pith. sign in

arxiv: 2605.09253 · v4 · pith:2QE6M6KHnew · submitted 2026-05-10 · 💻 cs.CL · cs.AI

Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation

Yuxuan Jiang , Runchao Li , Shubhashis Roy Dipta , Dawei Li , Zhao Yang This is my paper

Pith reviewed 2026-07-02 23:32 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords on-policy distillationrock tokenshigh-loss tokensreasoning performancecausal interventiontoken weightingmodel alignmentgradient norms
0
0 comments X

The pith

Rock Tokens persist with high loss after on-policy distillation saturates yet add negligible value to reasoning.

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

The paper investigates high-loss tokens during on-policy distillation, where the per-token KL objective should drive all mismatches toward zero as training converges. It identifies Rock Tokens as a persistent subset that remains high-loss even after apparent saturation and can comprise up to 18 percent of generated outputs. Despite supplying a large fraction of gradient norms, causal interventions show these tokens contribute almost nothing to the model's downstream reasoning performance. The findings indicate that substantial optimization effort targets discourse and structural elements the student does not need to acquire.

Core claim

Even after OPD training reaches apparent saturation, a substantial subset of tokens continues to exhibit persistently high loss; these tokens, which we term Rock Tokens, can account for up to 18% of the tokens in generated outputs. Rock Tokens provide a disproportionately large share of total gradient norms yet remain stagnant throughout training, resisting teacher-driven corrections. Through causal intervention, these tokens provide negligible functional contribution to the model's actual reasoning performance.

What carries the argument

Rock Tokens: the subset of tokens that maintain persistently high per-token loss under the KL objective in on-policy distillation, resist correction, and show negligible causal impact on reasoning.

If this is right

  • A large share of gradient updates is directed at structural and discourse residuals the student model cannot or need not internalize.
  • Strategically bypassing Rock Tokens can streamline the alignment process.
  • Uniform token weighting is not required for effective distillation.
  • Large-scale model distillation can adopt a more efficient paradigm by de-emphasizing non-functional tokens.

Where Pith is reading between the lines

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

  • The same high-loss non-functional tokens may appear in reinforcement learning with verifiable rewards and other alignment methods.
  • Training objectives could be redesigned to down-weight or ignore Rock Tokens from the start.
  • Inference-time filtering of Rock Tokens might reduce output length without harming task performance.

Load-bearing premise

Causal interventions that measure the functional contribution of Rock Tokens do so without altering other tokens or overall model dynamics.

What would settle it

Selectively masking or replacing only Rock Tokens during generation and observing a measurable drop in reasoning accuracy on held-out tasks would falsify the negligible-contribution claim.

Figures

Figures reproduced from arXiv: 2605.09253 by Dawei Li, Runchao Li, Shubhashis Roy Dipta, Yuxuan Jiang, Zhao Yang.

Figure 1
Figure 1. Figure 1: The lifecycle and functional impact of Rock Tokens in OPD. (a) Phenomenon: Identifi￾cation of optimization-resistant tokens. (b) Mechanism: Causal evidence of structural redundancy via token knock-out. (c) Utility: Performance parity achieved through strategic gradient sparsification. is inferred from policy entropy or reward signals, OPD provides a direct measure of student-teacher mismatch through per-to… view at source ↗
Figure 2
Figure 2. Figure 2: Empirical identification and stability of Rock Tokens. (a) Per-token KL ℓbv vs. frequency on N=500 MATH-500 trajectories: rare tokens are noise-dominated, while the Rock Score R(v) isolates true Rock Tokens (red) at the upper edge of stable frequency bands. (b) Per-sequence Rock-Token density (median 18.5%). (c) Cumulative loss coverage (blue) and selection stability (red) vs. cutoff K; K=100 balances repr… view at source ↗
Figure 3
Figure 3. Figure 3: Per-token gradient geometry and persistence under training. (a) Per-token logit￾gradient magnitude ∥g¯t∥ by group: rocks are an order of magnitude smaller than rare high-KL tokens. (b) Cosine alignment with the frequency-balanced descent direction Gbalanced: rocks are positively aligned, with a tail reaching cos > 0.3. (c) Per-token mean KL paired across two training checkpoints (log-log). Points below the… view at source ↗
Figure 4
Figure 4. Figure 4: Knockout effect on the |R| e = 200 screened Rock-Token candidates, sorted by ∆. Each bar is a single candidate; height is the accuracy change when its logit is masked at decode time. The shaded grey band marks the categorization threshold |∆| < ε = 0.01. Bars outside the band that pass the paired-bootstrap test (α = 0.05, 10,000 resamples) are coloured by category (Strong Pillar in red; Strong Stumbling in… view at source ↗
Figure 5
Figure 5. Figure 5: Average accuracy across AIME24, AIME25, and HMMT25 during OPD training. Each 200 training steps correspond to 8,000 prompts, with 4 rollouts per prompt. 4.1 RQ3: What is the genuine functional contribution of Rock Tokens to model training? The functional redundancy of Rock Tokens at in￾ference prompts a critical question: do their persis￾tent high-loss signals provide essential constraints, or are they mer… view at source ↗
Figure 6
Figure 6. Figure 6: Pillarhood is not predicted by entropy, frequency, or loss. MATH-500 knockout ∆ for each of the |R| e = 200 screened candidates (Strong Pillars in red) plotted against six candidate predictors: post- and pre-OPD student entropy, teacher entropy, log-frequency, rock rate, and mean post-OPD KL. Annotated r, p are Pearson correlations over all 200 candidates. None reach |r| > 0.07. Multiple-testing considerat… view at source ↗
Figure 6
Figure 6. Figure 6: Pillarhood is not predicted by entropy, frequency, or loss. MATH-500 knockout ∆ for each of the |R| e = 200 screened candidates (Strong Pillars in red) plotted against six candidate predictors: post- and pre-OPD student entropy, teacher entropy, log-frequency, rock rate, and mean post-OPD KL. Annotated r, p are Pearson correlations over all 200 candidates. None reach |r| > 0.07. “Do”; IFEval includes “ too… view at source ↗
read the original abstract

