Pith. sign in

REVIEW 3 major objections 4 minor 23 references

OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

T0 review · 3 major / 4 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 OmniDelta claims that deciding where a fixed token-compression budget is spent — between audio and video, then across audio segments and video frames — is a distinct and solvable problem from deciding which tokens to keep, and that solving

desk verdict Worth engaging, but the skill-router evidence is circular—the transferable gains in the ablations come from the intra-modal allocator, not the signature skill-driven module. read the letter →

arxiv 2607.25669 v1 pith:WDLJVET2 submitted 2026-07-28 cs.AI

Haoyang Huang , Wenjie Huang , Tianqi Xu , Hongyaoxing Gu , Kang Tan , Yikai Fu , Yuhao Shen , Tianyu Liu
show 9 more authors
Baolin Zhang Jun Zhang Xinyi Hu Jun Dai Shuang Ge Lei Chen Yue Li Mingchen Wang Meng Zhang
This is my paper
classification cs.AI
keywords tokencompressionbudgetallocationomni-modalLLMskill-poolroutingaudio-videopruningtraining-freeinferencequery-awareacceleration
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

OmniDelta sets out to show that token compression in omni-modal LLMs has a neglected upstream problem: how to divide a fixed retained-token budget between audio and video, and then among audio segments and video frames, before any token-pruning rule runs. The paper argues that direct query-to-audio/video similarity is a near-useless modality router (close to chance at every representation depth tested), and that uniform intra-modal budgets waste tokens even when an answer-aware oracle performs the pruning. As a replacement, OmniDelta routes the query through offline-built audio and video skill pools matched in the model's text-embedding space, then redistributes each modality budget using per-unit complexity and cross-unit redundancy, with an exact guarantee that the total retained ratio never changes. On four audio-video benchmarks and two Qwen2.5-Omni model sizes, this placement change improves compressed accuracy over the OmniZip baseline at both 25% and 20% retention, and at 25% retention on the 7B model it cuts GPU memory by 22% and gives a 1.64x end-to-end speedup. A sympathetic reader would take the paper as establishing that budget placement is a first-class axis of omni-modal compression, orthogonal to and stackable with existing token-selection methods.

What carries the argument

Central mechanism: a two-level allocator that never changes the total retained-token count. First, audio and video skill pools — lexical entries like 'rhythm,' 'who is speaking,' 'appearance,' 'read on-screen text' — are embedded once in the model's text-embedding space; the query is mean-pooled, matched Top-K per pool, and the score gap g = r_v − r_a gives a signed bias m_q = tanh(g/2), transferring at most λ_q·K tokens between modalities. Second, within each modality, local units (one-second audio segments, single frames) get an intra-unit complexity score C_i and a token-aligned redundancy score R_i to the previous unit; a centered score z_i lets redundant units only shed budget and compl

What would settle it

Build a benchmark of audio- and video-oriented questions whose wording deliberately avoids WorldSense-style cues (e.g., 'timbre,' 'prosody,' 'parallax,' 'occlusion'), label the required modality by hand, and measure OmniDelta's routing accuracy and end-to-end accuracy at 25% retention; if routing accuracy falls toward the ~60.8% direct-similarity ceiling and the accuracy gains vanish, the transfer assumption fails. Alternatively, compare OmniDelta's intra-modal budgets against the paper's own Oracle-Oracle split on its 50-clip diagnostic: if the complexity/redundancy proxy places budgets no cl

Watch

Extended reading notes

Core claim

OmniDelta claims that where a fixed token budget is spent — across audio/video and across segments/frames — is a distinct problem from which tokens to keep. Direct query-to-audio/video similarity routes near chance (60.8% at best), and even an answer-aware oracle pruner fails under uniform local budgets, so misallocation is not fixable by pruning. The training-free allocator shifts budget between modalities via a bounded signed bias from Top-K query-skill similarities, then adjusts per-unit keep counts using redundancy-minus-complexity scores, conserving the exact retained ratio. At 25% retention on Qwen2.5-Omni-7B it reaches 44.2% WorldSense (OmniZip: 43.2%) with 1.64x speedup and 22% less

Load-bearing premise

The inter-modal router assumes that fixed lexical skill entries built offline from WorldSense questions and answer candidates, matched to an arbitrary query by cosine similarity in the model's text-embedding space, reliably signal whether that query needs audio or video evidence — if the skill vocabulary does not transfer to new domains, the top-level allocation collapses to the fixed prior and the reported gains disappear.

