Pith. sign in

REVIEW 2 major objections 5 minor 39 references

Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning

T0 review · 2 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Turn-level hindsight gaps, reweighted to a mean-one trajectory profile, beat the rollout-relative baseline in all eight tests.

desk verdict Clean incremental method with a sensible allocation rule; the empirical claim is plausible but single-seed evidence makes the 'all eight' headline premature. read the letter →

arxiv 2608.07371 v1 pith:Z75IBECO submitted 2026-08-07 cs.LG cs.CL

Haoyu Zheng , Yun Zhu , Qing Wang , Wenqiao Zhang This is my paper · ORCID
classification cs.LGcs.CL
keywords agenticreinforcementlearningmulti-turnLLMagentshindsightdistillationtrajectory-relativecreditassignmentGRPOsparserewardWebShopALFWorld
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TRIAL addresses a granularity problem in training multi-turn language agents from sparse outcome rewards: one completed rollout contains many possible hindsight signals, but nothing in the reward says how much weight each decision turn deserves. The paper's answer is to make the allocation trajectory-relative: for each turn, score the same realized response under an ordinary context and under a context augmented with that turn's realized consequence, then take the signed log-probability gap as token-level revision. Turn-level mean absolute gaps are normalized across the trajectory so the multipliers have an eligible-token-weighted mean of one, redistributing dense supervision toward turns whose policy assessment changes most under hindsight while keeping the average multiplier fixed. If the paper is right, this gives multi-turn agent training a cheap, training-only way to calibrate dense supervision; the reported evidence is higher success than the GRPO baseline in all eight backbone/environment/metric combinations, including a WebShop success rise from 56.4% to 75.2%.

What carries the argument

The load-bearing object is the trajectory-relative turn profile built from a frozen scoring snapshot. For turn $k$, the outcome view $z_k = \mathrm{OutcomeView}(\tau,k)$ exposes the realized post-action consequence, and the augmented prompt $P_k^+ = \mathrm{Aug}(P_k,z_k)$ scores the same recorded response tokens; eligible tokens are generated response tokens, with prompt and padding tokens masked out. The token gap $\Delta_{k,t}$ is a clipped, stop-grad log-probability difference between the hindsight-conditioned and ordinary contexts. The turn score $s_k = \frac{1}{n_k}\sum_t m_{k,t}|\Delta^{\mathrm{old}}_{k,t}|$ aggregates absolute gaps, and the weight $w_k = s_k/s_\tau$ with $s_\tau = \frac{1}{N_\tau}\sum_j n_j s_j$ satisfies $\sum_k n_k w_k = N_\tau$, giving the multipliers an eligible-token-weighted mean of one. That identity is what makes the method a redistribution rather than a rescaling: it reallocates dense supervision across turns while fixing the average multiplier, and the signed gaps still set update direction and local strength.

What would settle it

A decisive check would be to compute, across a batch of rollouts, the correlation between each turn's absolute hindsight gap and a counterfactual estimate of that turn's causal contribution to the final reward; if the correlation is not positive—or if, on a task with deliberately inserted high-gap but causally irrelevant turns, TRIAL falls below the uniform-multiplier control—the trajectory-relative profile is not carrying the claimed benefit.

Watch

Extended reading notes

Core claim

The central discovery is that hindsight supervision for a completed multi-turn rollout should be allocated by comparing each turn's revision mass with its token share, not by applying dense feedback uniformly or by selecting turns ad hoc. In TRIAL, the realized response at turn $k$ is scored both with the ordinary history and with a hindsight-augmented prompt that exposes the post-action consequence; the clipped, stop-gradient log-probability gap $\Delta_{k,t}$ fixes the sign and local strength of each token update. Aggregating absolute gaps gives $s_k$, and normalizing by the trajectory average $s_\tau$ gives $w_k = s_k/s_\tau$, so the turn weights satisfy $\sum_k n_k w_k = N_\tau$ and have eligible-token-weighted mean one. The actor loss uses $w_k \Delta_{k,t}$ as a dense term added to the unchanged GRPO objective, with a relative clamp so the dense contribution cannot dominate the outcome gradient. The paper's evidence for this mechanism is that TRIAL outperforms GRPO on every reported backbone/environment/metric combination, beats or ties the best of five other methods in six of eight aggregates, and beats both a unit-multiplier control and a shuffled-multiplier control on all four controlled WebShop/ALFWorld metrics.

