{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:32AGNVESK7C6YI24VJUC4G546R","short_pith_number":"pith:32AGNVES","schema_version":"1.0","canonical_sha256":"de8066d49257c5ec235caa682e1bbcf446f38f4239e1848df90b5c9d030327c8","source":{"kind":"arxiv","id":"2502.08524","version":1},"attestation_state":"computed","paper":{"title":"LLM Pretraining with Continuous Concepts","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Andrew Cohen, Ilia Kulikov, Jack Lanchantin, Jane Yu, Janice Lan, Jason Weston, Jihoon Tack, Shibo Hao, Xian Li, Yuandong Tian","submitted_at":"2025-02-12T16:00:11Z","abstract_excerpt":"Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts continuous concepts learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling "},"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":"2502.08524","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-02-12T16:00:11Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"7fe977c8f17f7b8ba023c59cff72900b5b66bd6e5d01f452accf094fb260ed69","abstract_canon_sha256":"0b47c7623be47e9f9c7e19780f79d3b11ebbbcfdd9776dbeeff9a3fe66a0317d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:13:24.359162Z","signature_b64":"BHOF7WE1zE1SkpxskD1hRlTHSx1UkwgnFiPMX8nS21p8XOBohPduKz42d6Q7+NSaO+OA9YDQ38+2OjoRp0frDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"de8066d49257c5ec235caa682e1bbcf446f38f4239e1848df90b5c9d030327c8","last_reissued_at":"2026-07-05T10:13:24.358671Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:13:24.358671Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LLM Pretraining with Continuous Concepts","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Andrew Cohen, Ilia Kulikov, Jack Lanchantin, Jane Yu, Janice Lan, Jason Weston, Jihoon Tack, Shibo Hao, Xian Li, Yuandong Tian","submitted_at":"2025-02-12T16:00:11Z","abstract_excerpt":"Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts continuous concepts learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.08524","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/2502.08524/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":"2502.08524","created_at":"2026-07-05T10:13:24.358731+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.08524v1","created_at":"2026-07-05T10:13:24.358731+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.08524","created_at":"2026-07-05T10:13:24.358731+00:00"},{"alias_kind":"pith_short_12","alias_value":"32AGNVESK7C6","created_at":"2026-07-05T10:13:24.358731+00:00"},{"alias_kind":"pith_short_16","alias_value":"32AGNVESK7C6YI24","created_at":"2026-07-05T10:13:24.358731+00:00"},{"alias_kind":"pith_short_8","alias_value":"32AGNVES","created_at":"2026-07-05T10:13:24.358731+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.06648","citing_title":"Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.24956","citing_title":"NITP: Next Implicit Token Prediction for LLM Pre-training","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24956","citing_title":"NITP: Next Implicit Token Prediction for LLM Pre-training","ref_index":48,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27734","citing_title":"Learn from your own latents and not from tokens: A sample-complexity theory","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30343","citing_title":"Unlocking the Working Memory of Large Language Models for Latent Reasoning","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2606.07604","citing_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","ref_index":127,"is_internal_anchor":false},{"citing_arxiv_id":"2504.02181","citing_title":"A Survey of Scaling in Large Language Model Reasoning","ref_index":189,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20075","citing_title":"CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06285","citing_title":"LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG","ref_index":31,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17892","citing_title":"LEPO: Latent Reasoning Policy Optimization for Large Language Models","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.21027","citing_title":"HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering","ref_index":95,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R","json":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R.json","graph_json":"https://pith.science/api/pith-number/32AGNVESK7C6YI24VJUC4G546R/graph.json","events_json":"https://pith.science/api/pith-number/32AGNVESK7C6YI24VJUC4G546R/events.json","paper":"https://pith.science/paper/32AGNVES"},"agent_actions":{"view_html":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R","download_json":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R.json","view_paper":"https://pith.science/paper/32AGNVES","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.08524&json=true","fetch_graph":"https://pith.science/api/pith-number/32AGNVESK7C6YI24VJUC4G546R/graph.json","fetch_events":"https://pith.science/api/pith-number/32AGNVESK7C6YI24VJUC4G546R/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R/action/timestamp_anchor","attest_storage":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R/action/storage_attestation","attest_author":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R/action/author_attestation","sign_citation":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R/action/citation_signature","submit_replication":"https://pith.science/pith/32AGNVESK7C6YI24VJUC4G546R/action/replication_record"}},"created_at":"2026-07-05T10:13:24.358731+00:00","updated_at":"2026-07-05T10:13:24.358731+00:00"}