Editorial extensions

If this is right

  • At 25% token retention on Qwen2.5-Omni-7B, OmniDelta uses 22.0% less GPU memory (30.0G to 23.4G), achieves a 1.69x time-to-first-token speedup and a 1.64x end-to-end speedup, while scoring 44.2% on WorldSense versus 43.2% for the OmniZip compressor it builds on.
  • The same allocator improves or matches compressed baselines across all four benchmarks (WorldSense, AVUT, VideoMME, DailyOmni) and on both the 3B and 7B Qwen2.5-Omni models, at both 25% and 20% retention.
  • Budget allocation is orthogonal to token selection: OmniDelta preserves the sample-level retained ratio exactly, so it can be layered on top of any pruning backend without changing the compression factor.
  • The query-level budget shift is the most sensitive hyperparameter (averages range 43.28%–44.29% across the coefficient sweep), so inter-modal transfer strength needs more careful calibration than the intra-modal shifts.
  • Skill-pool routing is the mechanism behind the inter-modal gain: 87.5% routing accuracy on the 400-query diagnostic versus at most 60.8% for direct similarity, translating to 44.0% versus 43.5% WorldSense under the same pruning backend.

Reading between the lines

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

  • The WorldSense-built skill lexicon is the transfer risk: if a target benchmark uses vocabulary the pools never saw, inter-modal allocation would degrade toward the fixed prior. A direct test is rebuilding pools from each benchmark's own questions and measuring the routing-accuracy delta.
  • The oracle diagnostic — even answer-aware pruning cannot fix a uniform budget — implies an upper bound on any fixed-budget compressor, and suggests the intra-modal proxies (complexity, redundancy) are replaceable: a better predictor of where evidence sits could push the same allocator closer to the Oracle-Oracle scores (3.55 video, 3.40 audio) from the Uniform-Oracle scores (3.25, 3.14).
  • Because the allocator is training-free and pruning-agnostic, it should transfer to other omni-modal models as long as their text-embedding space is reused for skill matching; the main hidden risk is embedding-space drift between model versions, which would silently break the cached skill bank.
  • The paper only reports 20–25% retention; whether hierarchical allocation still beats uniform budgets at gentler compression (e.g., 50%) would clarify whether the effect is specific to extreme pruning or general across the cost-accuracy tradeoff.
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

3 major / 4 minor

Summary. The paper proposes OmniDelta, a training-free hierarchical budget allocator for omni-modal LLM token compression. Given a fixed total retained-token budget, OmniDelta first shifts tokens between audio and video modalities using query-skill cosine similarity, where the skill pools are constructed offline by GPT-5.5 from WorldSense questions and answer candidates. It then redistributes each modality's budget across audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets are used with an OmniZip-style pruning backend. Experiments on WorldSense, AVUT, VideoMME, and DailyOmni with Qwen2.5-Omni-7B/3B report improved compressed accuracy over OmniZip, with a 1.64x end-to-end speedup at 25% retention on the 7B model. The Appendix provides allocator pseudocode, conservation/boundedness proofs, and full diagnostic protocols.

