Visual token pruning drops single-score ranking: graph propagation plus degree correction keeps 98.7% accuracy at 30% tokens on LLaVA-NeXT

Which visual tokens to keep is no longer decided by each token's individual saliency score alone; tokens exchange scores before the final ranking. Earlier selectors drew on visual redundancy, language-model attention or task-conditioned relevance, and graph methods related visual tokens and propagated importance, but initial scores did not determine the final ranking on their own. Across four backbones, ten benchmarks and seven pruning ratios, with all baselines re-implemented under one protocol, the method achieves the highest retention at most backbone-budget settings, and the margin grows through the moderate-to-aggressive regime.

The selector first compares text and visual states, then propagates scores on a visual graph, and finally corrects the ranking by neighborhood count. Propagation uses row normalization to fix the total weight leaving each source, while degree correction separately adjusts the ranking using neighborhood counts rather than weighted row sums. On LLaVA-NeXT, removing degree correction drops retention from 99.43% to 91.88%, and dense normalization gives 91.74%, against 92.12% for random selection.

One trained checkpoint serves every evaluated budget through explicit retention-ratio conditioning. Previously each budget typically required training a separate selector. The selector has about M parameters and is trained end-to-end through the frozen LLaVA-NeXT-8B backbone using only the language-modeling objective, with training ratios covering several retention levels and evaluation covering seven.

The report covers four backbones, ten benchmarks and seven pruning ratios, with all baselines re-implemented inside one evaluation pipeline, and comparisons are restricted to these evaluated methods and settings. The selector runs once before the decoder, unselected visual tokens are physically removed, and retained tokens preserve their original position identifiers. On LLaVA-NeXT, retaining 30% of candidate tokens yields a reported net prefill speedup and an end-to-end FLOPs reduction. Coverage observations distinguish a retention ratio from the amount and distribution of surviving evidence, and coverage-aware selection is named as a direction for testing.

The reported numbers follow a per-point best-hyperparameter protocol, a performance envelope rather than a fixed-deployment curve. The method is overtaken by FasterVLM and VisionZip on LLaVA-1.5 and by G-Prune on InternVL3. Uniform initialization already preserves most of the full method's performance, so the ablations do not establish learned initialization as the main gain source. A propagation-depth sweep shows additional propagation is not uniformly beneficial, and the coverage observations do not isolate the cause of accuracy differences.