STAR-Pro: Stage-Wise Token Adaptive Reduction with Progressive Refinement for Efficient Large Vision-Language Models
Yichen Guo, Tinghao Wang, Qizhe Zhang, Lingbei Meng, Yuan Zhang, Jiajun Cao, Hao Jiang, Chenwei Wu, et al.
cs.CV · cs.AI
Quick verdict
STAR-Pro: training-free two-stage visual token pruning for LVLMs with strong wide-scope results; wins mostly hold up, but margins lack significance testing.
Summary
STAR-Pro is a training-free framework for pruning visual tokens in large vision-language models. Motivated by two empirical findings — that pre-LLM selectors lose feature-space coverage under aggressive budgets, and that text-to-visual attention importance shifts across decoder layers — it combines an Adaptive Stage, which uses pivoted QR to build an over-budget coverage candidate pool, with a Progressive Stage, which re-ranks survivors by averaged text-to-visual attention at two calibrated decoder layers under a layer-average token budget. Evaluation spans seven LVLMs (LLaVA-1.5/NeXT 7B/13B, LLaVA-Video-7B, Qwen3-VL-8B, InternVL3-8B) and 18 benchmarks against nine training-free baselines. Reported highlights: on LLaVA-Video-7B at 90.5% token reduction, 92.7% baseline retention and a measured 2.24x end-to-end speedup; best-or-tied aggregates in 11 of 12 LLaVA settings; InternVL3 leads all eight reported benchmarks at both budgets. Ablations cover stage necessity, the pool multiplier alpha, a 100-schedule pruning-layer sweep, and token-allocation robustness. Limitations: no statistical significance testing or error bars despite 0.1-0.5-point margins; architecture-specific schedules (pruning layers, per-crop quotas) calibrated case-by-case; VQAv2, VizWiz, and MM-Vet are listed as evaluated but never reported; video and advanced-architecture tables use reduced baseline sets; and fixed-resolution configurations may not reflect dynamic-resolution deployments.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Pre-LLM token selectors lose substantial feature-space coverage under aggressive budgets, and the text-attended visual token set drifts across decoder layers, making one-shot pruning unreliable.
- A pivoted-QR coverage pool followed by attention-based progressive refinement (STAR-Pro) achieves best-or-tied aggregates in 11 of 12 LLaVA settings and leads all eight reported InternVL3 benchmarks at both budgets.
- On LLaVA-Video-7B, STAR-Pro cuts visual tokens by 90.5% while retaining 92.7% of baseline performance with a measured 2.24x end-to-end speedup.
- The Adaptive-Stage pivoted-QR cost (40.6 ms) is negligible relative to the 872.9 ms decoder-prefill saving at T=16 on LLaVA-Video-7B.
Strengths
Unusually broad evaluation scope
Seven LVLMs (LLaVA-1.5/NeXT 7B/13B, LLaVA-Video-7B, Qwen3-VL-8B, InternVL3-8B), 18 benchmarks, nine training-free baselines, three compression levels (Tables 1-3, 6-10); the 11-of-12 Table 1 claim verifies against the extracted numbers.
Motivating analyses have documented protocols
Appendix E derives the coverage ratio eta via the Moore-Penrose projector with a Ky Fan spectral envelope (30 MME images) and defines layer continuity via averaged pairwise IoU over 500 images (Fig 2).
Credible efficiency accounting
Matched-compute ablations are extensive
Localized honest hedging
Weaknesses
No statistical discipline for small margins
Benchmark-coupled schedule calibration
Setup versus reporting mismatch
Baseline sets shrink in harder settings
Progressive Stage's marginal value is smaller than claimed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
STAR-Pro (arXiv 2609.05916) is a training-free visual token pruning framework for LVLMs. Two empirical findings motivate it: (1) pre-LLM selectors (CDPruner, DivPrune, VisionZip) lose feature-space coverage under aggressive budgets, measured by the retention ratio eta with a rigorous projector/Ky-Fan derivation (Fig 2a, Appendix E.1); and (2) the top-K text-attended visual token set drifts across decoder layers (IoU continuity well below 1; Fig 2b; 500 MME images). The method: a pivoted-QR Adaptive Stage builds a 2T coverage pool; a Progressive Stage prunes nested survivor sets at two decoder layers by averaged text-to-visual attention under an exact layer-average budget (Eq. 5). Headline claims: best-or-tied in 11 of 12 LLaVA settings (Table 1); LLaVA-Video-7B at 90.5% reduction retains 92.7% with a measured 2.24x speedup (Tables 2, 4); InternVL3-8B leads all eight benchmarks at both budgets (Table 3). I re-verified the 11-of-12 claim and the InternVL3 sweep from the extracted tables; both hold at reported precision.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is STAR-Pro and how much faster does it make vision-language models?
STAR-Pro is a training-free method that reduces visual tokens in large vision-language models in two stages: a pivoted-QR coverage selection before the decoder, then progressive attention-based refinement inside it. The paper reports a 2.24x end-to-end speedup on LLaVA-Video-7B with 90.5% of visual tokens removed, retaining 92.7% of unpruned accuracy. Speedups depend on model and budget; a competitor (VScan) is reported as faster on LLaVA-NeXT but with lower retained accuracy.
Does STAR-Pro require training or model modifications?
No. It is training-free and needs no architectural changes — only per-architecture hyperparameters: a candidate-pool multiplier (alpha=2 by default) and two pruning decoder layers with survivor counts, all tabulated for the seven evaluated models.
Which models and benchmarks were used to evaluate STAR-Pro?
Seven LVLMs (LLaVA-1.5-7B/13B, LLaVA-NeXT-7B/13B, LLaVA-Video-7B, Qwen3-VL-8B-Instruct, InternVL3-8B) against nine training-free pruning baselines, using 15 listed image benchmarks (10 per-benchmark columns shown for LLaVA models) and 3 video benchmarks (MLVU, LongVideoBench, Video-MME).
Is the STAR-Pro code publicly available?
Yes. The paper states that implementation code, evaluation configurations, and schedule-validation utilities are available at github.com/EasonAI-5589/starpro.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure panels (Figs 2, 4-7) are images; their per-cell values are not recoverable from the extracted text, so the Fig 5(a) ablation gaps and Fig 6 surface values rely on the authors' textual descriptions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.