Significance. The paper identifies and formalizes a genuinely under-explored problem: before deciding which tokens to prune, how should a fixed retained-token budget be allocated across modalities and temporal units? The controlled diagnostic studies in Sec. 2 are thoughtful, and the formal properties in Appendix B—exact budget conservation, bounded and monotonic inter-modal transfer, and feasibility of integer allocation—are correct and useful. The method is training-free, adds only linear overhead, and is compatible with existing pruning backends. If the skill-routing component transferred across benchmarks, the framework would be practically valuable. However, the current evidence supports the transferability of the intra-modal reallocation much more strongly than the central 'skill-driven' inter-modal contribution, and the most important empirical claims rest on WorldSense-derived skill pools and WorldSense-tuned hyperparameters.

major comments (3)
  1. [§3.2.1, App. E.1, Table 3] The skill pools are built by having GPT-5.5 extract audio/video task types and keywords from WorldSense questions and answer candidates (App. E.1), and the inter-modal router is then evaluated on WorldSense (Fig. 3, Table 3). This makes WorldSense an in-distribution test for the skill lexicon. More importantly, the ablation in Table 3 shows that adding the inter-modal A/V router alone (ID 'A/V only') improves WorldSense by +0.8 but decreases AVUT by -0.4 and VideoMME by -0.5, with only +0.3 on DailyOmni. Thus the skill-driven component does not transfer to benchmarks not used for pool construction; the held-out gains in Table 2 are attributable to the intra-modal reallocations. Since the title and abstract make 'skill-driven' the central contribution, the evidence does not support the general claim that query-skill routing improves OmniLLM compression. The authors should either construct
  2. [§4.1, Fig. 7] The three shift coefficients (lambda_q, lambda_a, lambda_v) are selected from a sweep on WorldSense, and the primary WorldSense results in Tables 1 and 3 are reported on the same benchmark. No held-out validation of these hyperparameters is presented. Given that the gains over OmniZip are small (e.g., +1.0 on WorldSense at 25% for 7B), the reported WorldSense improvement may partly reflect selection on the test set. The authors should use a validation split for coefficient selection or report the full grid results on the held-out benchmarks; the current Fig. 7 shows only WorldSense and therefore cannot establish that the chosen coefficients generalize.
  3. [Table 4] Table 4 compares efficiency at different retained-token ratios: DyCoke at 50%, OmniZip at 30%, and OmniDelta at 25%. Consequently, the statement that OmniDelta provides 'the largest acceleration' among compressed baselines is partly an artifact of the more aggressive budget setting. The Pareto-frontier claim in Fig. 1 and the Abstract would be stronger if efficiency were compared at matched retained ratios, or if full accuracy-latency tradeoff curves over ratios were shown for each method. Please add matched-ratio efficiency numbers or justify why different operating points are the appropriate comparison.
minor comments (4)
  1. [Tables 1, 2, 4] No variance or statistical significance information is reported. Several accuracy deltas are small (0.4–1.0 points), and the random-pruning baseline is stochastic. If evaluation is deterministic, say so; otherwise report repeated-run means and standard deviations or a significance test.
  2. [Appendix E] For reproducibility, the full skill pools should be released. Appendix E.2 shows only selected entries, and no code or data link is provided. The query_modality_skill_pool.json file mentioned in E.1 is not included.
  3. [Appendix B.4] Proposition 4 assumes an unobservable convex loss function and proves only that uniform allocation is suboptimal under heterogeneous marginal gains. The text acknowledges that the proposed signals are proxies, but the section heading 'Why Heterogeneous Budgets Are Necessary' could be misread as an optimality guarantee for OmniDelta. Please add an explicit caveat that no optimality claim is made for the proxy-based allocator.
  4. [Eq. (8)] The aligned token features tilde{x}_{i,j} are introduced as 'after resampling adjacent units to a common length,' but the resampling procedure (interpolation? selection? pooling?) is not specified. Please define it precisely.

Circularity Check

1 steps flagged · score 6.0 of 10

Skill-pool router is fitted on WorldSense and its only positive evidence is WorldSense; the held-out ablations show the skill-driven component does not transfer, so the skill-routing prediction is in-sample.