While recent work in Reinforcement Learning with Verifiable Rewards (RLVR) has shown that a small subset of critical tokens disproportionately drives reasoning gains, an analogous token-level understanding of On-Policy Distillation (OPD) remains largely unexplored. In this work, we investigate high-loss tokens, a token type that--as the most direct signal of student-teacher mismatch under OPD's per-token KL objective--should progressively diminish as training converges according to existing studies; however, our empirical analysis shows otherwise. Even after OPD training reaches apparent saturation, a substantial subset of tokens continues to exhibit persistently high loss; these tokens, which we term Rock Tokens, can account for up to 18\% of the tokens in generated outputs. Our investigation reveals two startling paradoxes. First, despite their high occurrence frequency providing a disproportionately large share of total gradient norms, Rock Tokens themselves remain stagnant throughout training, resisting teacher-driven corrections. Second, through causal intervention, we find that these tokens provide negligible functional contribution to the model's actual reasoning performance. These findings suggest that a vast amount of optimization bandwidth is spent on structural and discourse residuals that the student model cannot or need not internalize. By deconstructing these dynamics, we demonstrate that strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper examines token-level dynamics in On-Policy Distillation (OPD), identifying a persistent subset of high-loss tokens termed 'Rock Tokens' that continue to exhibit elevated loss after apparent training saturation. These tokens are reported to comprise up to 18% of generated outputs, account for a disproportionate share of gradient norms, yet causal interventions indicate they contribute negligibly to downstream reasoning performance. The authors conclude that standard OPD wastes optimization effort on these structural residuals and that selectively bypassing them can improve distillation efficiency.

Significance. If the empirical observations and intervention results hold under rigorous controls, the work identifies a concrete inefficiency in current OPD practice and supplies a practical lever for reducing unnecessary gradient computation. The strength lies in the combination of observational statistics with targeted interventions; however, the absence of reported dataset sizes, run counts, or explicit checks for intervention side-effects limits the immediate actionability of the efficiency claim.

major comments (3)
  1. [§4.2] §4.2 (Causal Intervention Protocol): the description of the masking/forced-logit interventions does not report any quantitative verification that non-rock token losses, next-token probabilities, or overall sequence-level KL divergence remain unchanged after the intervention; without such checks the claim of 'negligible functional contribution' rests on an untested isolation assumption.
  2. [§5.1, Table 3] §5.1 and Table 3: the reported 18% occurrence figure and gradient-norm share are presented without accompanying standard errors, number of evaluation runs, or dataset cardinality; this makes it impossible to assess whether the 'up to 18%' figure is stable or sensitive to post-hoc token selection criteria.
  3. [§4.3] §4.3 (Reasoning Performance Metrics): the performance comparison after rock-token intervention lacks an explicit control condition in which an equal number of randomly chosen tokens (matched for position or frequency) are altered; without this baseline it is unclear whether the observed lack of degradation is specific to rock tokens or an artifact of any token-level perturbation.
minor comments (2)
  1. [Abstract / §3] The abstract states that rock tokens 'resist teacher-driven corrections' but provides no explicit definition or equation for the per-token loss threshold used to classify them; a short formal definition in §3 would improve reproducibility.
  2. [Figure 2] Figure 2 caption refers to 'gradient norm share' without clarifying whether the plotted quantity is normalized per sequence or globally; adding this detail would prevent misinterpretation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point-by-point below and indicate planned revisions to improve clarity and rigor.

read point-by-point responses

  1. Referee: [§4.2] §4.2 (Causal Intervention Protocol): the description of the masking/forced-logit interventions does not report any quantitative verification that non-rock token losses, next-token probabilities, or overall sequence-level KL divergence remain unchanged after the intervention; without such checks the claim of 'negligible functional contribution' rests on an untested isolation assumption.

