{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:HVONQ4JDUBXOGVDVSBGHTGUV7A","short_pith_number":"pith:HVONQ4JD","canonical_record":{"source":{"id":"2603.19294","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-10T21:00:05Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"8fdfba85b3e55e2ca0764ff3f5f1036d3511394e5b141dabdb9c42fffa7dc91c","abstract_canon_sha256":"5baac0306ed355e8ac347a9fed882e498f70cbff830c853bd7aa626a90944280"},"schema_version":"1.0"},"canonical_sha256":"3d5cd87123a06ee35475904c799a95f81df9e4268d472b01ee1ea9baf990069a","source":{"kind":"arxiv","id":"2603.19294","version":3},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.19294","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"arxiv_version","alias_value":"2603.19294v3","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.19294","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_12","alias_value":"HVONQ4JDUBXO","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_16","alias_value":"HVONQ4JDUBXOGVDV","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_8","alias_value":"HVONQ4JD","created_at":"2026-05-29T01:04:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:HVONQ4JDUBXOGVDVSBGHTGUV7A","target":"record","payload":{"canonical_record":{"source":{"id":"2603.19294","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-10T21:00:05Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"8fdfba85b3e55e2ca0764ff3f5f1036d3511394e5b141dabdb9c42fffa7dc91c","abstract_canon_sha256":"5baac0306ed355e8ac347a9fed882e498f70cbff830c853bd7aa626a90944280"},"schema_version":"1.0"},"canonical_sha256":"3d5cd87123a06ee35475904c799a95f81df9e4268d472b01ee1ea9baf990069a","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-29T01:04:37.183514Z","signature_b64":"tZCKIzp+CgO0oCDT48zVYUyBvHuguMFky4FRpQL9hlKHPuzXJ6SALlg7MZyTx1jk52N2bm4nfIaTvKlpFo5MAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"3d5cd87123a06ee35475904c799a95f81df9e4268d472b01ee1ea9baf990069a","last_reissued_at":"2026-05-29T01:04:37.182899Z","signature_status":"signed_v1","first_computed_at":"2026-05-29T01:04:37.182899Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2603.19294","source_version":3,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-29T01:04:37Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"8unsxPVmdQjdzZAxfrXeI3c9+n4+cpEh0HcW7Z2qCHONHfHKeRU0OF5Xk0sTNUcF8zC8weJ3k8q5868UTNo+Cg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T11:43:43.702022Z"},"content_sha256":"97056dfe586f066490ccbf31e2357c2ce980643754cf8f7f0d76625cc5129347","schema_version":"1.0","event_id":"sha256:97056dfe586f066490ccbf31e2357c2ce980643754cf8f7f0d76625cc5129347"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:HVONQ4JDUBXOGVDVSBGHTGUV7A","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Haoran Li, Hyunji Nam, Natasha Jaques","submitted_at":"2026-03-10T21:00:05Z","abstract_excerpt":"While post-training has successfully improved large language models (LLMs) across a variety of domains, these gains heavily rely on human-labeled data or external verifiers. Existing data has already been exploited, and new data is expensive to collect. Moreover, true intelligence goes far beyond verifiable tasks. Therefore, we need self-improvement frameworks that are less dependent on external signals and more broadly applicable to both verifiable and non-verifiable domains. We propose **Mutual Information Preference Optimization (MIPO)**, a contrastive data augmentation method that construc"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"using Direct Preference Optimization (DPO) to learn from this paired data maximizes pointwise conditional mutual information (MI), under the base LLM, between prompts and model responses","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That responses generated by conditioning on random unrelated prompts form sufficiently informative negatives such that DPO on the resulting pairs actually maximizes the desired pointwise conditional mutual information and produces measurable downstream gains","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tasks without new data or oversight.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5d66e2c8e01d9fba86dc0a2854d1d223fe5d618f37bba02bfbf446fb7541e9eb"},"source":{"id":"2603.19294","kind":"arxiv","version":3},"verdict":{"id":"d004cad5-a734-4e1b-93b2-e0fc7ff29d3c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T12:49:19.894854Z","strongest_claim":"using Direct Preference Optimization (DPO) to learn from this paired data maximizes pointwise conditional mutual information (MI), under the base LLM, between prompts and model responses","one_line_summary":"MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tasks without new data or oversight.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That responses generated by conditioning on random unrelated prompts form sufficiently informative negatives such that DPO on the resulting pairs actually maximizes the desired pointwise conditional mutual information and produces measurable downstream gains","pith_extraction_headline":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.19294/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":2,"snapshot_sha256":"c95ff5f55d646463a3c8564f3625f562ee1de91fc0232f7e5e6f7f98add2bdc2"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"d004cad5-a734-4e1b-93b2-e0fc7ff29d3c"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-29T01:04:37Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"W1MwOpf05329FC/MR2ryDRQ8Zq9HwQAPrMURHcIEgA1jZr9yPe/5so6OEuSP2JQuOnMNaGMtXH2YOA4kZHlhAg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-10T11:43:43.703001Z"},"content_sha256":"6968b92f3d5aa7f0ea41146e3bea5f4297dcb4297a75b754627d068e79b68ee1","schema_version":"1.0","event_id":"sha256:6968b92f3d5aa7f0ea41146e3bea5f4297dcb4297a75b754627d068e79b68ee1"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/bundle.json","state_url":"https://pith.science/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-10T11:43:43Z","links":{"resolver":"https://pith.science/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