Pith. sign in

REVIEW 3 major objections 6 minor 27 references

A selective on-policy distillation method that trains a small vision-language robot on only the teacher's uncertain tokens beats standard fine-tuning and distillation baselines in social navigation.

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 ↗

SOPD claims selective entropy-based on-policy distillation improves lightweight VLM social navigation, but its entropy mask is mathematically vacuous as printed.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Sensible distillation idea, but Eq. (5) as written makes the selective mask a no-op — the ablation contradicts the math, so the paper needs correction and code before it can be taken seriously. the 3 major comments →

arxiv 2607.19850 v1 pith:KAUIX55Q submitted 2026-07-22 cs.RO

SOPD-SocialNav: Selective On-Policy Distillation for Vision-Language Social Navigation

Xinyu Zhang , Zishuo Wang , Ling Xiao This is my paper
classification cs.RO
keywords vision-language modelsocial navigationknowledge distillationon-policy distillationentropy-based token selectionJensen-Shannon divergencelightweight VLMhuman-aware navigation
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

This paper argues that when a large vision-language model (VLM) teaches a smaller VLM how to navigate socially, the student learns best if it is trained only on the teacher's high-uncertainty tokens — the places where the teacher weighs multiple plausible actions — rather than on all tokens equally. The proposed method, SOPD, has the student generate its own responses, asks a frozen teacher for token-level probabilities, filters tokens by teacher entropy, and aligns the two distributions with a temperature-controlled Jensen-Shannon divergence. Across two social-navigation benchmarks, the authors report consistent gains over supervised fine-tuning, off-policy distillation, and standard on-policy distillation in action accuracy and semantic similarity of perception and reasoning. If true, a 3-billion-parameter VLM could carry much of the social-reasoning capability of an 8-billion-parameter teacher while running on a modest robot. The core mechanism is one equation: an entropy threshold that decides which tokens generate training signal.

Core claim

The paper's central claim is that distilling only the tokens where a frozen teacher VLM is uncertain — via temperature-softened Jensen-Shannon divergence, on responses the student generated itself — transfers social-navigation knowledge better than uniform-token distillation. On SNEI and MUSON, the authors report action accuracy of 0.889 and 0.953, the best perception and reasoning cosine similarities among all baselines, and a real robot that bypasses a face-to-face conversation and a queue instead of cutting through. The intended conclusion is that a lightweight VLM can become socially aware by concentrating training on the teacher's high-uncertainty decisions rather than diluting the grad

What carries the argument

The load-bearing mechanism is the entropy-gated token mask coupled with a temperature-controlled Jensen-Shannon divergence. The student samples a response under its current policy; the frozen teacher scores each token; the teacher's softened softmax entropy H_i is compared with a threshold τ; only tokens with H_i > τ contribute to the distillation loss. High-entropy tokens are assumed to correspond to socially ambiguous choices — yielding, avoiding, negotiating a gap — while low-entropy tokens are routine motion that would otherwise dominate the gradient. The temperature λ sharpens or flattens both distributions and stabilizes the JSD alignment.

Load-bearing premise

The entire claimed advantage rests on the entropy mask actually selecting high-uncertainty tokens; but as written, Eq. (5) lacks the minus sign that makes entropy non-negative, so every token has H_i ≤ 0 and the threshold τ=1.5 never admits a single token — the SOPD loss in Eq. (7) is identically zero unless the implementation differs from the text.

What would settle it

Check the implementation against Eqs. (5)-(7): compute H_i = Σ_k P(i) log P(i) for any teacher softmax — it is ≤ 0 — so with τ=1.5 the mask is always false and the selective loss contributes nothing. If the reported SNEI/MUSON gains (0.889/0.953 action accuracy) are real, the code must have a minus sign, a reversed inequality, or a different threshold. Re-running the ablation of Table II with the corrected entropy definition would settle whether the gains come from the entropy filter or from the JSD objective alone.

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

If this is right

  • A 3-billion-parameter VLM distilled this way can run at 0.984 frames per second on a wheeled robot, making social reasoning feasible on-device.
  • Masking routine tokens prevents long stretches of trivial forward motion from swamping the few socially critical decisions in a trajectory.
  • Temperature-controlled JSD outperforms reverse-KL and weighted-JSD objectives in the ablation, indicating that mode-seeking collapse is a real risk that this objective avoids.
  • Entropy filtering and the JSD objective are complementary: each alone improves over SFT, and their combination yields the largest gains, per the paper's ablation.

