Pith. sign in

REVIEW 4 major objections 6 minor 25 references

Recall Before You Rank: Similarity-Guided Top-$K$ Reuse for Efficient Long-Context Attention

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read ReTopK speeds up dynamic Top-K attention by reusing past retrieval choices, delivering 3.07x speedup at 128K with only 0.50% perplexity loss.

desk verdict Useful recall-based top-K attention with a plausible algorithm, but the headline speedup is likely inflated by asymmetric kernel fusion; needs a matched-fused baseline and the missing reuse baselines. read the letter →

arxiv 2607.27692 v1 pith:I7FRUQEQ submitted 2026-07-30 cs.CL cs.LG

Wenshuai Yao , Wenyong Zhou , Hanyong Shao , Yizhe Chen , Zhiyuan Ning , Yuannuo Feng , Ru Huang , Kechao Tang This is my paper
classification cs.CLcs.LG
keywords efficientattentionTop-Ksparselong-contextinferenceKVcacheretrievalreusequerysimilaritytraining-freemethoddecodingacceleration
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

ReTopK targets the index-discovery bottleneck of exact Top-K attention: even when aggregation is restricted to K keys, finding those keys still requires scoring every query against the entire KV cache. The paper's central claim is that similar queries attend to overlapping key supports, so the support selected for a new query can be recalled from a small cache of previous query–support pairs, unioned with a recent-token window, and then reranked with exact current-query scores. This training-free recall-before-rerank procedure keeps the full KV cache, reuses only index selections, and makes the selector's work independent of context length, with similarity fallback and periodic refresh as safeguards. If correct, the main linear-in-context cost of sparse attention—full-history scoring and global Top-K selection—disappears from most decoding steps; the paper reports a 3.07x attention speedup at 128K context with K=512 and only a 0.50% perplexity increase over exact Top-K.

What carries the argument

The mechanism is a recall-before-rerank selector: a bounded per-head query–support cache (normalized queries plus the Top-K index sets selected at those steps) is searched for the R most similar cached queries; their supports are unioned with a recent window of W positions, and only the resulting M_t ≤ RK+W candidates are scored with the current query and reranked to the final K indices. The cache stores only index selections—never historical scores, weights, or outputs—so the complete KV cache remains intact. A similarity fallback (ρmax < τ invokes full exact Top-K) and periodic refresh (every Tr steps) are the reliability mechanisms that keep reuse from degrading.

What would settle it

Force the reuse path unconditionally (disable fallback and refresh), then measure the overlap between ReTopK's candidate set and the exact Top-K support on a held-out long-context corpus; if highly similar cached queries (cosine similarity above 0.85) recover less than roughly half the exact support on most steps, the central reuse assumption is false and the speedup will not generalize.

Watch

Extended reading notes

Core claim

The central discovery is an empirical regularity plus a design that exploits it: for each attention head, the exact Top-K support of a query is largely recoverable from the supports of a few historically cached queries that are most similar to it, plus a small window of recent positions. ReTopK maintains a bounded per-head FIFO cache of normalized queries and their selected index sets; on a reuse step it computes cosine similarity against the cache, takes the R most similar entries, unions their stored Top-K supports with the last W positions, computes exact QK scores only on that candidate set, and reranks to select exactly K indices. Because the recalled indices are only candidates, the fi

Load-bearing premise

The load-bearing premise is empirical: similar query vectors reliably attend to overlapping sets of keys, so a small cache of lookalike queries plus a recent-token window covers most of the exact Top-K support; if query drift weakens that link, fallback fires more often and the promised speedup shrinks.

Editorial extensions

If this is right

  • Index discovery, not aggregation, becomes the cost that matters: at long contexts full-history QK scoring and global Top-K selection account for 87–97% of exact Top-K latency, and ReTopK's reuse path removes both on most steps.
  • If the reuse regularity holds, sparse attention decoders can stop scanning the KV cache at every step without evicting any keys, so retrieval quality remains query-dependent rather than fixed by a streaming window.
  • The method transfers across model families and scales without retuning: 82.5–89.6% of head–token steps take the reuse path, with perplexity changes within about 2.8% of exact Top-K and speedups of 1.26–2.66x.
  • Support recall of roughly 57–79% is enough: because reranking uses exact current scores, unmatched low-weight keys contribute little, so 99%+ of attention mass and about 97% output cosine can be retained.
  • End-to-end decoding gains widen as the shared non-attention work shrinks, reaching about 3.7x at multi-million-token contexts in the paper's measurements.

