{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:CW7IWZ46YM7VWQNJOBNGPBBPUG","short_pith_number":"pith:CW7IWZ46","schema_version":"1.0","canonical_sha256":"15be8b679ec33f5b41a9705a67842fa1981380046c4fd4fb2995a6d2eef2fecf","source":{"kind":"arxiv","id":"2601.15197","version":6},"attestation_state":"computed","paper":{"title":"LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LangForce restores language grounding in vision-language-action models by maximizing conditional pointwise mutual information between instructions and actions.","cross_cats":["cs.CL","cs.CV","cs.RO"],"primary_cat":"cs.AI","authors_text":"Bin Yu, Changti Wu, Cong Huang, Kai Chen, Laurence T. Yang, Shijie Lian, Xiaopeng Lin, Yuzhuo Miao, Zhaolong Shen","submitted_at":"2026-01-21T17:15:22Z","abstract_excerpt":"Vision-Language-Action (VLA) models have shown promise in robot manipulation but often struggle to generalize to new instructions or complex multi-task scenarios. We identify a critical pathology in current training paradigms where goal-driven data collection creates a dataset bias. In such datasets, language instructions are highly predictable from visual observations alone, causing the conditional mutual information between instructions and actions to vanish, a phenomenon we term Information Collapse. Consequently, models degenerate into vision-only policies that ignore language constraints "},"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":true},"canonical_record":{"source":{"id":"2601.15197","kind":"arxiv","version":6},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2026-01-21T17:15:22Z","cross_cats_sorted":["cs.CL","cs.CV","cs.RO"],"title_canon_sha256":"a2f940e430f04da099d006def40320a211a8eaabe42cdfdf8b048208a6a7d653","abstract_canon_sha256":"5d8f3b3c3e63e88ea90b00a8b588f3289b724b72ca3ba6584ba3dfe04fad01cd"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:09:24.337849Z","signature_b64":"1vTHYmnF5OOzfPdgqTe2nFLSyahdghFqHuSMtcixZCuLaMeb2sywIJ4zo1JDSYVWRhD8rhZvEqCmrfGUfvJDBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"15be8b679ec33f5b41a9705a67842fa1981380046c4fd4fb2995a6d2eef2fecf","last_reissued_at":"2026-05-18T03:09:24.337228Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:09:24.337228Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LangForce restores language grounding in vision-language-action models by maximizing conditional pointwise mutual information between instructions and actions.","cross_cats":["cs.CL","cs.CV","cs.RO"],"primary_cat":"cs.AI","authors_text":"Bin Yu, Changti Wu, Cong Huang, Kai Chen, Laurence T. Yang, Shijie Lian, Xiaopeng Lin, Yuzhuo Miao, Zhaolong Shen","submitted_at":"2026-01-21T17:15:22Z","abstract_excerpt":"Vision-Language-Action (VLA) models have shown promise in robot manipulation but often struggle to generalize to new instructions or complex multi-task scenarios. We identify a critical pathology in current training paradigms where goal-driven data collection creates a dataset bias. In such datasets, language instructions are highly predictable from visual observations alone, causing the conditional mutual information between instructions and actions to vanish, a phenomenon we term Information Collapse. Consequently, models degenerate into vision-only policies that ignore language constraints "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Without requiring new data, LangForce significantly improves generalization... including an 11.3% improvement on the challenging OOD SimplerEnv benchmark.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identified information collapse is the dominant cause of poor OOD performance and that maximizing conditional PMI via the dual-branch architecture will reliably correct it without introducing new failure modes or requiring careful hyperparameter tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LangForce prevents vision-language-action models from ignoring language by maximizing conditional pointwise mutual information via Bayesian decomposition with latent action queries.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LangForce restores language grounding in vision-language-action models by maximizing conditional pointwise mutual information between instructions and actions.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9355ce0d1517b1d25726560d9f452f287cbbeb6630412fe28969e01a72e53486"},"source":{"id":"2601.15197","kind":"arxiv","version":6},"verdict":{"id":"44b4db63-88ca-4edb-a2c5-f47d5236710c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T11:56:02.585654Z","strongest_claim":"Without requiring new data, LangForce significantly improves generalization... including an 11.3% improvement on the challenging OOD SimplerEnv benchmark.","one_line_summary":"LangForce prevents vision-language-action models from ignoring language by maximizing conditional pointwise mutual information via Bayesian decomposition with latent action queries.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identified information collapse is the dominant cause of poor OOD performance and that maximizing conditional PMI via the dual-branch architecture will reliably correct it without introducing new failure modes or requiring careful hyperparameter tuning.","pith_extraction_headline":"LangForce restores language grounding in vision-language-action models by maximizing conditional pointwise mutual information between instructions and actions."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"cb93c14ce138b235765f5cadb0d8806e4d2ef8f0c09bbee112fbb4c6887b7190"},"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":"2601.15197","created_at":"2026-05-18T03:09:24.337326+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.15197v6","created_at":"2026-05-18T03:09:24.337326+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.15197","created_at":"2026-05-18T03:09:24.337326+00:00"},{"alias_kind":"pith_short_12","alias_value":"CW7IWZ46YM7V","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"CW7IWZ46YM7VWQNJ","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"CW7IWZ46","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2603.22003","citing_title":"VP-VLA: Visual Prompting as an Interface for Vision-Language-Action Models","ref_index":24,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13757","citing_title":"FrameSkip: Learning from Fewer but More Informative Frames in VLA Training","ref_index":8,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":1,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG","json":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG.json","graph_json":"https://pith.science/api/pith-number/CW7IWZ46YM7VWQNJOBNGPBBPUG/graph.json","events_json":"https://pith.science/api/pith-number/CW7IWZ46YM7VWQNJOBNGPBBPUG/events.json","paper":"https://pith.science/paper/CW7IWZ46"},"agent_actions":{"view_html":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG","download_json":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG.json","view_paper":"https://pith.science/paper/CW7IWZ46","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.15197&json=true","fetch_graph":"https://pith.science/api/pith-number/CW7IWZ46YM7VWQNJOBNGPBBPUG/graph.json","fetch_events":"https://pith.science/api/pith-number/CW7IWZ46YM7VWQNJOBNGPBBPUG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG/action/storage_attestation","attest_author":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG/action/author_attestation","sign_citation":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG/action/citation_signature","submit_replication":"https://pith.science/pith/CW7IWZ46YM7VWQNJOBNGPBBPUG/action/replication_record"}},"created_at":"2026-05-18T03:09:24.337326+00:00","updated_at":"2026-05-18T03:09:24.337326+00:00"}