Where Pith is reading between the lines

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

  • Editorial extension: if the mechanism is sound, entropy-gated on-policy distillation should generalize to other sequential embodied tasks — manipulator grasp selection, lane merging, or crowd crossing — where a small number of tokens carry most of the social or safety cost.
  • Editorial extension: the real-world check is qualitative and limited to two scenarios; a stronger test would measure passing distance and pedestrian reactions while ablating the entropy gate.
  • Editorial extension: the fixed threshold τ suggests an adaptive variant that scales with the entropy distribution of each episode, so socially rich scenes contribute more tokens than empty corridors.
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

3 major / 6 minor

Summary. Zhang et al. propose SOPD-SocialNav, a method for distilling a large vision-language model (Qwen3-VL-8B) into a lightweight student (Qwen2.5-VL-3B) for social robot navigation. After supervised fine-tuning of the teacher on a social-navigation dataset, the student generates on-policy response sequences; the teacher and student produce token-level distributions. An entropy mask defined on teacher distributions selects tokens with entropy above a threshold, and a temperature-scaled Jensen-Shannon divergence loss is applied to those tokens. The paper reports action accuracy, perception cosine similarity, and reasoning cosine similarity on SNEI and MUSON, and presents qualitative real-world deployments on a Scout Mini robot.

Significance. The potential significance is real: if SOPD works as claimed, it offers a practical recipe for compressing large VLMs for socially compliant navigation, and the idea of focusing distillation on high-uncertainty decision tokens is timely. The benchmark comparisons and ablations address a problem of growing interest. However, the central selection mechanism is written in a way that cannot produce any distillation signal, and the empirical evidence lacks variance estimates and is accompanied by no code or data release. Consequently, the current manuscript does not yet substantiate its headline claims.

major comments (3)
  1. [III-B, Eq. (5)-(7)] The entropy-based selection mechanism is inoperative as written. Eq. (5) defines H_i = Σ_k \tilde P_t(i) log \tilde P_t(i) without a minus sign. For any valid probability distribution, every term p log p ≤ 0, so H_i ≤ 0 for all tokens. With threshold τ = 1.5, the indicator in Eq. (6) is false for every token, and the SOPD loss in Eq. (7) is identically zero on every minibatch. This directly contradicts Table II, where adding the mask (Ours vs. w/o m_i) changes MUSON Action-Acc from 0.893 to 0.953. Either Eq. (5) has a sign typo that must be corrected and documented, or the implementation differs from the text. No code, seeds, or per-token selection statistics are provided to resolve this ambiguity. Because the entire claimed advantage over standard OPD rests on this selection mechanism, the issue is load-bearing.
  2. [Tables I–III] The empirical comparison reports no error bars, confidence intervals, or multiple-seed results. The claim that SOPD 'consistently outperforms' baselines is not supported by a margin as small as 0.002 (MUSON Per-cos in Table I: 0.938 vs. 0.936, as acknowledged in Sec. IV-B). The data come from a single random 8:2 split (Sec. IV-A), so the observed differences may be within split-to-split variation. Please report mean±std over at least three seeds and, ideally, multiple data splits. This is central to the paper's empirical claim.
  3. [IV-D and Eq. (4)] The 'Weighted JSD' baseline in Table III is not defined. Eq. (4) is the standard Jensen-Shannon divergence with a uniform mixture, M = (\tilde P_t + \tilde P_s)/2, which is not a 'weighted' variant in any conventional sense. The text says 'The weighted Jensen–Shannon divergence used in our objective follows the formulation of OPD [24]', but the formula is not reproduced. 'Temperature JSD' is also not explicitly distinguished; temperature scaling in Eqs. (2)-(3) applies to all divergence variants. Please provide explicit definitions of every objective compared in Table III, or the divergence ablation is not interpretable.
minor comments (6)
  1. [III-B, Eq. (5)] Index mismatch: the sum runs over k but the term contains \tilde P_t(i) instead of \tilde P_t(k). Please correct the notation.
  2. [II-A] Typo: 'DW A' should be 'DWA' (Dynamic Window Approach).
  3. [Tables II and III] The captions should specify the dataset (MUSON, presumably) and clarify whether the SNEI results are omitted. The reader is left to infer this from the text.
  4. [IV-A] Please clarify what the SFT baseline is: is it the student model fine-tuned with supervised instruction tuning on the dataset? The description of the teacher being SFT-trained could be conflated with the baseline.
  5. [IV-E] The real-world evaluation is qualitative and based on two scenarios (Figs. 3 and 4). Adding quantitative metrics such as clearance distance, success rate, or a human evaluation would substantially strengthen the deployment claim.
  6. [Fig. 1] The equations embedded in Figure 1 are garbled in the provided manuscript; please ensure all subscripts and symbols are legible.