    Authors: We acknowledge that the original manuscript did not include explicit post-intervention statistics confirming invariance for non-rock tokens. The protocol targeted only rock-token positions, but we agree that quantitative verification is needed to substantiate the isolation claim. In the revised version we will add measurements of non-rock token losses, next-token probabilities, and sequence-level KL divergence before and after intervention. revision: yes

  2. Referee: [§5.1, Table 3] §5.1 and Table 3: the reported 18% occurrence figure and gradient-norm share are presented without accompanying standard errors, number of evaluation runs, or dataset cardinality; this makes it impossible to assess whether the 'up to 18%' figure is stable or sensitive to post-hoc token selection criteria.

    Authors: The 18% figure is the maximum observed value across our primary evaluation sets. We agree that reporting dataset cardinality, number of runs, and standard errors would allow better assessment of stability. The revision will specify the evaluation dataset sizes, run counts, and include error bars or ranges for the occurrence and gradient-norm statistics in §5.1 and Table 3. revision: yes

  3. Referee: [§4.3] §4.3 (Reasoning Performance Metrics): the performance comparison after rock-token intervention lacks an explicit control condition in which an equal number of randomly chosen tokens (matched for position or frequency) are altered; without this baseline it is unclear whether the observed lack of degradation is specific to rock tokens or an artifact of any token-level perturbation.

    Authors: We agree that a matched random-token control is necessary to establish specificity. The revised manuscript will include an additional control experiment in which an equal number of randomly selected tokens (matched for position or frequency) receive the same intervention, allowing direct comparison of performance impact. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical categorization and intervention study

full rationale

The manuscript contains no derivation chain, equations, fitted parameters renamed as predictions, or load-bearing self-citations. Rock Tokens are introduced as an empirical label for persistently high-loss tokens observed after OPD saturation; their frequency (up to 18%), gradient-norm share, and negligible functional contribution are measured directly via data inspection and causal interventions. These steps are externally falsifiable and do not reduce to the paper's own inputs by construction. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the empirical definition of Rock Tokens via loss thresholds and the validity of causal interventions; no free parameters are explicitly fitted in the abstract, and the only invented entity is the label itself.

axioms (1)
  • domain assumption High-loss tokens should progressively diminish as OPD training converges, per existing studies
    The abstract contrasts the observed persistence against this expectation drawn from prior work.
invented entities (1)
  • Rock Tokens no independent evidence
    purpose: Label for the subset of persistently high-loss tokens that resist teacher-driven correction
    New term introduced to describe the observed phenomenon.

pith-pipeline@v0.9.1-grok · 5796 in / 1289 out tokens · 37918 ms · 2026-07-02T23:32:26.945878+00:00 · methodology

Review history (3 revisions) →

discussion (0)

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

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems

    cs.SE 2026-05 unverdicted novelty 6.0

    About 18.2% of structurally flagged skill pairs represent genuine compositional safety risks in agent skill registries, with exploitation gated by host model behavior.

  2. Bridging Reasoning Trajectories in On-Policy Distillation via Near-Future Guidance

    cs.CL 2026-05 unverdicted novelty 6.0

    TOPD augments on-policy distillation by using near-future trajectory signals to suppress non-divergent high-loss tokens and distribute guidance, raising average accuracy from 47.8% to 52.2% on reasoning benchmarks.

  3. Bridging Reasoning Trajectories in On-Policy Distillation via Near-Future Guidance

    cs.CL 2026-05 unverdicted novelty 6.0

    TOPD improves on-policy distillation for LLM reasoning by using near-future guidance to identify divergent states, raising average accuracy from 47.8% to 52.2% on math benchmarks including AIME24 and AIME25.

  4. A Formula-Driven Survey and Research Agenda for On-Policy Distillation

    cs.AI 2026-06 unverdicted novelty 4.0

    A survey creates a taxonomy for on-policy distillation in LLMs that separates temporal credit assignment from vocabulary-level probability routing.

  5. RankVR: Low-Rank Structure Perception and Value Recalibration for Robust Composed Image Retrieval

    cs.CV 2026-06 unverdicted novelty 4.0

    RankVR introduces GSCP and ASVC modules to improve CIR robustness by decoupling clean samples via low-rank structure and dynamically scoring triplet value in noisy datasets.

  6. IMAGINE: Adaptive Schema-Imagery Enhanced Composition for Composed Video Retrieval

    cs.CV 2026-06 unverdicted novelty 4.0

    IMAGINE uses adaptive schema-imagery via dynamic multimodal prototypes to incorporate implicit semantics into composed video retrieval, claiming SOTA results on CVR and CIR benchmarks.

