{"paper":{"title":"Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Omnimodal LLMs encode premise-perception mismatches in hidden states but almost never reject the conflicting claims in their outputs.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Fanyi Pu, Kaichen Zhang, Shuo Sun, Trung Nguyen Quang, Yiming Gao, Ziwei Liu","submitted_at":"2026-05-13T16:14:44Z","abstract_excerpt":"When an omnimodal large language model accepts a question whose textual premise contradicts what it actually sees or hears, does the failure lie in perception or in action? Recent omnimodal models are positioned as perception-grounded agents that jointly process video, audio, and text, yet a basic form of grounding remains untested: catching a textual claim that conflicts with the model's own sensory input. We introduce IMAVB, a curated 500-clip benchmark of long-form movies with a 2x2 design crossing target modality (vision, audio) and premise condition (standard, misleading), which lets us m"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"hidden states reliably encode premise-perception mismatches even when the same models almost never reject the false claim in their outputs","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the IMAVB benchmark's curated misleading premises cleanly isolate grounding failures without introducing confounds from clip selection, question phrasing, or model-specific training data, and that probe-based detection of hidden-state mismatches accurately reflects functional encoding.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Omnimodal LLMs encode premise-perception mismatches in hidden states yet almost never reject false textual claims, exposing a representation-action gap that is modality-asymmetric and prompt-resistant.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Omnimodal LLMs encode premise-perception mismatches in hidden states but almost never reject the conflicting claims in their outputs.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"00b3ef468fc7f11ebb122098271f2a83121b4db0528edfff7e855bc1b2bccb7d"},"source":{"id":"2605.13737","kind":"arxiv","version":1},"verdict":{"id":"009deb09-86c0-4403-82bf-ac51277f9fda","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:03:22.061288Z","strongest_claim":"hidden states reliably encode premise-perception mismatches even when the same models almost never reject the false claim in their outputs","one_line_summary":"Omnimodal LLMs encode premise-perception mismatches in hidden states yet almost never reject false textual claims, exposing a representation-action gap that is modality-asymmetric and prompt-resistant.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the IMAVB benchmark's curated misleading premises cleanly isolate grounding failures without introducing confounds from clip selection, question phrasing, or model-specific training data, and that probe-based detection of hidden-state mismatches accurately reflects functional encoding.","pith_extraction_headline":"Omnimodal LLMs encode premise-perception mismatches in hidden states but almost never reject the conflicting claims in their outputs."},"references":{"count":99,"sample":[{"doi":"","year":2026,"title":"OpenAI GPT-5 System Card","work_id":"ca87689a-0d29-4476-b504-b65dbbb08af4","ref_index":1,"cited_arxiv_id":"2601.03267","is_internal_anchor":true},{"doi":"","year":2026,"title":"A new era of intelligence with Gemini 3, 2026","work_id":"70678dd4-c18e-407f-a968-a0b8180cf72a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Qwen3-omni technical report, 2025","work_id":"2421f421-da09-44ef-b96f-2a64fba6a7ef","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Amos Azaria and Tom M. Mitchell. The internal state of an llm knows when it’s lying, 2023","work_id":"ed116ea6-70d9-4c0f-90db-16fe14dcc3a7","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Discovering latent knowledge in language models without supervision","work_id":"db51027f-63a1-4026-b9c6-6df48de0cf76","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":99,"snapshot_sha256":"a67f6b2d0dce6c7852eba2d59c852fa2e4a799e1074b11a14dc3994b61ca4a2f","internal_anchors":3},"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"}