  1. fitted input called prediction [Eqs. (4)-(6), Sec. 3.2.1; Sec. 2.1 Fig. 3; Appendix E]
    "We therefore construct modality skill pools offline with GPT-5.5-xhigh. From WorldSense, it identifies representative audio-oriented task types ... It extracts diagnostic keywords for these tasks and expands them with semantically related terms. ... We use GPT-5.5-xhigh to categorize WorldSense queries ... and sample 200 audio-oriented and 200 video-oriented queries to construct a balanced diagnostic set of 400 queries."

    The only fitted component of the inter-modal router is the skill bank S_a,S_v, built by GPT-5.5 from WorldSense questions and answer candidates (Appendix E). Eq. (4) scores a query by cosine similarity to the top-k entries of this bank, and Eq. (6) shifts modality budgets from that score. The paper then reports routing accuracy on 400 WorldSense queries and end-to-end WorldSense accuracy as evidence. These are in-sample: the test queries come from the same corpus used to select and expand the lexical entries, so high similarity and accuracy are expected by construction. Table 3 confirms the lack of transfer: adding the A/V skill router alone changes held-out accuracy by -0.4 (AVUT), -0.5 (VideoMME), and +0.3 (DailyOmni), while WorldSense improves +0.8.

full rationale

The strongest circular component is the skill-driven inter-modal router: its lexical pools are constructed offline from WorldSense Q/A, and the routing and end-to-end WorldSense results used to support it are measured on the same benchmark. The 25.5-28.6 pp routing advantage and the 43.2->44.2 WorldSense gain are therefore partly in-sample fits rather than independent predictions. The shift coefficients (lambda_q, lambda_a, lambda_v) are also calibrated on WorldSense (Fig. 7, Sec. 4.1), further tying the headline numbers to the construction corpus. However, the full OmniDelta also improves on AVUT, VideoMME, and DailyOmni, which were not used in pool construction or coefficient selection, and those gains come primarily from the content-aware intra-modal allocation. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled via citation. The Table 3 held-out degradation of the A/V-only router is a generalization failure rather than a formal circularity, but it reinforces that the skill-driven mechanism lacks independent evidence. Overall, one prediction (skill-based routing and its WorldSense payoff) reduces by construction to the WorldSense-derived inputs, while the broader Pareto claim retains independent intra-modal content; hence a partial-circularity score of 6.

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

The method has six explicit or implicit free parameters (three shift coefficients, prior budget split, Top-k, skill-pool contents), plus several domain assumptions about representation geometry, perceptual proxies, and LLM judges. The theory in Appendix B proves only budget conservation and boundedness, not optimality, so the practical behavior rests on these empirical choices.

free parameters (6)
  • lambda_q (query shift coefficient) = 0.05 (7B), 0.15 (3B)
    Controls strength of inter-modal budget shift; selected from a 27-config grid sweep on WorldSense (Figure 7).
  • lambda_a (audio shift coefficient) = 0.20 (7B), 0.10 (3B)
    Controls intra-modal budget shift for audio segments; selected from grid sweep on WorldSense.
  • lambda_v (video shift coefficient) = 0.30 (7B), 0.40 (3B)
    Controls intra-modal budget shift for video frames; selected from grid sweep on WorldSense.
  • Prior modality budget (K0a, K0v) = Derived from OmniZip allocation: 50% audio keep rate, 20% video keep rate at 25% total retention
    Hand-chosen starting point for modal split; not swept or justified independently.
  • k (Top-k skills) = Not specified in method section (k=3 used in diagnostic)
    The number of nearest skill entries used in Eq. (4) is never stated for the actual router, making replication ambiguous.
  • Skill pool contents = 90 audio entries, 122 video entries (generated by GPT-5.5 from WorldSense)
    The lexical entries are the core routing signal; their exact composition is a fitted artifact of the benchmark and the generator.
assumptions (6)
  • domain assumption Information loss D_i(k) is differentiable, decreasing, and convex in k (Assumption 1, Appendix B.4)
    Used to prove uniform allocation is suboptimal (Prop. 4); not verified against actual token-loss curves.
  • domain assumption The query primarily encodes task intent, while modality tokens encode sample-specific content (Sec. 2.1)
    Motivates the failure of direct query-token similarity; supported only by the paper's own diagnostic.
  • domain assumption Cosine similarity between query text and skill-pool text embeddings indicates required modality (Sec. 3.2.1)
    This is the load-bearing transfer assumption of the inter-modal router; no theoretical or cross-domain guarantee is given.
  • domain assumption Local complexity (intra-unit token diversity) and temporal redundancy (inter-unit similarity) are valid proxies for information density (Sec. 3.2.2)
    Heuristic used to redistribute budgets; never validated against oracle-optimal budgets on real OmniLLM tokens.
  • domain assumption GPT-5.5/Gemini-3.1 scoring in the diagnostics is a valid measure of answer quality (Sec. 2.2, Appendix D.2)
    The controlled diagnostics rely on LLM judges to score answers; no human validation or inter-judge agreement is reported.
  • domain assumption The OmniZip pruning backend and DPC-KNN operate correctly as described in prior work (Sec. 3.2.3)
    The downstream pruning rules are treated as black-box, but their correctness and implementation details are not re-derivable from this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs." pith.science (2026). https://pith.science/paper/WDLJVET2