Reading between the lines

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

  • If the similarity–overlap regularity is robust, ReTopK's cache could be built during chunked prefill rather than only during decoding, extending the same recall-before-rerank idea to the first pass over a long context.
  • Because reuse quality varies by head, an adaptive per-head choice of cache size, number of retrieved queries, and fallback threshold could push the speed/quality Pareto frontier beyond the fixed defaults reported here.
  • The candidate set ReTopK produces is a natural, query-informed signal for which KV entries matter; under a hard memory budget, the union of recalled supports could double as an eviction policy, a use the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes ReTopK, a training-free approximate dynamic Top-K attention method for long-context decoding. ReTopK maintains a bounded per-head FIFO cache of normalized query embeddings together with the support indices selected for those queries. At each decoding step it retrieves the R most similar cached queries by cosine similarity, unions their stored supports with a recent W-token window, scores only the resulting compact candidate set (size M_t <= RK+W), reranks to K, and performs sparse softmax-value aggregation over the reranked support. A similarity-based fallback (threshold tau) and a periodic exact refresh (interval Tr) invoke full-history Exact Top-K when reuse is deemed unreliable. The paper reports quality and efficiency results on PG19, RULER NIAH, and LongBench across Qwen2.5-7B, Llama-3.1-8B, and Qwen2.5-14B, claiming attention speedups up to 3.07x at 128K context with only a 0.50% perplexity increase over Exact Top-K. It also provides a fused GPU implementation, kernel-level ablation, sensitivity analysis, long-horizon stability experiments, and end-to-end decoding measurements.

Significance. If the speedup claim holds under matched implementations, ReTopK is a useful contribution: it targets the index-discovery bottleneck of sparse attention without evicting KV entries and without any training. The recall-before-rerank design is simple and well motivated by the query-similarity/support-overlap evidence in Figures 2--3. The empirical study is thorough: three model families, three benchmarks, component ablation, hyperparameter sensitivity, long-horizon stability, a five-stage kernel fusion analysis, and end-to-end scaling to very long contexts. The complexity analysis for the reuse path is correct and the paper is transparent about the fact that fallback and refresh steps perform full-history Exact Top-K. The main caveat is that the headline 3.07x attention speedup is measured against an Exact Top-K baseline whose kernel-fusion level is not reported; Table 4 shows that ReTopK's own fusion improvements alone account for up to 2.57x, so the claimed algorithmic speedup is not cleanly separated from implementation engineering.

