{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:6AF7F2KURPOYRPTIRADAVDSXT3","short_pith_number":"pith:6AF7F2KU","schema_version":"1.0","canonical_sha256":"f00bf2e9548bdd88be6888060a8e579ed38fe90620204461800696a07e6b4559","source":{"kind":"arxiv","id":"2505.21200","version":1},"attestation_state":"computed","paper":{"title":"Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Bizhe Bai, Jia Hao, Jialin Zheng, Peng Ye, Tao Chen, Xinyi Wang, Xudong Tan, Yaoxin Yang","submitted_at":"2025-05-27T13:47:18Z","abstract_excerpt":"Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-stemming from large-scale token computation and autoregressive decoding-poses significant challenges for real-time deployment and edge applications. While prior work has primarily focused on architectural optimization, we take a different perspective by identifying a dual form of redundancy in VLA models: (i) high similarity across consecutive action steps, and (ii) substantial redundancy in visual tokens. Motivated "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2505.21200","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-05-27T13:47:18Z","cross_cats_sorted":[],"title_canon_sha256":"9c6660d904ab4622b9554f8a47f9f5dc973707bfc86da484042d6b72ff931569","abstract_canon_sha256":"4353a226e7d4c1053028d1b9c82bb1c04ea0af10f3f90d0fef2e062b8ae08b48"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:10:38.646587Z","signature_b64":"u7JjH6OHezxBnq6cC2x0Rpkq9vHveuzVm9kCAdoDXCxfteLCPx6U7AJVbU6Xr7jc7ePxvl3EZfuqbmyq2lmpCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"f00bf2e9548bdd88be6888060a8e579ed38fe90620204461800696a07e6b4559","last_reissued_at":"2026-07-05T11:10:38.646097Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:10:38.646097Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Bizhe Bai, Jia Hao, Jialin Zheng, Peng Ye, Tao Chen, Xinyi Wang, Xudong Tan, Yaoxin Yang","submitted_at":"2025-05-27T13:47:18Z","abstract_excerpt":"Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-stemming from large-scale token computation and autoregressive decoding-poses significant challenges for real-time deployment and edge applications. While prior work has primarily focused on architectural optimization, we take a different perspective by identifying a dual form of redundancy in VLA models: (i) high similarity across consecutive action steps, and (ii) substantial redundancy in visual tokens. Motivated "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.21200","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2505.21200/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2505.21200","created_at":"2026-07-05T11:10:38.646154+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.21200v1","created_at":"2026-07-05T11:10:38.646154+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.21200","created_at":"2026-07-05T11:10:38.646154+00:00"},{"alias_kind":"pith_short_12","alias_value":"6AF7F2KURPOY","created_at":"2026-07-05T11:10:38.646154+00:00"},{"alias_kind":"pith_short_16","alias_value":"6AF7F2KURPOYRPTI","created_at":"2026-07-05T11:10:38.646154+00:00"},{"alias_kind":"pith_short_8","alias_value":"6AF7F2KU","created_at":"2026-07-05T11:10:38.646154+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":15,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.22907","citing_title":"Improving Robotic Imitation Learning via Trajectory Standardization","ref_index":1,"is_internal_anchor":false},{"citing_arxiv_id":"2606.22794","citing_title":"UniFS: Unified Fast-to-Slow Hierarchical Architecture for Vision-Language-Action Models","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2606.03188","citing_title":"GeoSem-WAM: Geometry- and Semantic-Aware World Action Models","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13548","citing_title":"AttenA+: Rectifying Action Inequality in Robotic Foundation Models","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31382","citing_title":"Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2606.27755","citing_title":"Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.29438","citing_title":"ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2603.14371","citing_title":"OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2508.13073","citing_title":"Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey","ref_index":115,"is_internal_anchor":false},{"citing_arxiv_id":"2511.18082","citing_title":"ActDistill: General Action-Guided Self-Derived Distillation for Efficient Vision-Language-Action Models","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2603.01581","citing_title":"KERV: Kinematic-Rectified Speculative Decoding for Embodied VLA Models","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11459","citing_title":"Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13548","citing_title":"AttenA+: Rectifying Action Inequality in Robotic Foundation Models","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11459","citing_title":"Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05323","citing_title":"VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success","ref_index":10,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3","json":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3.json","graph_json":"https://pith.science/api/pith-number/6AF7F2KURPOYRPTIRADAVDSXT3/graph.json","events_json":"https://pith.science/api/pith-number/6AF7F2KURPOYRPTIRADAVDSXT3/events.json","paper":"https://pith.science/paper/6AF7F2KU"},"agent_actions":{"view_html":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3","download_json":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3.json","view_paper":"https://pith.science/paper/6AF7F2KU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.21200&json=true","fetch_graph":"https://pith.science/api/pith-number/6AF7F2KURPOYRPTIRADAVDSXT3/graph.json","fetch_events":"https://pith.science/api/pith-number/6AF7F2KURPOYRPTIRADAVDSXT3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3/action/storage_attestation","attest_author":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3/action/author_attestation","sign_citation":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3/action/citation_signature","submit_replication":"https://pith.science/pith/6AF7F2KURPOYRPTIRADAVDSXT3/action/replication_record"}},"created_at":"2026-07-05T11:10:38.646154+00:00","updated_at":"2026-07-05T11:10:38.646154+00:00"}