Load-bearing premise

The load-bearing premise is that a turn's mean absolute log-probability gap between hindsight-conditioned and ordinary scoring is a trustworthy measure of that turn's revision evidence; the paper itself notes the score measures the magnitude of a hindsight-conditioned policy discrepancy, not the causal importance of the turn to the final reward.

Editorial extensions

If this is right

  • Sparse outcome rewards no longer force a choice between no intermediate feedback and heuristic dense feedback: a completed trajectory itself supplies calibrated turn-level supervision.
  • The method adds no inference-time cost, because all hindsight views, gaps, and profiles are discarded after training and deployment keeps the ordinary policy.
  • The mean-one profile makes the approach a pure reallocation, so changes in performance are attributable to where supervision is concentrated rather than to an overall loss-scale change.
  • The profile's increasing dispersion during training suggests the policy learns to concentrate hindsight on a shrinking set of decisive turns, which could inform trajectory-length and budgeting decisions.
  • The WebShop success rate and dense task score both improve, indicating the gain is not only about completing more tasks but about better partial credit on unfinished ones.

Reading between the lines

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

  • The same budget-preserving normalization could be applied to other dense revision signals, such as step-level process rewards or learned token-importance scores; the paper tests it only on hindsight-conditioned log-probability gaps.
  • Because the turn score is explicitly not a causal importance measure, a natural next test is to combine or compare it with counterfactual turn-removal estimates; the paper leaves that combination open.
  • The frozen pre-update scoring snapshot means the profile is built from the old policy at each update; one could investigate whether refreshing the scoring snapshot more frequently changes allocations early in training.
  • The two evaluated environments are text-based with discrete serializable post-action states; the protocol's portability to partially observed or continuous-action settings, where outcome views are harder to serialize, remains untested.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes TRIAL, a training-time objective for multi-turn agentic reinforcement learning that augments GRPO with dense hindsight supervision. For each turn, it computes a clipped signed log-probability gap between an ordinary context and a hindsight-conditioned context over the same realized response tokens (Eq. 4), aggregates absolute gaps into a per-turn score (Eq. 5), and normalizes these scores across the trajectory so that the eligible-token-weighted mean multiplier is one (Eq. 6). The dense loss (Eq. 7) is added to the GRPO objective with a scalar clamp relative to the detached GRPO loss magnitude (Eq. 8). Experiments on WebShop and ALFWorld with Qwen2.5-3B and Qwen3-1.7B compare TRIAL against GRPO and four hindsight-distillation baselines, plus Uniform and Permuted profile controls. The paper claims TRIAL outperforms GRPO on all eight backbone-environment-metric combinations and is best or tied-best among six methods on six of them, and that trajectory-relative allocation provides gains beyond dense hindsight alone.

Significance. If the empirical claims hold, TRIAL is a simple and well-motivated modification of dense hindsight distillation: the mean-one allocation identity is clean, the training-only hindsight pathway leaves deployment unchanged, and the Uniform/Permuted ablation design is well suited to isolate the effect of source-aligned relative allocation. The authors also release code and state their protocol in detail, which are clear strengths. The main barrier to acceptance is statistical rather than conceptual: the headline and ablation claims rest entirely on single-seed point estimates without variance reporting or significance testing, and the paper's own Limitations section acknowledges this. Given that several ALFWorld margins are within plausible binomial noise, the categorical 'all eight' claim is not yet established. The core method is defensible and the missing evidence is obtainable, so the appropriate outcome is a major revision rather than rejection.

