{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:2LXGIWHQJO4ZQ75ZHM7ANHHAZL","short_pith_number":"pith:2LXGIWHQ","schema_version":"1.0","canonical_sha256":"d2ee6458f04bb9987fb93b3e069ce0caf5460e370e1278035a3320487b238820","source":{"kind":"arxiv","id":"2506.04706","version":1},"attestation_state":"computed","paper":{"title":"Line of Sight: On Linear Representations in VLLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Achyuta Rajaram, Arthur Conmy, Jacob Andreas, Sarah Schwettmann","submitted_at":"2025-06-05T07:30:58Z","abstract_excerpt":"Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show that the features are causal by performing targeted edits on the model output. In order to increase the diversity of the studied linear features, we train multimodal Sparse Autoencoders (SAEs), creating"},"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":"2506.04706","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2025-06-05T07:30:58Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"3da183bd85f1d632ecadb5c1e38f25980583d23d5d1bfa16863a8b7cfac9a776","abstract_canon_sha256":"e4a6480c6e32b6e5f49daf3b8ffbf40da3ab951032ba73d1afca421919c01650"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:16:19.191830Z","signature_b64":"kRrCLKpaLQFWKX7YNZk+pEfyfKf0RfvBw+rHfy7A7USQ+t/wRn9uWJ+s1iW0JtzDF0ovT1bNZLezZb6VECsQDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d2ee6458f04bb9987fb93b3e069ce0caf5460e370e1278035a3320487b238820","last_reissued_at":"2026-07-05T11:16:19.191383Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:16:19.191383Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Line of Sight: On Linear Representations in VLLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Achyuta Rajaram, Arthur Conmy, Jacob Andreas, Sarah Schwettmann","submitted_at":"2025-06-05T07:30:58Z","abstract_excerpt":"Language models can be equipped with multimodal capabilities by fine-tuning on embeddings of visual inputs. But how do such multimodal models represent images in their hidden activations? We explore representations of image concepts within LlaVA-Next, a popular open-source VLLM. We find a diverse set of ImageNet classes represented via linearly decodable features in the residual stream. We show that the features are causal by performing targeted edits on the model output. In order to increase the diversity of the studied linear features, we train multimodal Sparse Autoencoders (SAEs), creating"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.04706","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/2506.04706/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":"2506.04706","created_at":"2026-07-05T11:16:19.191448+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.04706v1","created_at":"2026-07-05T11:16:19.191448+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.04706","created_at":"2026-07-05T11:16:19.191448+00:00"},{"alias_kind":"pith_short_12","alias_value":"2LXGIWHQJO4Z","created_at":"2026-07-05T11:16:19.191448+00:00"},{"alias_kind":"pith_short_16","alias_value":"2LXGIWHQJO4ZQ75Z","created_at":"2026-07-05T11:16:19.191448+00:00"},{"alias_kind":"pith_short_8","alias_value":"2LXGIWHQ","created_at":"2026-07-05T11:16:19.191448+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.07148","citing_title":"Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Models","ref_index":42,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL","json":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL.json","graph_json":"https://pith.science/api/pith-number/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/graph.json","events_json":"https://pith.science/api/pith-number/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/events.json","paper":"https://pith.science/paper/2LXGIWHQ"},"agent_actions":{"view_html":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL","download_json":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL.json","view_paper":"https://pith.science/paper/2LXGIWHQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.04706&json=true","fetch_graph":"https://pith.science/api/pith-number/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/graph.json","fetch_events":"https://pith.science/api/pith-number/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/action/storage_attestation","attest_author":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/action/author_attestation","sign_citation":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/action/citation_signature","submit_replication":"https://pith.science/pith/2LXGIWHQJO4ZQ75ZHM7ANHHAZL/action/replication_record"}},"created_at":"2026-07-05T11:16:19.191448+00:00","updated_at":"2026-07-05T11:16:19.191448+00:00"}