Reference graph

Works this paper leans on

93 extracted references · 43 canonical work pages · cited by 5 Pith papers · 18 internal anchors

  1. [1]

    Agarwal, N

    R. Agarwal, N. Vieillard, Y . Zhou, P. Stanczyk, S. R. Garea, M. Geist, and O. Bachem. On-policy distillation of language models: Learning from self-generated mistakes. InThe twelfth international conference on learning representations, 2024

    2024

  2. [2]

    Al Nazi, S

    Z. Al Nazi, S. R. Dipta, and S. Kar. † dagger: Distractor-aware graph generation for executable reasoning in math problems.arXiv e-prints, pages arXiv–2601, 2026

    2026

  3. [3]

    J. Cao, Y . Ma, X. Li, Q. Ren, and X. Chen. Task-specific efficiency analysis: When small language models outperform large language models, 2026. URLhttps://arxiv.org/abs/2603.21389

    work page arXiv 2026

  4. [4]

    W. Chen, L. Wu, Y . Hu, Z. Li, Z. Cheng, Y . Qian, L. Zhu, Z. Hu, L. Liang, Q. Tang, Z. Liu, and H. Yang. Autoneural: Co-designing vision-language models for npu inference, 2025. URL https: //arxiv.org/abs/2512.02924

    work page arXiv 2025

  5. [5]

    Y . Chen, P. Qian, S. Wang, S. Zhang, H. Xu, S. Lin, and X. Wei. Does rag know when retrieval is wrong? diagnosing context compliance under knowledge conflict, 2026. URL https://arxiv.org/abs/2605. 14473

    2026

  6. [6]

    Cheng, S

    D. Cheng, S. Huang, X. Zhu, B. Dai, X. Zhao, Z. Zhang, and F. Wei. Reasoning with exploration: An entropy perspective. InProceedings of the AAAI Conference on Artificial Intelligence, 2026

    2026

  7. [7]

    Cheng, X

    K. Cheng, X. Qi, Z. Cheng, and L. Lai. V olatility persistence and model choice in cross-market volatility forecasting.Available at SSRN 6610278, 2026. 12

    2026

  8. [8]

    Cheng, X

    K. Cheng, X. Qi, Z. Cheng, L. Lai, and X. Liu. Regime-dependent volatility dynamics: Evidence from time-series analysis. InProceedings of the 2026 3rd International Conference on Applied Economics, Management Science and Social Development (AEMSS 2026), pages 179–189. Atlantis Press, 2026. ISBN 978-94-6239-672-2. doi: 10.2991/978-94-6239-672-2_18. URL htt...

    work page doi:10.2991/978-94-6239-672-2_18 2026

  9. [9]

    Sustainable Hybrid Document-Routed Retrieval for Financial RAG: Resolving the Robustness-Precision Trade-off

    Z. Cheng, L. Lai, and Y . Liu. Resolving the robustness-precision trade-off in financial rag through hybrid document-routed retrieval, 2026. URLhttps://arxiv.org/abs/2603.26815

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    Cheng, L

    Z. Cheng, L. Lai, Y . Liu, K. Cheng, and X. Qi. Enhancing financial report question-answering: A retrieval- augmented generation system with reranking analysis, 2026. URL https://arxiv.org/abs/2603. 16877

    2026

  11. [11]

    Cheng, L

    Z. Cheng, L. Lai, Y . Liu, and Y . Sun. Toward sustainable on-device intelligence: A survey on energy- efficient rag systems with small language models.Available at SSRN 6698538, 2026

    2026

  12. [12]

    Y . Dai, M. Chen, and Z. Zuo. Neighbors in space: Satellite imagery and chinese b-share discount.China Economic Review, 82:102063, 2023

    2023

  13. [13]

    Deepseek-v4: Towards highly efficient million-token context intelligence.Technical Report, 2026

    DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence.Technical Report, 2026

    2026

  14. [14]

    Dekoninck, N

    J. Dekoninck, N. Jovanovi’c, T. Gehrunger, K. Rognvalddson, I. Petrov, C. Sun, and M. T. Vechev. Beyond benchmarks: Matharena as an evaluation platform for mathematics with llms. 2026

    2026

  15. [15]

    S. R. Dipta, D. Bis, K. Zhou, L. Wang, B. Z. Yao, C. Guo, and R. Sarikaya. Pa3: Policy-aware agent alignment through chain-of-thought.arXiv preprint arXiv:2603.14602, 2026

    work page arXiv 2026

  16. [16]

    S. R. Dipta, K. Mahbub, and N. Najjar. Ganitllm: Difficulty-aware bengali mathematical reasoning through curriculum-grpo.arXiv preprint arXiv:2601.06767, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  17. [17]