major comments (2)
  1. [§4.2, Tables 1–2; §6] The headline claim—'TRIAL outperforms GRPO across all eight combinations'—is supported only by single-seed point estimates with no error bars, confidence intervals, or significance tests, a limitation the paper itself acknowledges in §6. The margins are not all large relative to evaluation-set noise: for example, the Qwen3-1.7B ALFWorld Seen gain is 68.6% versus 62.9% on 140 games, a difference on the order of 1.4 binomial standard errors, and several other aggregate and family-level margins are comparable to plausible seed-to-seed variation under stochastic rollouts. Because the abstract and §4.2 make a categorical universal claim, multi-seed runs with variance reporting or a paired significance test are required before the claim can be accepted. This is fixable within the paper's scope, but it is load-bearing for the main conclusion.
  2. [§4.3, Table 2; §A.2] The controlled profile study is the direct evidence that trajectory-relative allocation, rather than dense hindsight alone, drives the gains, but it is also based on single runs. The TRIAL-versus-Permuted differences are 2.2 and 4.5 points on ALFWorld Seen/Unseen (140 and 134 games) and 18.4 and 9.4 points on WebShop; the ALFWorld margins are within plausible noise for one seed, so the conclusion that 'TRIAL's consistent lead supports applying revision to the turn whose policy assessment changes' is not yet established. In addition, the Permuted procedure is underspecified: the manuscript does not state whether the permutation is per trajectory, how self-assignments and ties are handled, or which random seed is used. Multi-seed means with error bars for all four profile conditions and a precise description of the permutation are necessary to support the paper's central contribution over unit-weighted dense hindsight.
minor comments (5)
  1. [§3.2, Eq. (4)] The phrase 'let clipb(u) cliputo[−b,b]' appears to be a typo; it should read 'let clip_b(u) denote clipping u to [−b,b]'.
  2. [Table 1 caption] The caption says bold and underlining mark the best and second-best results, but the rendered table does not show underlining; please use a visible marker or revise the caption.
  3. [§4.4 and Figure 4] The labels 'INFORMATIVE DISCOVERY', 'INEFFECTIVE ACTION', and the class-averaged multipliers are interpretive and are based on 21 successful trajectories from a single batch; the SEM in Figure 4(b) is descriptive, and the text should present these as qualitative observations rather than mechanistic conclusions.
  4. [§4.1 and Table 3] No sensitivity analysis is reported for λg, c, α, or the auxiliary activation step; since the mean-one normalization is claimed to make the method robust to gap scale, a short sensitivity table would strengthen the paper.
  5. [Abstract and §4.2] The phrase 'all eight combinations' refers to eight aggregate metrics; because ALFWorld task-family columns are separate, rephrasing as 'eight aggregate metrics' would avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TRIAL's objective is a self-contained reweighting construction, and its empirical claims rest on held-out benchmarks.

full rationale

TRIAL is a training-objective design rather than a derived prediction, so the circularity tests reduce to whether any reported 'prediction' is equivalent to its inputs by construction. The turn score in Eq. 5 is defined as the eligible-token mean absolute log-probability gap between ordinary and hindsight-conditioned contexts, and Eq. 6 normalizes that score by its token-weighted trajectory average, yielding the identity sum_k n_k w_k = N_tau. This mean-one property is an explicit algebraic normalization, not a fitted parameter or an empirical discovery; the paper openly states that 'the token-weighted trajectory average removes the common gap scale,' and it presents the identity as a calibration control rather than as a result derived from hidden assumptions. The empirical claims are benchmarked on held-out evaluation sets: WebShop's complete official 500-goal test split and ALFWorld's official Seen and Unseen splits, with training-time validation on a separate fixed 128-instance diagnostic subset. Hyperparameters such as the auxiliary coefficient, gap clipping bound, and activation step are fixed in Table 3 and are not fitted to the test outcomes. The paper's own Limitations section acknowledges that 'Our reported runs use a single seed; cross-seed variance and statistical robustness remain unmeasured,' and Section 3.2 explicitly says the turn score 'measures the magnitude of a hindsight-conditioned policy discrepancy, not the causal importance of turn k to the final reward'; both are evidentiary limitations that affect confidence in the magnitude and generality of the gains, but neither is a circular step. Self-citations appear only in related-work context and are not load-bearing for the proposed objective or its evaluation. No equation in the paper reduces to its own input by definition, and no fitted quantity is renamed as a prediction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method has no invented physical entities. Its load-bearing assumptions are that hindsight-conditioned log-probability gaps are informative, that the mean absolute gap is the right turn-level aggregator, and that post-action evidence can be serialized and aligned per turn. The hyperparameters are environment-specific hand choices.