major comments (4)
  1. [Implementation settings; Table 4; Figure 9] The paper states that attention latency is measured with 'matched kernels', but the Exact Top-K baseline's kernel-fusion level is never specified. Figure 9 decomposes Exact Top-K into separate components (Full-context QK, Exact Top-K select, Sparse Softmax-SV, Split-attention merge), suggesting an unfused or partially fused baseline. Table 4 shows ReTopK's own L1-to-L5 fusion improvements are 2.57x at 64K and 2.29x at 128K. Since the claimed 3.07x speedup at 128K is attention-only (end-to-end speedup in Table 5 is only 1.13x), the baseline comparison is load-bearing. Please report the exact kernel structure of the Exact Top-K baseline and remeasure against a fused Exact Top-K baseline (e.g., fused full-context QK + top-K selection + sparse softmax-V).
  2. [Methodology, Eq. (14); Figure 9] The complexity statement that a reuse step 'replaces both full-history QK scoring and global selection over L_t entries with bounded cache lookup and candidate processing' is correct only for the reuse path. On fallback and refresh steps the method executes full-history Exact Top-K with O(L_t d) work. The paper's broader phrasing that the selector work 'does not grow with context length' is therefore not a worst-case statement; the measured speedup depends on the fallback/refresh mixture, which is why ReTopK's latency still grows with context in Figure 9. Please state the amortized/expected complexity with respect to the exact-path rate and give the worst-case bound explicitly.
  3. [Experiments; Figure 7; Table 1] The default hyperparameters C=W=32, R=4, tau=0.85, Tr=128 are selected by one-at-a-time sensitivity analysis on PG19 at 64K, and the same configuration is then used for the headline 128K 0.50% PPL increase and for the NIAH/LongBench results. Because the selection and evaluation are on overlapping data, the reported operating point may be optimistically selected. The cross-model results in Table 2 mitigate this concern, but the paper should describe the selection protocol precisely (e.g., whether NIAH/LongBench were touched during tuning) and, ideally, report a validation-split or holdout model/context pair to establish out-of-sample behavior.
  4. [Figure 6; fallback design] The method's reliability rests on the empirical correlation between query cosine similarity and support overlap, but the paper does not characterize the tail behavior of reuse fidelity. Figure 6 reports average support recall of only 56.9% on reuse-path pairs with 91.6% retained mass; this is encouraging, but the paper does not state the fraction of head--token pairs with low retained mass or low head-output cosine. Since reranking cannot recover indices absent from the candidate set, and the fallback triggers only on raw similarity rather than predicted recall, there may be input segments where reuse silently degrades. Please report quantiles of retained mass/head-output cosine and a calibration-style analysis of similarity versus recall to justify the tau threshold.
minor comments (6)
  1. [Abstract; Table 1] The abstract says ReTopK achieves the 'lowest PG19 perplexity and highest NIAH and LongBench scores among the evaluated approximate methods' without specifying the configuration. With the default tau=0.85, NIAH at 16K is 85.3 versus 96.3 for tau=0.90. Please state the configuration used for the headline claim.
  2. [Table 2 versus Table 1] Cross-model results use K=1024 at 128K while the headline Table 1 uses K=512. The 128K PPL deltas in Table 2 are therefore not directly comparable to the 0.50% figure in the abstract; please label this clearly.
  3. [Figure 5 versus Figure 6] Figure 5 reports 78.9% support recall and 99.45% retained mass for one representative head, while Figure 6 reports 56.9% average recall and 91.6% mass over all heads. The caption should explicitly note that Figure 5 is a single favorable head.
  4. [Equation (13)] The first branch of the path rule uses Tr>0 and t_dec in {Tr,2Tr,...}; for Tr=0 the condition is undefined. The notation should be written so that Tr=0 disables scheduled refresh without an undefined expression.
  5. [Table 5] The 'Non-attn.' column is the shared decoder latency as a fraction of Exact Top-K latency; please define this in the caption or text, since it explains why the end-to-end speedup grows with context length.
  6. [General] No code or data availability statement appears in the paper. Given the reproducibility-sensitive nature of kernel-level speedup claims, a code release or detailed kernel pseudocode would strengthen the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ReTopK's quality and speed claims are measured against an independently computed Exact Top-K baseline, with no load-bearing self-citation or fitted quantity masquerading as a prediction.

full rationale

The paper's derivation chain is empirical and measurement-based rather than definitional. The central observation (Eqs. 1-3, Figure 2) defines Exact Top-K support, overlap, and coverage, and then measures query cosine similarity against these independently computed quantities; the relation is used as motivation, not as an output of the method. ReTopK's candidate construction (Eqs. 5-9) and reranking (Eqs. 10-11) are defined purely from cached index sets, current-query scores, and a recent window; no output quantity is defined in terms of the fitted hyperparameters. The fallback and refresh rule (Eq. 13) simply switches between the reuse path and a full Exact Top-K computation, which is independently defined in Eq. 1. Quality metrics (PG19 perplexity, NIAH, LongBench) and speedups are all reported relative to Exact Top-K as an external baseline. Hyperparameters (C, W, R, tau, Tr) are selected by sensitivity analysis on the same benchmarks, but this is tuning rather than circularity: the Exact Top-K 'ground truth' is not constructed from ReTopK's parameters. There are no self-citations that carry a load-bearing argument, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. The skeptic concern about the Exact Top-K baseline's kernel-fusion level is a measurement-fairness issue, not a circularity of the type defined here, because it does not make any ReTopK result equal to its input by construction. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method introduces no new theoretical entities; it relies on a set of empirically validated regularity assumptions about query similarity and support overlap. Its free parameters are hyperparameters selected by hand/sensitivity analysis.

