Pith. sign in

REVIEW 9 cited by

TokenCarve: Information-Preserving Visual Token Compression in Multimodal Large Language Models

Not yet reviewed by Pith; the record is open.

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

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.10501 v1 pith:WXJFORND submitted 2025-03-13 cs.CV

Xudong Tan , Peng Ye , Chongjun Tu , Jianjian Cao , Yaoxin Yang , Lin Zhang , Dongzhan Zhou , Tao Chen This is my paper
classification cs.CV
keywords tokencompressiontokencarvemultimodalperformancetokensvisualwhile
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Multimodal Large Language Models (MLLMs) are becoming increasingly popular, while the high computational cost associated with multimodal data input, particularly from visual tokens, poses a significant challenge. Existing training-based token compression methods improve inference efficiency but require costly retraining, while training-free methods struggle to maintain performance when aggressively reducing token counts. In this study, we reveal that the performance degradation of MLLM closely correlates with the accelerated loss of information in the attention output matrix. This insight introduces a novel information-preserving perspective, making it possible to maintain performance even under extreme token compression. Based on this finding, we propose TokenCarve, a training-free, plug-and-play, two-stage token compression framework. The first stage employs an Information-Preservation-Guided Selection (IPGS) strategy to prune low-information tokens, while the second stage further leverages IPGS to guide token merging, minimizing information loss. Extensive experiments on 11 datasets and 2 model variants demonstrate the effectiveness of TokenCarve. It can even reduce the number of visual tokens to 22.2% of the original count, achieving a 1.23x speedup in inference, a 64% reduction in KV cache storage, and only a 1.54% drop in accuracy. Our code is available at https://github.com/ShawnTan86/TokenCarve.

Discussion (0). Sign in to comment.

Forward citations

Cited by 9 Pith papers

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

  1. Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference

    cs.CV 2026-07 accept novelty 7.0 of 10

    For text-rich MLLMs, accuracy at a fixed token budget does not determine whether the surviving tokens retain spatial coverage of the OCR region that answers the query.

  2. OmniPack: Unified Token Compression for Efficient Omni-modal Large Language Models

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A training-free two-stage token compression framework preserves 92.9% of Qwen2.5-Omni-7B's audio-visual understanding accuracy while using 6.8% of the original multimodal-token FLOPs.

  3. Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Same-accuracy pruned vision-language prefixes can retain very different amounts of traceable OCR support (PosECR 0.620 vs 0.270 for Target vs Random at 30% budget), so pruning evaluation should report spatial provenan...

  4. Spectral-Progressive Thought Flow for Lightweight Multimodal Reasoning

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    SpecFlow represents intermediate visual thoughts in fixed-size DCT space and uses classifier-free guidance to steer updates from textual thoughts, achieving up to 2.1x lower computation and KV cache costs.

  5. Keep What Audio Cannot Say: Context-Preserving Token Pruning for Omni-LLMs

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    ContextGuard prunes 55% of tokens in Qwen2.5-Omni 7B while matching full performance on five of six audio-visual benchmarks by preserving audio-irrecoverable visual context.

  6. ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Frozen visual tokens with sparse cross-attention at selected layers preserve 98% accuracy while reducing vision-side FLOPs to 4% in LLaVA-1.5 models.

  7. OmniZip: Audio-Guided Dynamic Token Compression for Fast Omnimodal Large Language Models

    cs.CV 2025-11 conditional novelty 6.0 of 10

    OmniZip introduces an audio-guided dynamic token compression framework that achieves 3.42X inference speedup and 1.4X memory reduction for omnimodal LLMs without any training.

  8. OmniRefine: Alignment-Aware Cooperative Compression for Efficient Omnimodal Large Language Models

    cs.AI 2026-05 unverdicted novelty 5.0 of 10

    OmniRefine introduces alignment-aware chunk refinement via similarity and dynamic programming followed by modality-cooperative token compression, achieving near-baseline accuracy at 44% token retention on WorldSense.

  9. Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models

    cs.CV 2026-03 unverdicted novelty 5.0 of 10

    AOT reduces visual tokens in VLLMs via intra-frame and inter-frame anchors with local-global optimal transport, delivering competitive benchmark performance and efficiency gains in a training-free way.

Pith tools