A","bundle":"https://pith.science/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/bundle.json","state":"https://pith.science/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/state.json","well_known_bundle":"https://pith.science/.well-known/pith/HVONQ4JDUBXOGVDVSBGHTGUV7A/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:HVONQ4JDUBXOGVDVSBGHTGUV7A","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"5baac0306ed355e8ac347a9fed882e498f70cbff830c853bd7aa626a90944280","cross_cats_sorted":["cs.AI","cs.CL"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-10T21:00:05Z","title_canon_sha256":"8fdfba85b3e55e2ca0764ff3f5f1036d3511394e5b141dabdb9c42fffa7dc91c"},"schema_version":"1.0","source":{"id":"2603.19294","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2603.19294","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"arxiv_version","alias_value":"2603.19294v3","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.19294","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_12","alias_value":"HVONQ4JDUBXO","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_16","alias_value":"HVONQ4JDUBXOGVDV","created_at":"2026-05-29T01:04:37Z"},{"alias_kind":"pith_short_8","alias_value":"HVONQ4JD","created_at":"2026-05-29T01:04:37Z"}],"graph_snapshots":[{"event_id":"sha256:6968b92f3d5aa7f0ea41146e3bea5f4297dcb4297a75b754627d068e79b68ee1","target":"graph","created_at":"2026-05-29T01:04:37Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"using Direct Preference Optimization (DPO) to learn from this paired data maximizes pointwise conditional mutual information (MI), under the base LLM, between prompts and model responses"},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That responses generated by conditioning on random unrelated prompts form sufficiently informative negatives such that DPO on the resulting pairs actually maximizes the desired pointwise conditional mutual information and produces measurable downstream gains"},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tasks without new data or oversight."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight."}],"snapshot_sha256":"5d66e2c8e01d9fba86dc0a2854d1d223fe5d618f37bba02bfbf446fb7541e9eb"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c95ff5f55d646463a3c8564f3625f562ee1de91fc0232f7e5e6f7f98add2bdc2"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2603.19294/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"While post-training has successfully improved large language models (LLMs) across a variety of domains, these gains heavily rely on human-labeled data or external verifiers. Existing data has already been exploited, and new data is expensive to collect. Moreover, true intelligence goes far beyond verifiable tasks. Therefore, we need self-improvement frameworks that are less dependent on external signals and more broadly applicable to both verifiable and non-verifiable domains. We propose **Mutual Information Preference Optimization (MIPO)**, a contrastive data augmentation method that construc","authors_text":"Haoran Li, Hyunji Nam, Natasha Jaques","cross_cats":["cs.AI","cs.CL"],"headline":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight.","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-10T21:00:05Z","title":"Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.19294","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-15T12:49:19.894854Z","id":"d004cad5-a734-4e1b-93b2-e0fc7ff29d3c","model_set":{"reader":"grok-4.3"},"one_line_summary":"MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tasks without new data or oversight.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Maximizing mutual information between prompts and responses lets LLMs improve personalization and problem solving without new data or oversight.","strongest_claim":"using Direct Preference Optimization (DPO) to learn from this paired data maximizes pointwise conditional mutual information (MI), under the base LLM, between prompts and model responses","weakest_assumption":"That responses generated by conditioning on random unrelated prompts form sufficiently informative negatives such that DPO on the resulting pairs actually maximizes the desired pointwise conditional mutual information and produces measurable downstream gains"}},"verdict_id":"d004cad5-a734-4e1b-93b2-e0fc7ff29d3c"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:97056dfe586f066490ccbf31e2357c2ce980643754cf8f7f0d76625cc5129347","target":"record","created_at":"2026-05-29T01:04:37Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"5baac0306ed355e8ac347a9fed882e498f70cbff830c853bd7aa626a90944280","cross_cats_sorted":["cs.AI","cs.CL"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2026-03-10T21:00:05Z","title_canon_sha256":"8fdfba85b3e55e2ca0764ff3f5f1036d3511394e5b141dabdb9c42fffa7dc91c"},"schema_version":"1.0","source":{"id":"2603.19294","kind":"arxiv","version":3}},"canonical_sha256":"3d5cd87123a06ee35475904c799a95f81df9e4268d472b01ee1ea9baf990069a","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"3d5cd87123a06ee35475904c799a95f81df9e4268d472b01ee1ea9baf990069a","first_computed_at":"2026-05-29T01:04:37.182899Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-29T01:04:37.182899Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"tZCKIzp+CgO0oCDT48zVYUyBvHuguMFky4FRpQL9hlKHPuzXJ6SALlg7MZyTx1jk52N2bm4nfIaTvKlpFo5MAA==","signature_status":"signed_v1","signed_at":"2026-05-29T01:04:37.183514Z","signed_message":"canonical_sha256_bytes"},"source_id":"2603.19294","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:97056dfe586f066490ccbf31e2357c2ce980643754cf8f7f0d76625cc5129347","sha256:6968b92f3d5aa7f0ea41146e3bea5f4297dcb4297a75b754627d068e79b68ee1"],"state_sha256":"1f6d265845d3ec533efd1cb1c18860dfa36dd51faa0f5c91b9806c6c519084b6"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"NE0yYzJoQOPiBqsQMmcktfV7jyGLD8ttZ/9GyLZWkdovzWgNm1XrsDjUZHlbpND81wTJ6nvVmLKIw0MrpAnpAw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-10T11:43:43.708087Z","bundle_sha256":"79e49ec62b09fcc5e159af56b02d6a8733767cfe5aa3cd6d4ae5b606476c09a9"}}