free parameters (6)
  • similarity threshold tau = 0.85 (default; 0.90 also evaluated)
    Controls trade-off between reuse and exact fallback; chosen via sensitivity analysis on PG19 at 64K.
  • cache capacity C = 32
    Number of historical query–support pairs kept per head; chosen by hand via sensitivity analysis.
  • recent window W = 32
    Number of recent tokens added to candidate set; default chosen via sensitivity analysis.
  • retrieval count R = 4
    Number of most similar cached queries retrieved; default chosen via sensitivity analysis.
  • refresh interval Tr = 128
    Steps between exact refreshes; default chosen via sensitivity analysis.
  • active token budget K = 512 (also 1024 at 128K in cross-model runs)
    Number of tokens attended to after reranking; shared across baselines and selected by the authors.
assumptions (5)
  • domain assumption Similar queries attend to overlapping Top-K supports.
    Empirical observation motivating the method (Figure 2), not derived from model properties.
  • domain assumption A small bounded cache of C=32 historical queries per head is sufficient to contain similar queries for future decoding steps.
    Assumed in the cache design; validated only empirically across the tested benchmarks.
  • domain assumption Reranking the union of recalled supports plus a recent window preserves enough attention mass to keep quality near Exact Top-K.
    Central to the recall-before-rerank design; empirically measured but not proven.
  • standard math The Exact Top-K support is the appropriate ground-truth for quality.
    Exact Top-K is defined as selecting the K largest QK scores; used as the reference in all quality comparisons.
  • domain assumption The GPU kernels for the baseline and ReTopK are comparably optimized.
    Speedup numbers assume matched kernel quality; the paper states 'matched' but does not describe the baseline implementation in enough detail to verify.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recall Before You Rank: Similarity-Guided Top-$K$ Reuse for Efficient Long-Context Attention." pith.science (2026). https://pith.science/paper/I7FRUQEQ

@misc{pith2026260727692,
  author       = {Pith},
  title        = {Pith review of: Recall Before You Rank: Similarity-Guided Top-$K$ Reuse for Efficient Long-Context Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7FRUQEQ}},
  note         = {Machine review of arXiv:2607.27692}
}
abstract

Top-$K$ sparse attention reduces the cost of Softmax and value aggregation by attending to only a small subset of key--value (KV) entries. However, identifying this subset still requires scoring the current query against the full KV cache and performing global Top-$K$ selection, leaving selector cost linear in context length and limiting the practical efficiency of sparse attention for long-context decoding. In this paper, we introduce ReTopK, a training-free method that accelerates dynamic Top-$K$ attention by reusing historical retrieval decisions. ReTopK builds on the observation that similar queries often attend to overlapping supports and that partially overlapping supports can still preserve most of the Exact Top-$K$ attention mass. For each attention head, it maintains a bounded cache of historical query--support pairs, retrieves the most similar cached queries for each new query, unions their stored supports with a recent window, and reranks only the resulting compact candidate set using exact current-query scores. A similarity-based fallback invokes full-history Exact Top-$K$ when reuse is unreliable, while periodic exact refreshes limit cache drift. ReTopK retains the complete KV cache and reuses only selected indices, rather than historical scores, attention weights, or outputs. Across 16K--128K contexts, ReTopK achieves the lowest PG19 perplexity and the highest NIAH and LongBench scores among the evaluated approximate methods. At 128K with $K=512$, ReTopK incurs only a 0.50\% perplexity increase over Exact Top-$K$ while accelerating attention computation by $3.07\times$.

Figures

Figures reproduced from arXiv: 2607.27692 by the authors.