    Y . Fu, H. Huang, K. Jiang, Y . Zhu, and D. Zhao. Revisiting on-policy distillation: Empirical failure modes and simple fixes.ArXiv preprint, 2026

    2026

  18. [18]

    Z. Fu, Y . Hu, Q. Yang, S. Zhang, Z. Chen, and Z. Li. Air-know: Arbiter-calibrated knowledge-internalizing robust network for composed image retrieval, 2026. URLhttps://arxiv.org/abs/2604.19386

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    L. Gao, J. Tow, B. Abbasi, S. Biderman, S. Black, A. DiPofi, C. Foster, L. Golding, J. Hsu, A. Le Noac’h, H. Li, K. McDonell, N. Muennighoff, C. Ociepa, J. Phang, L. Reynolds, H. Schoelkopf, A. Skowron, L. Sutawika, E. Tang, A. Thite, B. Wang, K. Wang, and A. Zou. The language model evaluation harness, 2024

    2024

  20. [20]

    Y . Gao, Y . Lu, Z. Zhang, J. Nie, S. Yu, and Q. Xuan. Dspc: Dual-stage progressive compression framework for efficient long-context reasoning. InICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 19387–19391, 2026. doi: 10.1109/ICASSP55912.2026. 11460600

    work page doi:10.1109/icassp55912.2026 2026

  21. [21]

    E. Guha, R. Marten, S. S. Keh, N. Raoof, G. Smyrnis, H. Bansal, M. Nezhurina, J.-P. Mercat, T. Vu, Z. Sprague, A. Suvarna, B. Feuer, L. Chen, Z. Khan, E. Frankel, S. Grover, C. Choi, N. Muennighoff, S. Su, W. Zhao, J. Yang, S. Pimpalgaonkar, K. Sharma, C. C.-J. Ji, Y . Deng, S. Pratt, V . Ramanujan, J. Saad-Falcon, J. Li, A. Dave, A. Albalak, K. Arora, B....

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.04178 2025

  22. [22]

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    J. Guo, X. Luo, J. Zheng, Y . Wang, K.-W. Chang, W. Wang, and J. Liu. Quantized-tinyllava: A new multimodal foundation model enables efficient split learning. InarXiv preprint arXiv:2511.23402, 2025

    work page arXiv 2025

  24. [24]

    How do llms use their depth?arXiv preprint arXiv:2510.18871,

    A. Gupta, J. Yeung, G. Anumanchipalli, and A. A. Ivanova. How do llms use their depth?arXiv.org, 2025. doi: 10.48550/arxiv.2510.18871

    work page doi:10.48550/arxiv.2510.18871 2025

  25. [25]

    Hübotter, F

    J. Hübotter, F. Lübeck, L. Behric, A. Baumann, M. Bagatella, D. Marta, I. Hakimi, I. Shenfeld, T. K. Buening, C. Guestrin, et al. Reinforcement learning via self-distillation.ArXiv preprint, 2026. 13

    2026

  26. [26]

    MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents

    D. Jiang, Y . Li, G. Li, and B. Li. Magma: A multi-graph based agentic memory architecture for ai agents. arXiv preprint arXiv:2601.03236, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  27. [27]

    Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations

    D. Jiang, Y . Li, S. Wei, J. Yang, A. Kishore, A. Zhao, D. Kang, X. Hu, F. Chen, Q. Li, et al. Anatomy of agentic memory: Taxonomy and empirical analysis of evaluation and system limitations.arXiv preprint arXiv:2602.19320, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [29]

    SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models

    Y . Jiang and F. Ferraro. Scribe: Structured mid-level supervision for tool-using language models, 2026. URLhttps://arxiv.org/abs/2601.03555

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [30]

    DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Models

    Y . Jiang, D. Li, and F. Ferraro. Drp: Distilled reasoning pruning with skill-aware step decomposition for efficient large reasoning models.arXiv preprint arXiv:2505.13975, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [31]

    Kunstner, R

    F. Kunstner, R. Yadav, A. Milligan, M. Schmidt, and A. Bietti. Heavy-tailed class imbalance and why adam outperforms gradient descent on language models, 2024. URL https://arxiv.org/abs/2402.19449

    work page arXiv 2024

  31. [32]

    L. Lai, Z. Cheng, K. Cheng, and X. Qi. Do transformers always win? an empirical study of semantic embeddings for short-text e-commerce reviews. In2026 9th International Symposium on Big Data and Applied Statistics (ISBDAS), pages 525–529, 2026. doi: 10.1109/ISBDAS69350.2026.11484350

    work page doi:10.1109/isbdas69350.2026.11484350 2026

  32. [33]