@misc{pith2026260725669,
  author       = {Pith},
  title        = {Pith review of: OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDLJVET2}},
  note         = {Machine review of arXiv:2607.25669}
}
read the original abstract

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 3 linked inside Pith

  1. [1]

    Inspect all samples individually; do not randomly sample

  2. [2]

    Classify from the question and answer candidates, without using the ground-truth answer or a model prediction

  3. [3]

    Yudong Yang, Jimin Zhuang, Guangzhi Sun, Changli Tang, Yixuan Li, Peihan Li, Yifan Jiang, Wei Li, Zejun Ma, and Chao Zhang

    URLhttps://arxiv.org/abs/2605.18041. Yudong Yang, Jimin Zhuang, Guangzhi Sun, Changli Tang, Yixuan Li, Peihan Li, Yifan Jiang, Wei Li, Zejun Ma, and Chao Zhang. AVUT: Audio-centric video understanding benchmark without text shortcut, 2025. URL https://arxiv.org/abs/2503.19951. Hanrong Ye, Chao-Han Huck Yang, Arushi Goel, Wei Huang, Ligeng Zhu, Yuanhang Su...

    arXiv 2025

  4. [4]

    If the answer candidates reinforce the modality, mention this only in the reason

    If several cues occur, record the most diagnostic phrase and optionally record secondary cues. If the answer candidates reinforce the modality, mention this only in the reason

  5. [5]

    How does the rhythm of the pipa music change at the beginning of the video?

    If more than 200 high-confidence samples are available for one modality, prioritize explicit modality cues, modality- specific candidates, high confidence, and diversity across tasks and domains. For every inspected sample, return its identifiers, question, candidates, candidate modality (audio, video, or mixed/ambiguous), confidence from 1 to 5, primary ...

  6. [6]

    A visible instrument does not make a sound-related skill visual, and the presence of a speaker does not make a speech-understanding skill visual

    Audit every entry for cross-modal ambiguity. A visible instrument does not make a sound-related skill visual, and the presence of a speaker does not make a speech-understanding skill visual. Retain a cue only when it reliably indicates the evidence needed to answer the query

  7. [7]

    Prefer high-confidence questions whose required modality is unambiguous, and exclude questions for which both modalities are necessary or either modality alone is plausible

  8. [10]

    Uniform budget–Random pruning:allocate 25% to every view and randomly retain image regions within each view

Show all 23 references
  1. [11]

    3.First-only:spend the complete budget on the first view and remove the remaining views

    Weighted budget–Random pruning:use the fixed view-level allocation[0.70, 0.15, 0.10, 0.05] and randomly retain regions within each view. 3.First-only:spend the complete budget on the first view and remove the remaining views

  2. [12]

    Uniform budget–Oracle pruning:keep the budget uniform, but use the reference answer to retain the regions that best preserve answer-relevant evidence in each view

  3. [13]

    Cover removed regions with a mask

    Oracle budget–Oracle pruning:use the reference answer to determine both the budget distribution and retained regions while keeping the same total budget. Cover removed regions with a mask. For each variant, answer the question using only its retained visible content; do not co...