Circularity Check

0 steps flagged

No substantive circularity: the central claim is an empirical comparison, and the self-citations (MUSON, MAction-SocialNav, etc.) are benchmarks/related work, not load-bearing justifications. The Eq. (5) sign issue is a correctness/reproducibility problem, not circularity.

full rationale

I walked the paper's derivation chain. The method is defined by Eqs. (1)-(7): on-policy sampling, temperature-softened teacher/student distributions, JSD alignment, and an entropy-based binary mask. None of these quantities is defined in terms of the SNEI/MUSON metrics (Acc, Per-cos, Rea-cos), and no parameter is fitted to the evaluation labels to produce the reported predictions. The claimed improvements are empirical results from Table I, not derived identities. The paper's own-group citations — MUSON [26], MAction-SocialNav [13], SocialNav-MoE [14], E-SocialNav [15] — are used as benchmarks and related work; they do not supply the mathematical justification for SOPD and are not invoked as a uniqueness theorem or an ansatz that forces the method. Therefore the central claim has independent empirical content. Separately, Eq. (5) writes H_i = Σ_k \tilde P_t(i) log \tilde P_t(i) without a minus sign, so for a probability distribution H_i ≤ 0, and with τ = 1.5 the mask in Eq. (6) is always false, making L_SOPD in Eq. (7) identically zero as written. This is an internal inconsistency that makes the reported ablation gains unreproducible from the text, but it is not circularity: the loss is not equal to its input by construction, and the error is a sign/implementation bug rather than a definitional reduction. Thus no circular step is exhibited, and the score reflects only the minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The method introduces no new physical entities; its load-bearing free choices are the temperature and threshold, and the key axiom is that teacher entropy marks social decisions. The sign error in Eq. (5) makes the threshold axiom untestable as printed.

free parameters (2)
  • Temperature λ = 2.0
    Set in §IV-A to soften teacher/student logits in Eqs. (2)-(3); no sensitivity analysis, and the ablation compares JSD vs other divergences, not other λ values.
  • Entropy threshold τ = 1.5
    Set in §IV-A and used in Eq. (6); controls which tokens contribute. No sensitivity sweep, and it is inconsistent with Eq. (5) as written.
axioms (3)
  • standard math Shannon entropy is non-positive for p log p without a minus sign; the paper's H_i as written cannot exceed τ = 1.5.
    Eq. (5) omits the negative sign; this is the load-bearing failure of the selection mechanism.
  • domain assumption Teacher predictive entropy is a valid proxy for social informativeness of decision tokens.
    Introduced in §III-B without analysis of which tokens are selected or dropped; high entropy can also mark function words or noise.
  • domain assumption On-policy student-generated responses approximate the deployment distribution.
    Standard OPD premise adopted in §III-B; not tested against off-policy distillation in the ablations.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SOPD-SocialNav: Selective On-Policy Distillation for Vision-Language Social Navigation." pith.science (2026). https://pith.science/paper/KAUIX55Q

@misc{pith2026260719850,
  author       = {Pith},
  title        = {Pith review of: SOPD-SocialNav: Selective On-Policy Distillation for Vision-Language Social Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KAUIX55Q}},
  note         = {Machine review of arXiv:2607.19850}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-language models have shown strong potential for social robot navigation by leveraging rich semantic understanding of complex environments and human behaviors. However, large scale VLMs are difficult to deploy on resource-constrained robotic platforms, while lightweight VLMs often lack sufficient social reasoning capability. To address this problem, we propose SOPD-SocialNav, a selective on-policy distillation (SOPD) method that transfers social navigation knowledge from a large teacher VLM to a lightweight student VLM. SOPD introduces an entropy-based token selection mechanism that uses teacher uncertainty to identify socially informative decision tokens, while suppressing gradients from low-entropy tokens corresponding to trivial navigation states. A temperature-controlled Jensen-Shannon divergence objective is then used to align the student and teacher distributions on the selected tokens. Experiments on the SNEI and MUSON benchmarks demonstrate that SOPD consistently outperforms supervised fine-tuning, off-policy distillation, and standard on-policy distillation baselines in action prediction, perception consistency, and reasoning consistency. Real-world deployment on a Scout Mini robot further shows that the distilled model can generate more socially appropriate navigation behaviors in conversational and queuing scenarios. These results suggest that SOPD is an effective strategy for building lightweight yet socially aware VLM-based navigation systems.