    Y . Lei, Y . Wang, Y . Zhang, B. Guan, D. Zhu, C. Wang, Z. Hao, and T. Shi. Sheet as token: A graph- enhanced representation for multi-sheet spreadsheet understanding, 2026. URL https://arxiv.org/ abs/2605.05811

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [34]

    D. Li, Z. Tan, T. Chen, and H. Liu. Contextualization distillation from large language model for knowledge graph completion. InFindings of the Association for Computational Linguistics: EACL 2024, pages 458–477, 2024

    2024

  34. [35]

    D. Li, B. Jiang, L. Huang, A. Beigi, C. Zhao, Z. Tan, A. Bhattacharjee, Y . Jiang, C. Chen, T. Wu, K. Shu, L. Cheng, and H. Liu. From generation to judgment: Opportunities and challenges of LLM- as-a-judge. In C. Christodoulopoulos, T. Chakraborty, C. Rose, and V . Peng, editors,Proceedings of the 2025 Conference on Empirical Methods in Natural Language P...

    work page doi:10.18653/v1/2025.emnlp-main.138 2025

  35. [36]

    D. Li, B. Jiang, L. Huang, A. Beigi, C. Zhao, Z. Tan, A. Bhattacharjee, Y . Jiang, C. Chen, T. Wu, et al. From generation to judgment: Opportunities and challenges of llm-as-a-judge. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 2757–2791, 2025

    2025

  36. [37]

    D. Li, R. Sun, Y . Huang, M. Zhong, B. Jiang, J. Han, X. Zhang, W. Wang, and H. Liu. Preference leakage: A contamination problem in llm-as-a-judge.arXiv preprint arXiv:2502.01534, 2025

    work page arXiv 2025

  37. [38]

    X. Li, Y . Ma, Y . Huang, X. Wang, Y . Lin, and C. Zhang. Synergized data efficiency and compression (sec) optimization for large language models. In2024 4th International Conference on Electronic Information Engineering and Computer Science (EIECS), pages 586–591, 2024. doi: 10.1109/EIECS63941.2024. 10800533

    work page doi:10.1109/eiecs63941.2024 2024

  38. [39]

    X. Li, Y . Ma, K. Ye, J. Cao, M. Zhou, and Y . Zhou. Hy-Facial: hybrid feature extraction by dimensionality reduction methods for enhanced facial expression classification. In W. Osten and E. Mamut, editors, Eighteenth International Conference on Machine Vision (ICMV 2025), volume 14114, page 141140R. International Society for Optics and Photonics, SPIE, ...

    work page doi:10.1117/12.3096291 2025

  39. [40]

    Y . Li, Y . Lu, Z. Dong, C. Yang, Y . Chen, and J. Gou. Sglp: A similarity guided fast layer partition pruning for compressing large deep models.arXiv preprint arXiv:2410.14720, 2024

    work page arXiv 2024

  40. [41]

    Y . Li, K. Ding, C. Yang, H. Wang, H. Wang, H. Duan, J. Liu, and Y . Tian. Ddtime: Dataset distillation with spectral alignment and information bottleneck for time-series forecasting.arXiv preprint arXiv:2511.16715, 2025

    work page arXiv 2025

  41. [42]

    Y . Li, J. Dong, Z. Dong, C. Yang, Z. An, and Y . Xu. Srkd: Towards efficient 3d point cloud segmentation via structure-and relation-aware knowledge distillation.arXiv preprint arXiv:2506.17290, 2025

    work page arXiv 2025

  42. [43]

    Y . Li, J. Dong, C. Yang, S. Wen, P. Koniusz, T. Huang, Y . Tian, and Y .-S. Ong. Mmt-ard: Multimodal multi-teacher adversarial distillation for robust vision-language models.arXiv preprint arXiv:2511.17448, 2025. 14

    work page arXiv 2025

  43. [44]

    Y . Li, Q. Long, Y . Zhou, R. Zhang, Z. Ning, Z. Zhu, Y . Zhou, X. Wang, and M. Xiao. Comae: Comprehen- sive attribute exploration for zero-shot hashing.ICMR, 2025

    2025

  44. [45]

    Y . Li, C. Yang, J. Dong, Z. Yao, H. Xu, Z. Dong, H. Zeng, Z. An, and Y . Tian. Ammkd: Adaptive multi- modal multi-teacher distillation for lightweight vision-language models.arXiv preprint arXiv:2509.00039, 2025

    work page arXiv 2025

  45. [46]

    Y . Li, C. Yang, H. Zeng, Z. Dong, Z. An, Y . Xu, Y . Tian, and H. Wu. Frequency-aligned knowledge distillation for lightweight spatiotemporal forecasting. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262–7272, 2025

    2025

  46. [47]

