{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:FUC5QTYODKCYPCETJA4QVTIG4M","short_pith_number":"pith:FUC5QTYO","schema_version":"1.0","canonical_sha256":"2d05d84f0e1a8587889348390acd06e33be44db6b856baeddb4d47482cda9434","source":{"kind":"arxiv","id":"2503.14337","version":2},"attestation_state":"computed","paper":{"title":"PENCIL: Long Thoughts with Short Memory","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Chenxiao Yang, David McAllester, Nathan Srebro, Zhiyuan Li","submitted_at":"2025-03-18T15:14:14Z","abstract_excerpt":"While state-of-the-art LLMs have demonstrated great promise of using long Chains-of-Thought (CoT) to boost reasoning, scaling it up to more challenging problems at test-time is fundamentally limited by suboptimal memory usage -- intermediate computations accumulate indefinitely in context even when no longer needed for future thoughts. We introduce PENCIL, which incorporates a novel reduction mechanism into the autoregressive generation process that recursively cleans up intermediate thoughts based on patterns learned from training. By iteratively generating and erasing thoughts, PENCIL can th"},"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":"2503.14337","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-03-18T15:14:14Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"af00b24b721caf2786e78318bfdd77ec29864ebfc137d12b9a590dc123e2a292","abstract_canon_sha256":"01253a6fd7405de1460df0117401afe19bd5e4db9c6d968a367a10bd38ce1ca1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:32:26.389214Z","signature_b64":"hSqyjm7j7OeKQqMaGi+T+c0Qa0q5xSMbhmaaVSmGOXneESmzhFdS4jp2DWxZfquJYcGkhKuIUKp9doo1N11HBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2d05d84f0e1a8587889348390acd06e33be44db6b856baeddb4d47482cda9434","last_reissued_at":"2026-07-05T11:32:26.388715Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:32:26.388715Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PENCIL: Long Thoughts with Short Memory","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Chenxiao Yang, David McAllester, Nathan Srebro, Zhiyuan Li","submitted_at":"2025-03-18T15:14:14Z","abstract_excerpt":"While state-of-the-art LLMs have demonstrated great promise of using long Chains-of-Thought (CoT) to boost reasoning, scaling it up to more challenging problems at test-time is fundamentally limited by suboptimal memory usage -- intermediate computations accumulate indefinitely in context even when no longer needed for future thoughts. We introduce PENCIL, which incorporates a novel reduction mechanism into the autoregressive generation process that recursively cleans up intermediate thoughts based on patterns learned from training. By iteratively generating and erasing thoughts, PENCIL can th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.14337","kind":"arxiv","version":2},"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/2503.14337/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":"2503.14337","created_at":"2026-07-05T11:32:26.388774+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.14337v2","created_at":"2026-07-05T11:32:26.388774+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.14337","created_at":"2026-07-05T11:32:26.388774+00:00"},{"alias_kind":"pith_short_12","alias_value":"FUC5QTYODKCY","created_at":"2026-07-05T11:32:26.388774+00:00"},{"alias_kind":"pith_short_16","alias_value":"FUC5QTYODKCYPCET","created_at":"2026-07-05T11:32:26.388774+00:00"},{"alias_kind":"pith_short_8","alias_value":"FUC5QTYO","created_at":"2026-07-05T11:32:26.388774+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":11,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07423","citing_title":"The Optimal Sample Complexity of Learning Autoregressive Chain-of-Thought","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2606.21943","citing_title":"Modularized Reinforcement Learning on LLMs: From MDP Creation to Exploration and Learning","ref_index":242,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09731","citing_title":"Tight Sample Complexity of Transformers","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01532","citing_title":"Rethinking the Role of Positional Encoding: Sliding-Window Transformers without PE Remain Turing Complete","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20531","citing_title":"Pseudo-Formalization for Automatic Proof Verification","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20531","citing_title":"Pseudo-Formalization for Automatic Proof Verification","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14457","citing_title":"Stateful Reasoning via Insight Replay","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19358","citing_title":"Taming the Thinker: Conditional Entropy Shaping for Adaptive LLM Reasoning","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14457","citing_title":"Stateful Reasoning via Insight Replay","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05909","citing_title":"Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09852","citing_title":"MEMENTO: Teaching LLMs to Manage Their Own Context","ref_index":33,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M","json":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M.json","graph_json":"https://pith.science/api/pith-number/FUC5QTYODKCYPCETJA4QVTIG4M/graph.json","events_json":"https://pith.science/api/pith-number/FUC5QTYODKCYPCETJA4QVTIG4M/events.json","paper":"https://pith.science/paper/FUC5QTYO"},"agent_actions":{"view_html":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M","download_json":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M.json","view_paper":"https://pith.science/paper/FUC5QTYO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.14337&json=true","fetch_graph":"https://pith.science/api/pith-number/FUC5QTYODKCYPCETJA4QVTIG4M/graph.json","fetch_events":"https://pith.science/api/pith-number/FUC5QTYODKCYPCETJA4QVTIG4M/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M/action/storage_attestation","attest_author":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M/action/author_attestation","sign_citation":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M/action/citation_signature","submit_replication":"https://pith.science/pith/FUC5QTYODKCYPCETJA4QVTIG4M/action/replication_record"}},"created_at":"2026-07-05T11:32:26.388774+00:00","updated_at":"2026-07-05T11:32:26.388774+00:00"}