Figure 1
Figure 1. Long-context attention latency: (a) Exact Top- [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Query similarity versus Exact Top-K support over￾lap: (a) all historical pairs; (b) each query’s most similar predecessor; (c) coverage by the union of R retrieved sup￾ports. Reuse quality is not uniform across attention heads. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. (a) Full-history Exact Top-K selection. (b) ReTopK recall, candidate construction, and reranking. Dashed boxes and lines denote unselected entries and connections. Candidate construction and reranking. The supports as￾sociated with the retrieved cache entries are merged with a recent-token window, Lt = {max(1, Lt − W + 1), . . . , Lt} , (8) where W is the window size. The resulting candidate set is At = Unique [ c∈R… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: QK supports for Qwen2.5-7B layer 18, head 6 after [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: All-head ReTopK fidelity on Qwen2.5-7B for 4 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Long-horizon stability under (a) teacher forcing [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 4 linked inside Pith

  1. [1]

    Fu and Stefano Ermon and Atri Rudra and Christopher R

    Tri Dao and Daniel Y. Fu and Stefano Ermon and Atri Rudra and Christopher R. FlashAttention: Fast and Memory-Efficient Exact Attention with. Advances in Neural Information Processing Systems , volume =

  2. [2]

    International Conference on Learning Representations , year =

    Guangxuan Xiao and Yuandong Tian and Beidi Chen and Song Han and Mike Lewis , title =. International Conference on Learning Representations , year =

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    Zhenyu Zhang and Ying Sheng and Tianyi Zhou and Tianlong Chen and Lianmin Zheng and Ruisi Cai and Zhao Song and Yuandong Tian and Christopher R. Advances in Neural Information Processing Systems , volume =

  4. [4]

    Advances in Neural Information Processing Systems , volume =

    Yuhong Li and Yingbing Huang and Bowen Yang and Bharat Venkitesh and Acyr Locatelli and Hanchen Ye and Tianle Cai and Patrick Lewis and Deming Chen , title =. Advances in Neural Information Processing Systems , volume =

  5. [5]

    Second Conference on Language Modeling , year =

    Zefan Cai and Yichi Zhang and Bofei Gao and Yuliang Liu and Yucheng Li and Tianyu Liu and Keming Lu and Wayne Xiong and Yue Dong and Junjie Hu and Wen Xiao , title =. Second Conference on Language Modeling , year =

  6. [6]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Jiaming Tang and Yilong Zhao and Kan Zhu and Guangxuan Xiao and Baris Kasikci and Song Han , title =. Proceedings of the 41st International Conference on Machine Learning , series =

  7. [7]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Luka Ribar and Ivan Chelombiev and Luke Hudlass-Galley and Charlie Blake and Carlo Luschi and Douglas Orr , title =. Proceedings of the 41st International Conference on Machine Learning , series =

  8. [8]

    Rae and Anna Potapenko and Siddhant M

    Jack W. Rae and Anna Potapenko and Siddhant M. Jayakumar and Chloe Hillier and Timothy P. Lillicrap , title =. International Conference on Learning Representations , year =

Show all 25 references
  1. [9]

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

    Yushi Bai and Xin Lv and Jiajie Zhang and Hongchang Lyu and Jiankai Tang and Zhidian Huang and Zhengxiao Du and Xiao Liu and Aohan Zeng and Lei Hou and Yuxiao Dong and Jie Tang and Juanzi Li , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational...

  2. [10]

    First Conference on Language Modeling , year =

    Cheng-Ping Hsieh and Simeng Sun and Samuel Kriman and Shantanu Acharya and Dima Rekesh and Fei Jia and Yang Zhang and Boris Ginsburg , title =. First Conference on Language Modeling , year =

  3. [11]

    Advances in Neural Information Processing Systems , volume =

    Zichang Liu and Aditya Desai and Fangshuo Liao and Weitao Wang and Victor Xie and Zhaozhuo Xu and Anastasios Kyrillidis and Anshumali Shrivastava , title =. Advances in Neural Information Processing Systems , volume =

  4. [12]