free parameters (5)
  • Auxiliary coefficient lambda_g = 0.005 (WebShop), 0.01 (ALFWorld)
    Scales the dense hindsight loss in Eq. 8; chosen by hand per environment.
  • Gap clipping bound c = 2.0
    Clipping bound for token-level log-probability gaps in Eq. 4.
  • Relative clamp alpha = 1.0
    Bounds the dense contribution relative to the detached GRPO loss magnitude in Eq. 8.
  • Auxiliary activation step = 25
    Steps 1-24 use zero dense coefficient; chosen as warmup.
  • Learning rate = 1e-6
    A standard choice shared by all methods; influences optimization but is not benchmark-specific.
assumptions (5)
  • domain assumption The signed log-probability gap under a hindsight-conditioned context is a useful token-level learning signal.
    Eq. 4 defines the signal; the whole method assumes this gap carries revision information.
  • domain assumption The turn score s_k, mean absolute gap over eligible tokens, captures the right allocation evidence.
    Eq. 5 aggregates gaps per turn; this aggregation choice is assumed, not derived.
  • domain assumption The frozen pre-update policy snapshot provides a stable scoring reference.
    Section 3.2 sets pi_T = pi_old; if the snapshot moves, the allocation profile changes.
  • domain assumption Completed trajectories expose serializable post-action evidence alignable with the decision that produced it.
    Explicitly stated in Section 6 Limitations as an assumption of the method.
  • standard math GRPO's group-relative advantage normalization provides a valid outcome-level baseline.
    Eq. 1 from Shao et al. 2024; used as the sparse outcome objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning." pith.science (2026). https://pith.science/paper/Z75IBECO

@misc{pith2026260807371,
  author       = {Pith},
  title        = {Pith review of: Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z75IBECO}},
  note         = {Machine review of arXiv:2608.07371}
}
read the original abstract

Recent agentic reinforcement learning methods use hindsight to complement sparse outcome rewards. However, a completed rollout can yield many such signals, leaving their appropriate allocation across turns unclear. We introduce TRIAL, a trajectory-relative hindsight distillation framework with a unified turn-aligned scoring protocol. For each decision turn, TRIAL extracts an outcome view of that decision's realized consequence and evaluates the same response under ordinary and hindsight-conditioned contexts. The signed log-probability gap determines the direction and local strength of token-level supervision, while turn-level magnitudes are normalized jointly over the realized trajectory. The resulting allocation multipliers have an eligible-token-weighted mean of one, redistributing dense supervision across turns while fixing its average multiplier. Experiments on WebShop and ALFWorld with different backbones show that TRIAL outperforms GRPO across all eight combinations of backbone, environment, and evaluation metric, while achieving the best or tied-best performance among six methods on six of them. On WebShop with Qwen3-1.7B, TRIAL improves the success rate from 56.4% to 75.2% and the task score from 78.7% to 85.7%. Controlled ablations further show that trajectory-relative turn allocation provides substantial gains beyond those of dense hindsight distillation alone.

Figures

Figures reproduced from arXiv: 2608.07371 by the authors.

Figure 1
Figure 1. Complementary feedback granularities in a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. TRIAL augments GRPO on the same realized trajectory. GRPO’s across-rollout comparison yields one outcome signal Abi shared by all action tokens in rollout i. A frozen training-time scoring view evaluates realized tokens with hindsight derived from action–observation consequences; signed gaps determine token direction and local strength, while a normalized turn profile determines relative allocation. The aligned sign… view at source ↗
Figure 3
Figure 3. Qwen3-1.7B validation and allocation dynamics on ALFWorld. (a) Success on the fixed 128-game [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: TRIAL’s hindsight allocation on successful ALFWorld Pick-Two trajectories. (a) An 18-turn profile and three example decisions; T15 is the acquisition following T14’s discovery. (b) Mean multipliers by decision class across 21 successful trajectories from five task inst…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 12 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems , year =

    Hindsight Experience Replay , author =. Advances in Neural Information Processing Systems , year =

  2. [2]

    Robotics: Science and Systems XIV , year =

    Asymmetric Actor Critic for Image-Based Robot Learning , author =. Robotics: Science and Systems XIV , year =. doi:10.15607/RSS.2018.XIV.008 , url =

    work page doi:10.15607/rss.2018.xiv.008 2018

  3. [3]

    2024 , eprint =

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author =. 2024 , eprint =

    2024

  4. [4]

    2024 , eprint =

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes , author =. 2024 , eprint =

    2024

  5. [5]

    2504.20073 , archivePrefix =

    Zihan Wang and Kangrui Wang and Qineng Wang and Pingyue Zhang and Linjie Li and Zhengyuan Yang and Xing Jin and Kefan Yu and Minh Nhat Nguyen and Licheng Liu and Eli Gottlieb and Yiping Lu and Kyunghyun Cho and Jiajun Wu and Li Fei-Fei and Lijuan Wang and Yejin Choi and Manling Li , year =. 2504.20073 , archivePrefix =

    arXiv

  6. [6]

