{"paper":{"title":"See What Matters: Differentiable Grid Sample Pruning for Generalizable Vision-Language-Action Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Differentiable grid sampling lets VLA models compress visual tokens to under 10 percent while keeping full manipulation success.","cross_cats":["cs.CV"],"primary_cat":"cs.RO","authors_text":"Chang Xu, Chengbin Du, Chenghao Xia, Yanxiang Ma, Yixu Feng, Yunke Wang, Zinan Zhao","submitted_at":"2026-05-12T09:08:42Z","abstract_excerpt":"Vision-Language-Action (VLA) models have shown remarkable promise in robotics manipulation, yet their high computational cost hinders real-time deployment. Existing token pruning methods suffer from a fundamental trade-off: aggressive compression using pruning inevitably discards critical geometric details like contact points, leading to severe performance degradation. This forces a compromise, limiting the achievable compression rate and thus the potential speedup. We argue that breaking this trade-off requires rethinking compression as a geometry-aware, continuous token resampling in the vis"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"validating the lowest feasible visual token count reported to date, GridS achieves a 76% reduction in FLOPs with no degradation in the success rate.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That adaptively predicting a minimal set of salient coordinates and extracting features via differentiable interpolation will always preserve the critical geometric details (contact points, spatial relations) required for manipulation success across tasks and environments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GridS reduces visual tokens in VLA models to under 10% of the original count via task-aware differentiable resampling, delivering 76% lower FLOPs with no drop in task success rate on benchmarks and real robots.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Differentiable grid sampling lets VLA models compress visual tokens to under 10 percent while keeping full manipulation success.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b7a26ca9e662d38a871a961c34d9c9af25335b81066568a1e799cdd4343096e4"},"source":{"id":"2605.11817","kind":"arxiv","version":2},"verdict":{"id":"02473b66-5ced-4ed6-aa02-292631ff9758","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T05:37:23.048495Z","strongest_claim":"validating the lowest feasible visual token count reported to date, GridS achieves a 76% reduction in FLOPs with no degradation in the success rate.","one_line_summary":"GridS reduces visual tokens in VLA models to under 10% of the original count via task-aware differentiable resampling, delivering 76% lower FLOPs with no drop in task success rate on benchmarks and real robots.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That adaptively predicting a minimal set of salient coordinates and extracting features via differentiable interpolation will always preserve the critical geometric details (contact points, spatial relations) required for manipulation success across tasks and environments.","pith_extraction_headline":"Differentiable grid sampling lets VLA models compress visual tokens to under 10 percent while keeping full manipulation success."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11817/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T11:37:09.344165Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T09:01:18.728167Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:09:33.258817Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"5369e7ac0214a6872f8302180b02f9bfb6ef7fedf5ac01a1556e47fe9a8a1d7f"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"bc3aa9c3266c2b61e4fe9b2fd3c9ad46fd5fe3dd562c25bf4020765d7a7ffc98"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}