2025

  38. [38]

    CoRR , volume =

    Ziheng Li and Liu Kang and Feng Xiao and Luxi Xing and Qingyi Si and Zhuoran Li and Weikang Gong and Deqing Yang and Yanghua Xiao and Hongcheng Guo , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2601.07408 , eprinttype =. 2601.07408 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2601.07408 2026

  39. [39]

    Proceedings of the Twentieth European Conference on Computer Systems, EuroSys 2025, Rotterdam, The Netherlands, 30 March 2025 - 3 April 2025 , pages =

    Guangming Sheng and Chi Zhang and Zilingfeng Ye and Xibin Wu and Wang Zhang and Ru Zhang and Yanghua Peng and Haibin Lin and Chuan Wu , title =. Proceedings of the Twentieth European Conference on Computer Systems, EuroSys 2025, Rotterdam, The Netherlands, 30 March 2025 - 3 April 2025 , pages =. 2025 , url =. doi:10.1145/3689031.3696075 , timestamp =

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.