{"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"}