Figures

Figures reproduced from arXiv: 2607.19850 by Ling Xiao, Xinyu Zhang, Zishuo Wang.

Figure 1
Figure 1. Figure 1: Overview of SOPD. The proposed SOPD framework takes text–image pairs from social navigation datasets as input. After supervised fine-tuning, the teacher VLM is frozen, while the student interacts with the environment and generates responses under its current policy. The frozen teacher then evaluates the student responses to obtain token-level probability distributions. Based on these distributions, Entropy… view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison with other methods. SOPD can better capture [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison in a real-world conversational scenario. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison in a real-world queuing scenario. Two [PITH_FULL_IMAGE:figures/full_fig_p006_4.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

27 extracted references · 6 linked inside Pith

  1. [1]

    Social force model for pedestrian dynam- ics,

    D. Helbing and P. Molnar, “Social force model for pedestrian dynam- ics,”Physical review E, vol. 51, no. 5, p. 4282, 1995

    1995

  2. [2]

    Social robot navi- gation with adaptive proxemics based on emotions,

    B. Bilen, H. Kivrak, P. Uluer, and H. Kose, “Social robot navi- gation with adaptive proxemics based on emotions,”arXiv preprint arXiv:2401.17663, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Motion planning in dynamic environments using velocity obstacles,

    P. Fiorini and Z. Shiller, “Motion planning in dynamic environments using velocity obstacles,”The international journal of robotics re- search, vol. 17, no. 7, pp. 760–772, 1998

    1998

  4. [4]

    The dynamic window approach to collision avoidance,

    D. Fox, W. Burgard, and S. Thrun, “The dynamic window approach to collision avoidance,”IEEE robotics & automation magazine, vol. 4, no. 1, pp. 23–33, 1997

    1997

  5. [5]

    Learning social navigation from demonstra- tions with deep neural networks,

    Y . Yildirim and E. Ugur, “Learning social navigation from demonstra- tions with deep neural networks,”arXiv preprint arXiv:2404.11246, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Socially aware motion planning with deep reinforcement learning,

    Y . F. Chen, M. Everett, M. Liu, and J. P. How, “Socially aware motion planning with deep reinforcement learning,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 1343–1350

    2017

  7. [7]

    Vlm-social-nav: Socially aware robot navigation through scoring us- ing vision-language models,

    D. Song, J. Liang, A. Payandeh, A. H. Raj, X. Xiao, and D. Manocha, “Vlm-social-nav: Socially aware robot navigation through scoring us- ing vision-language models,”IEEE Robotics and Automation Letters, vol. 10, no. 1, pp. 508–515, 2024

    2024

  8. [8]

    A survey of on-policy distillation for large language models,

    M. Song and M. Zheng, “A survey of on-policy distillation for large language models,”arXiv preprint arXiv:2604.00626, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Minillm: Knowledge distil- lation of large language models,

    Y . Gu, L. Dong, F. Wei, and M. Huang, “Minillm: Knowledge distil- lation of large language models,” inProceedings of the International Conference on Learning Representations (ICLR), vol. 2024, 2024, pp. 32 694–32 717

    2024

  10. [10]

    K. P. Murphy,Machine learning: a probabilistic perspective. MIT press, 2012

    2012

  11. [11]

    Social-llava: Enhancing social robot navigation through human-language reasoning,

    A. Payandeh, D. Song, M. Nazeri, J. Liang, P. Mukherjee, A. H. Raj, Y . Kong, D. Manocha, and X. Xiao, “Social-llava: Enhancing social robot navigation through human-language reasoning,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 17 192–17 198

    2025

  12. [12]

    Autospatial: Visual-language reasoning for social robot navigation through efficient spatial reasoning learning,

    Y . Kong, D. Song, J. Liang, D. Manocha, Z. Yao, and X. Xiao, “Autospatial: Visual-language reasoning for social robot navigation through efficient spatial reasoning learning,” inProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 11 298–11 304

    2025

  13. [13]

    Maction-socialnav: Multi-action socially compliant navi- gation via reasoning-enhanced prompt tuning,