    Y . Li, H. Zeng, F. Zhang, C. Yang, Y . Li, and W. Ding. Efficient Medical Image Segmentation via Rein- forcement Learning-Driven K-Space Sampling.IEEE Transactions on Emerging Topics in Computational Intelligence, 2025. ISSN 2471-285X. doi: 10.1109/TETCI.2025.3621221

    work page doi:10.1109/tetci.2025.3621221 2025

  47. [48]

    Y . Li, K. Ding, C. Yang, S.-Y . Chen, and Y . Tian. Distilling time series foundation models for efficient forecasting. InICASSP, 2026

    2026

  48. [49]

    Y . Li, K. Li, X. Yin, Z. Yang, Z. Dong, Z. Yao, H. Xu, Y . Tian, and Y . Lu. Sepprune: Structured pruning for efficient deep speech separation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 31861–31869, 2026

    2026

  49. [50]

    Y . Li, S. Meng, C. Yang, W. Feng, J. Liu, Z. An, Y . Wang, and Y . Tian. A comprehensive survey of interaction techniques in 3d scene generation.Authorea Preprints, 2026

    2026

  50. [51]

    Y . Li, Y . Zuo, B. He, J. Zhang, C. Xiao, C. Qian, T. Yu, H. Gao, W. Yang, Z. Liu, and N. Ding. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.ArXiv preprint, 2026

    2026

  51. [52]

    Z. Li, Z. Chen, H. Wen, Z. Fu, Y . Hu, and W. Guan. Encoder: Entity mining and modification relation binding for composed image retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5101–5109, 2025

    2025

  52. [53]

    Z. Li, Z. Fu, Y . Hu, Z. Chen, H. Wen, and L. Nie. Finecir: Explicit parsing of fine-grained modification semantics for composed image retrieval.https://arxiv.org/abs/2503.21309, 2025

    work page arXiv 2025

  53. [54]

    Z. Li, Y . Hu, Z. Chen, Q. Huang, G. Qiu, Z. Fu, and M. Liu. Retrack: Evidence-driven dual-stream directional anchor calibration network for composed video retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 23373–23381, 2026

    2026

  54. [55]

    Z. Li, Y . Hu, Z. Chen, M. Zhang, Z. Fu, and L. Nie. Conesep: Cone-based robust noise-unlearning compositional network for composed image retrieval, 2026. URL https://arxiv.org/abs/2604. 20358

    2026

  55. [56]

    Z. Li, Y . Hu, Z. Chen, S. Zhang, Q. Huang, Z. Fu, and Y . Wei. Habit: Chrono-synergia robust progressive learning framework for composed image retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 6762–6770, 2026

    2026

  56. [57]

    Z. Li, Y . Hu, Z. Fu, Z. Chen, Y . Li, and L. Nie. Tema: Anchor the image, follow the text for multi- modification composed image retrieval, 2026. URLhttps://arxiv.org/abs/2604.21806

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  57. [58]

    Lightman, V

    H. Lightman, V . Kosaraju, Y . Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe. Let’s verify step by step. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024

    2024

  58. [59]

    J. Liu, T. Li, S. Du, X. Luo, H. Zeng, E. Tewolde, T. S. Lee, T. Wang, C. Kingsford, and V . Conitzer. The memory curse: How expanded recall erodes cooperative intent in llm agents.arXiv preprint arXiv:2605.08060, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  59. [60]

    Y . Liu, J. Zhang, and Y . Dai. Analyst following and greenwashing decision.Finance Research Letters, 58: 104510, 2023

    2023

  60. [61]

    Y . Liu, Z. Cheng, and L. Lai. Improving the completeness and comparability of segment disclosures: A large language model approach.Available at SSRN 6720239, 2026

    2026

  61. [62]

    Y . Lou, H. Hu, S. Ma, Z. Zhang, L. Wang, J. Ge, and X. Tao. Drf: Llm-agent dynamic reputation filtering framework. In T. Taniguchi, C. S. A. Leung, T. Kozuno, J. Yoshimoto, M. Mahmud, M. Doborjeh, and K. Doya, editors,Neural Information Processing, pages 127–141, Singapore, 2026. Springer Nature Singapore. ISBN 978-981-95-4384-7. 15

    2026

  62. [63]

    Lu and T

    K. Lu and T. M. Lab. On-policy distillation.Thinking Machines Lab: Connectionism, 2025. doi: 10.64434/tml.20251026. https://thinkingmachines.ai/blog/on-policy-distillation

    work page doi:10.64434/tml.20251026 2025

  63. [64]

    X. Luo, Y . Huang, H. Zeng, Y . Tao, X. Bao, F. Feng, A. L. Hopkirk, T. Pham, T. S. Bate, D. C. Saunders, et al. Cellink: integrating single-cell multi-omics data with weak feature linkage and imbalanced cell populations.Nucleic Acids Research, 53(22):gkaf1270, 2025

    2025

  64. [65]