    Group-in-Group Policy Optimization for

    Lang Feng and Zhenghai Xue and Tingcong Liu and Bo An , booktitle =. Group-in-Group Policy Optimization for. 2025 , url =

    work page 2025

  7. [7]

    2026 , eprint =

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models , author =. 2026 , eprint =

    2026

  8. [8]

    2026 , eprint =

    Reinforcement Learning via Self-Distillation , author =. 2026 , eprint =

    2026

Show all 39 references
  1. [9]

    Self-Distilled

    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 , year =. Self-Distilled. 2604.03128 , archivePrefix =

    arXiv

  2. [10]

    2604.10674 , archivePrefix =

    Hao Wang and Guozhi Wang and Han Xiao and Yufeng Zhou and Yue Pan and Jichao Wang and Ke Xu and Yafei Wen and Xiaohu Ruan and Xiaoxin Chen and Honggang Qi , year =. 2604.10674 , archivePrefix =

    arXiv

  3. [11]

    2604.14084 , archivePrefix =

    Yuanda Xu and Hejian Sang and Zhengze Zhou and Ran He and Zhipeng Wang and Alborz Geramifard , year =. 2604.14084 , archivePrefix =

    arXiv

  4. [12]

    2604.24005 , archivePrefix =

    Jiaqi Wang and Wenhao Zhang and Weijie Shi and Yaliang Li and James Cheng , year =. 2604.24005 , archivePrefix =

    arXiv

  5. [13]

    2026 , eprint =

    Self-Distilled Agentic Reinforcement Learning , author =. 2026 , eprint =

    2026

  6. [14]

    2605.27186 , archivePrefix =

    Haoyu Zheng and Yun Zhu and Shu Yuan and Shangming Chen and Qing Wang and Wenqiao Zhang and Jun Xiao and Yueting Zhuang , year =. 2605.27186 , archivePrefix =

    arXiv

  7. [15]

    2026 , eprint =

    What and When to Distill: Selective Hindsight Distillation for Multi-Turn Agents , author =. 2026 , eprint =

    2026

  8. [16]

    2605.27140 , archivePrefix =

    Yanfei Zhang and Xu Lin and Chenglin Wu , year =. 2605.27140 , archivePrefix =

    arXiv

  9. [17]

    2605.17873 , archivePrefix =

    Woongyeng Yeo and Yumin Choi and Taekyung Ki and Sung Ju Hwang , year =. 2605.17873 , archivePrefix =

    arXiv

  10. [18]

    2605.11853 , archivePrefix =

    Sijia Li and Yuchen Huang and Zifan Liu and Yanping Li and Jingjing Fu and Li Zhao and Jiang Bian and Ling Zhang and Jun Zhang and Rui Wang , year =. 2605.11853 , archivePrefix =

    arXiv

  11. [19]

    2606.26790 , archivePrefix =

    Shuo Yang and Jinyang Wu and Zhengxi Lu and Yuhao Shen and Fan Zhang and Lang Feng and Shuai Zhang and Haoran Luo and Zheng Lian and Zhengqi Wen and Jianhua Tao , year =. 2606.26790 , archivePrefix =

    arXiv

  12. [20]

    2606.09348 , archivePrefix =

    Yang Tian and Rui Wang and Xumeng Wen and Junjie Li and Shizhao Sun and Lei Song and Jiang Bian and Bo Zhao , year =. 2606.09348 , archivePrefix =

    arXiv

  13. [21]

    2603.18683 , archivePrefix =

    Zhicong Lu and Zichuan Lin and Wei Jia and Changyuan Tian and Deheng Ye and Peiguang Li and Li Jin and Nayu Liu and Guangluan Xu and Wei Feng , year =. 2603.18683 , archivePrefix =

  14. [22]

    2026 , eprint =

    Keep Policy Gradient in Charge: Sibling-Guided Credit Distillation for Long-Horizon Tool-Use Agents , author =. 2026 , eprint =

    2026

  15. [23]

    2606.11559 , archivePrefix =

    Haoran Liu and Yuwei Zhang and Xiyao Li and Bohan Lyu and Jingbo Shang , year =. 2606.11559 , archivePrefix =

    arXiv

  16. [24]