    Abdi and Dongsheng Li and Chin-Yew Lin and Yuqing Yang and Lili Qiu , title =

    Huiqiang Jiang and Yucheng Li and Chengruidong Zhang and Qianhui Wu and Xufang Luo and Surin Ahn and Zhenhua Han and Amir H. Abdi and Dongsheng Li and Chin-Yew Lin and Yuqing Yang and Lili Qiu , title =. Advances in Neural Information Processing Systems , volume =

  5. [13]

    Advances in Neural Information Processing Systems , volume =

    Di Liu and Meng Chen and Baotong Lu and Huiqiang Jiang and Zhenhua Han and Qianxi Zhang and Qi Chen and Chengruidong Zhang and Bailu Ding and Kai Zhang and Chen Chen and Fan Yang and Yuqing Yang and Lili Qiu , title =. Advances in Neural Information Processing Systems , volume =

  6. [14]

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

    Wei Wu and Zhuoshi Pan and Kun Fu and Chao Wang and Liyi Chen and Yunchu Bai and Tianfu Wang and Zheng Wang and Hui Xiong , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

    2025

  7. [15]

    International Conference on Learning Representations , year =

    Tri Dao , title =. International Conference on Learning Representations , year =

  8. [16]

    Advances in Neural Information Processing Systems , volume =

    Prajwal Singhania and Siddharth Singh and Shwai He and Soheil Feizi and Abhinav Bhatele , title =. Advances in Neural Information Processing Systems , volume =

  9. [17]

    Mahoney and Kurt Keutzer and Amir Gholami , title =

    Coleman Richard Charles Hooper and Sehoon Kim and Hiva Mohammadzadeh and Monishwaran Maheswaran and Sebastian Zhao and June Paik and Michael W. Mahoney and Kurt Keutzer and Amir Gholami , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Ling...

  10. [18]

    International Conference on Learning Representations , year =

    Guangxuan Xiao and Jiaming Tang and Jingwei Zuo and Junxian Guo and Shang Yang and Haotian Tang and Yao Fu and Song Han , title =. International Conference on Learning Representations , year =

  11. [19]

    Advances in Neural Information Processing Systems , volume =

    Yizhao Gao and Zhichen Zeng and DaYou Du and Shijie Cao and Peiyuan Zhou and Jiaxing Qi and Junjie Lai and Hayden So and Ting Cao and Fan Yang and Mao Yang , title =. Advances in Neural Information Processing Systems , volume =

  12. [20]

    arXiv preprint arXiv:2412.15115 , year =

    Qwen and others , title =. arXiv preprint arXiv:2412.15115 , year =

    arXiv

  13. [21]

    arXiv preprint arXiv:2501.15383 , year =

    An Yang and Bowen Yu and Chengyuan Li and Dayiheng Liu and Fei Huang and Haoyan Huang and Jiandong Jiang and Jianhong Tu and Jianwei Zhang and Jingren Zhou and Junyang Lin and Kai Dang and Kexin Yang and Le Yu and Mei Li and Minmin Sun and Qin Zhu and Rui Men and Tao He and We...

    arXiv

  14. [22]

    arXiv preprint arXiv:2407.21783 , year =

    Aaron Grattafiori and others , title =. arXiv preprint arXiv:2407.21783 , year =

    arXiv

  15. [23]

    arXiv preprint arXiv:2512.16391 , year =

    Dhruv Deshmukh and Saurabh Goyal and Nipun Kwatra and Ramachandran Ramjee , title =. arXiv preprint arXiv:2512.16391 , year =

  16. [24]

    arXiv preprint arXiv:2603.12201 , year =

    Yushi Bai and Qian Dong and Ting Jiang and Xin Lv and Zhengxiao Du and Aohan Zeng and Jie Tang and Juanzi Li , title =. arXiv preprint arXiv:2603.12201 , year =

  17. [25]

    arXiv preprint arXiv:2606.30389 , year =

    Tianyu Wang and Gourav Rattihalli and Aditya Dhakal and Junbo Li and Zhiwei Ren and Dejan Milojicic and Longfei Shangguan , title =. arXiv preprint arXiv:2606.30389 , year =

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.