    H. Meng, K. Huang, S. Wei, C. Ma, S. Yang, X. Wang, G. Wang, B. Ding, and J. Zhou. Sparse but critical: A token-level analysis of distributional shifts in rlvr fine-tuning of llms.ArXiv preprint, 2026

    2026

  65. [66]

    P. Qian, S. Wang, X. Wang, Y . Chen, W. Xu, Q. Yu, S. Lin, S. Zhang, J. You, and X. Wei. Relevant is not warranted: Evidence-force calibration for cited rag, 2026. URL https://arxiv.org/abs/2605.28044

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  66. [67]

    Shenfeld, M

    I. Shenfeld, M. Damani, J. Hübotter, and P. Agrawal. Self-distillation enables continual learning.ArXiv preprint, 2026

    2026

  67. [68]

    Z. Tan, D. Li, S. Wang, A. Beigi, B. Jiang, A. Bhattacharjee, M. Karami, J. Li, L. Cheng, and H. Liu. Large language models for data annotation and synthesis: A survey. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930–957, 2024

    2024

  68. [69]

    Y . Tong, D. Li, S. Wang, Y . Wang, F. Teng, and J. Shang. Can llms learn from previous mistakes? investigating llms’ errors to boost for reasoning. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3065–3080, 2024

    2024

  69. [70]

    S. Wang, L. Yu, C. Gao, C. Zheng, S. Liu, R. Lu, K. Dang, X.-h. Chen, J. Yang, Z. Zhang, Y . Liu, A. Yang, A. Zhao, Y . Yue, S. Song, B. Yu, G. Huang, and J. Lin. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning. InarXiv.org, 2025. doi: 10.48550/arxiv.2506.01939

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.01939 2025

  70. [71]

    S. Wang, L. Yu, C. Gao, C. Zheng, S. Liu, R. Lu, K. Dang, X.-H. Chen, J. Yang, Z. Zhang, et al. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  71. [72]

    Y . Wang, L. Yuan, Z. Chen, W. Zhang, X. Lin, and Q. Liu. Towards efficient shortest path counting on billion-scale graphs. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 2579–2592. IEEE, 2023

    2023

  72. [73]

    Y . Wang, L. Yuan, W. Zhang, X. Li, Z. Chen, and Q. Liu. Simpler is more: Efficient top-k nearest neighbors search on large road networks.Proc. VLDB Endow., 17(13):4683–4695, 2024

    2024

  73. [74]

    C.-C. Wu, Z. R. Tam, C.-Y . Lin, H.-y. Lee, and Y .-N. Chen. Mitigating forgetting in llm fine-tuning via low-perplexity token learning.ArXiv preprint, 2025

    2025

  74. [75]

    X. Wu, D. Jiang, F. Yu, Y . Tian, J. Tang, Q.-G. Chen, Y . Yang, and J. Lu. Adaptive debiasing tsallis entropy for test-time adaptation.arXiv preprint arXiv:2602.11743, 2026

    work page arXiv 2026

  75. [76]

    X. Xiao, B. Xia, B. Yang, B. Gao, B. Shen, C. Zhang, C. He, C. Lou, F. Luo, G. Wang, G. Xie, H. Zhang, H. Lv, H. Li, H. Chen, H.-M. Xu, H. Zhang, H. Liu, J. Duo, J. Wei, J. Xiao, J. Dong, J.-M. Shi, J. Hu, K. Bao, K. Zhou, L. Li, L. Zhao, L. Zhang, P. Li, Q. Chen, S.-y. Liu, S.-l. Yu, S. Cao, S. Chen, S. Yu, S. Liu, T.-Y . Zhou, W. Su, W. Wang, W. Ma, X. ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.02780 2026

  76. [77]

    Z. Xie, C. Wang, Y . Wang, S. Cai, S. Wang, and T. Jin. Chat-driven text generation and interaction for person retrieval. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5259–5270, 2025

    2025

  77. [78]

    Z. Xie, X. Liu, B. Zhang, Y . Lin, S. Cai, and T. Jin. Hvd: Human vision-driven video representation learning for text-video retrieval.arXiv preprint arXiv:2601.16155, 2026

    work page arXiv 2026

  78. [79]

    Z. Xie, B. Zhang, Y . Lin, and T. Jin. Delving deeper: Hierarchical visual perception for robust video-text retrieval.arXiv preprint arXiv:2601.12768, 2026. 16

    work page arXiv 2026

  79. [80]

    N. Xu, Y . Jiang, S. R. Dipta, and Z. Hengyuan. Learning how to use tools, not just when: Pattern-aware tool-integrated reasoning.MATH-AI @ NeurIPS 2025, 2025

    2025

  80. [81]

    Y . Xu, H. Sang, Z. Zhou, R. He, Z. Wang, and A. Geramifard. Tip: Token importance in on-policy distillation, 2026

    2026

Showing first 80 references.