    Z. Wang, X. Zhang, Z. Liu, T. Kawabata, D. Song, X. Xiao, and L. Xiao, “Maction-socialnav: Multi-action socially compliant navi- gation via reasoning-enhanced prompt tuning,”IEEE Robotics and Automation Letters, pp. 1–8, 2026

    2026

  14. [14]

    Socialnav-moe: A mixture-of- experts vision language model for socially compliant navigation with reinforcement fine-tuning,

    T. Kawabata, X. Zhang, and L. Xiao, “Socialnav-moe: A mixture-of- experts vision language model for socially compliant navigation with reinforcement fine-tuning,”arXiv preprint arXiv:2512.14757, 2025

    arXiv 2025

  15. [15]

    E-socialnav: Efficient socially compliant navigation with language models,

    L. Xiao, D. Song, X. Xiao, and T. Yamasaki, “E-socialnav: Efficient socially compliant navigation with language models,” inProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 20 077–20 081

    2026

  16. [16]

    Distilling the knowledge in a neural network,

    G. Hinton, “Distilling the knowledge in a neural network,” inProceed- ings of the Deep Learning and Representation Learning Workshop in Conjunction with NeurIPS, 2014

    2014

  17. [17]

    Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer,

    S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer,” inProceedings of the International Conference on Learning Representations (ICLR), 2017

    2017

  18. [18]

    Sequence-level knowledge distillation,

    Y . Kim and A. M. Rush, “Sequence-level knowledge distillation,” in Proceedings of the 2016 conference on empirical methods in natural language processing (EMNLP), 2016, pp. 1317–1327

    2016

  19. [19]

    Distilling step-by-step! outper- forming larger language models with less training data and smaller model sizes,

    C.-Y . Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y . Fujii, A. Ratner, R. Krishna, C.-Y . Lee, and T. Pfister, “Distilling step-by-step! outper- forming larger language models with less training data and smaller model sizes,” inProceedings of the Findings of the Association for Computational Linguistics: ACL 2023, 2023, pp. 8003–8017

    2023

  20. [20]

    Vl2lite: Task-specific knowledge distil- lation from large vision-language models to lightweight networks,

    J. Jang, C. Ma, and B. Lee, “Vl2lite: Task-specific knowledge distil- lation from large vision-language models to lightweight networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 30 073–30 083

    2025

  21. [21]

    Self-instruct: Aligning language models with self- generated instructions,

    Y . Wang, Y . Kordi, S. Mishra, A. Liu, N. A. Smith, D. Khashabi, and H. Hajishirzi, “Self-instruct: Aligning language models with self- generated instructions,” inProceedings of the 61st annual meeting of the association for computational linguistics (ACL), 2023, pp. 13 484– 13 508

    2023

  22. [22]

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe,

    Y . Li, Y . Zuo, B. He, J. Zhang, C. Xiao, C. Qian, T. Yu, H.-a. Gao, W. Yang, Z. Liuet al., “Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe,” in Proceedings of the ICML 2026 Workshop on F oundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning

    2026

  23. [23]

    Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling,

    W. Xu, R. Han, Z. Wang, L. Le, D. Madeka, L. Li, W. Wang, R. Agar- wal, C.-Y . Lee, and T. Pfister, “Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling,” in Proceedings of the International Conference on Learning Representa- tions (ICLR), vol. 2025, 2025, pp. 64 616–64 646

    2025

  24. [24]

    On-policy distillation of language models: Learning from self-generated mistakes,

    R. Agarwal, N. Vieillard, Y . Zhou, P. Stanczyk, S. Ramos Garea, M. Geist, and O. Bachem, “On-policy distillation of language models: Learning from self-generated mistakes,” inProceedings of the Inter- national Conference on Learning Representations (ICLR), vol. 2024, 2024, pp. 21 246–21 263

    2024

  25. [25]

    Self-distilled reasoner: On-policy self-distillation for large language models,

    S. Zhao, Z. Xie, M. Liu, J. Huang, G. Pang, F. Chen, and A. Grover, “Self-distilled reasoner: On-policy self-distillation for large language models,”arXiv preprint arXiv:2601.18734, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    Mu- son: A reasoning-oriented multimodal dataset for socially compliant navigation in urban environments,

    Z. Liu, X. Zhang, Z. Wang, T. Kawabata, X. Xiao, and L. Xiao, “Mu- son: A reasoning-oriented multimodal dataset for socially compliant navigation in urban environments,”arXiv preprint arXiv:2512.22867, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Geet al., “Qwen3-vl technical report,”arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.