  4. [14]

    3.First-only:retain all chunks from the first two segments and remove all chunks from the last two segments

    Weighted budget–Random pruning:retain[5, 4, 2, 1] chunks in the four segments and select chunks randomly within each segment. 3.First-only:retain all chunks from the first two segments and remove all chunks from the last two segments

  5. [15]

    Uniform budget–Oracle pruning:retain three chunks per segment, selecting the most important chunks according to the answer-aware ranking

  6. [16]

    Replace removed chunks with silence

    Oracle budget–Oracle pruning:distribute the 12-chunk budget freely across segments and retain the globally most answer-relevant chunks. Replace removed chunks with silence. Answer the question using only the current pruned audio, without access to the original clip, reference ...

  7. [17]

    Exclude mixed or ambiguous task types for which both modalities are indispensable

    Inspect the questions and answer candidates and identify representative task types that strongly require audio or video evidence. Exclude mixed or ambiguous task types for which both modalities are indispensable

  8. [18]

    Organize audio tasks around speech and dialogue, speaker and voice, music and rhythm, sound-event recognition, sound source, event counting, acoustic changes, loudness, and silence. Organize video tasks around object and scene recognition, appearance and attributes, spatial re...

  9. [19]

    Prefer general task cues that recur across samples rather than named entities, answer-specific values, or descriptions tied to one clip

    For each task type, select the smallest keywords or short query phrases that indicate the required evidence. Prefer general task cues that recur across samples rather than named entities, answer-specific values, or descriptions tied to one clip

  10. [20]

    voice”, “spoken words

    Expand each cue with useful synonyms, lexical variants, and short semantic paraphrases. For example, a speech task may contribute “voice”, “spoken words”, and “who is speaking”, while an OCR task may contribute “text”, “caption”, and “read on-screen text”

  11. [21]

    Use lowercase text, preserve meaningful multiword expressions, and avoid generic function words or the uninformative word “video” by itself

    Keep entries concise enough to represent a single skill. Use lowercase text, preserve meaningful multiword expressions, and avoid generic function words or the uninformative word “video” by itself

  12. [23]

    The two modality pools need not have identical sizes, but each should represent the task diversity observed in WorldSense

    Cover diverse task families rather than overpopulating one category. The two modality pools need not have identical sizes, but each should represent the task diversity observed in WorldSense. Produce four lists.audio_skillsandvideo_skillscontain compact words and lexical expre...

  13. [1057]

    Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang

    URLhttps://aclanthology.org/2026.findings-acl.1057/. Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, and Shanghang Zhang. SparseVLM: Visual token sparsification for efficient vision-langua...

    2026 arXiv

  14. [2025]

    Zining Wang, Zhihang Yuan, Yingjie Zhai, Wenshuo Li, Han Shu, Ruihao Gong, Jinyang Guo, and Xianglong Liu

    URLhttps://arxiv.org/abs/2510.13747. Zining Wang, Zhihang Yuan, Yingjie Zhai, Wenshuo Li, Han Shu, Ruihao Gong, Jinyang Guo, and Xianglong Liu. OmniFit: Bridging modalities via layer-adaptive token compression for omnimodal large language models. InInternational Conference on ...

    2026

  15. [2026]

    Xiaohu Huang, Hao Zhou, and Kai Han

    URLhttps://arxiv.org/abs/2502.04326. Xiaohu Huang, Hao Zhou, and Kai Han. PruneVid: Visual token pruning for efficient video large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 19959–19973, 2025. URL https://arxiv.org/abs/2412.1611...

    2025 arXiv

Pith tools

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