    2606.29476 , archivePrefix =

    Zibin Meng and Kani Chen , year =. 2606.29476 , archivePrefix =

    arXiv

  17. [25]

    doi:10.48550/arXiv.2606.27814 , url =

    Qitai Tan and Zefang Zong and Mo Li and Yipeng Shi and Yang Li and Peng Chen , year =. doi:10.48550/arXiv.2606.27814 , url =. 2606.27814 , archivePrefix =

    doi arXiv

  18. [26]

    doi:10.48550/arXiv.2606.19659 , url =

    Yuhang Zhou and Lizhu Zhang and Yifan Wu and Mingyi Wang and Bo Peng and Jiayi Liu and Xiangjun Fan and Zhuokai Zhao , year =. doi:10.48550/arXiv.2606.19659 , url =. 2606.19659 , archivePrefix =

    doi arXiv

  19. [27]

    2607.05804 , archivePrefix =

    Yuhang Zhou and Kai Zheng and Haoling Li and Dengyun Peng and Can Xu and Jingjing Chen , year =. 2607.05804 , archivePrefix =

    arXiv

  20. [28]

    2021 , eprint =

    Mohit Shridhar and Xingdi Yuan and Marc-Alexandre C. 2021 , eprint =

    2021

  21. [29]

    2022 , url =

    Shunyu Yao and Howard Chen and John Yang and Karthik Narasimhan , booktitle =. 2022 , url =

    2022

  22. [30]

    2024 , eprint =

    Qwen2.5 Technical Report , author =. 2024 , eprint =

    2024

  23. [31]

    2025 , eprint =

    Qwen3 Technical Report , author =. 2025 , eprint =

    2025

  24. [32]

    2017 , eprint =

    Proximal Policy Optimization Algorithms , author =. 2017 , eprint =

    2017

  25. [33]

    2501.00599 , archivePrefix =

    Yuqian Yuan and Hang Zhang and Wentong Li and Zesen Cheng and Boqiang Zhang and Long Li and Xin Li and Deli Zhao and Wenqiao Zhang and Yueting Zhuang and Jianke Zhu and Lidong Bing , year =. 2501.00599 , archivePrefix =

    arXiv

  26. [34]

    2510.23603 , archivePrefix =

    Yuqian Yuan and Wenqiao Zhang and Xin Li and Shihao Wang and Kehan Li and Wentong Li and Jun Xiao and Lei Zhang and Beng Chin Ooi , year =. 2510.23603 , archivePrefix =

  27. [35]

    2605.26102 , archivePrefix =

    Yuqian Yuan and Wentong Li and Zhaocheng Li and Yutong Lin and Juncheng Li and Siliang Tang and Jun Xiao and Yueting Zhuang and Wenqiao Zhang , year =. 2605.26102 , archivePrefix =

    arXiv

  28. [36]

    2403.13447 , archivePrefix =

    Wenqiao Zhang and Tianwei Lin and Jiang Liu and Fangxun Shu and Haoyuan Li and Lei Zhang and He Wanggui and Hao Zhou and Zheqi Lv and Hao Jiang and Juncheng Li and Siliang Tang and Yueting Zhuang , year =. 2403.13447 , archivePrefix =

    arXiv

  29. [37]

    2025 , url =

    Tianwei Lin and Wenqiao Zhang and Sijing Li and Yuqian Yuan and Binhe Yu and Haoyuan Li and Wanggui He and Hao Jiang and Mengze Li and Xiaohui Song and Siliang Tang and Jun Xiao and Hui Lin and Yueting Zhuang and Beng Chin Ooi , booktitle =. 2025 , url =

    2025

  30. [38]

    2605.30011 , archivePrefix =

    Mingjian Gao and Wenqiao Zhang and Yuqian Yuan and Yang Dai and Binhe Yu and Zheqi Lv and Haoyu Zheng and Jiaqi Zhu and Zhiqi Ge and Zixuan Wan and Siliang Tang and Yueting Zhuang , year =. 2605.30011 , archivePrefix =

    arXiv

  31. [39]

    Haoyu Zheng and Yun Zhu and Yuqian Yuan and Bo Yuan and Wenqiao Zhang and Siliang Tang and Jun Xiao , booktitle =

Pith tools

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