{"paper":{"title":"Conformal Thinking: Risk Control for Reasoning on a Compute Budget","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Distribution-free risk control sets upper and lower thresholds so LLMs stop reasoning early while keeping error rates below a user target.","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Alvin Zhang, Anushri Suresh, Benjamin Van Durme, Daniel Khashabi, Eric Nalisnick, Mehrdad Farajtabar, Rishi More, William Jurayj, Xi Wang","submitted_at":"2026-02-03T18:17:22Z","abstract_excerpt":"Reasoning Large Language Models (LLMs) enable test-time scaling, with dataset-level accuracy improving as the token budget increases, motivating adaptive reasoning -- spending tokens when they improve reliability and stopping early when additional computation is unlikely to help. However, setting the token budget, as well as the threshold for adaptive reasoning, is a practical challenge that entails a fundamental risk-accuracy trade-off. We re-frame the budget setting problem as risk control, limiting the error rate while minimizing compute. Our framework introduces an upper threshold that sto"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. For scenarios with multiple budget controlling criteria, we incorporate an efficiency loss to select the most computationally efficient exiting mechanism.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The validation set is representative of future test instances so that the distribution-free risk guarantees transfer; the parametric form chosen for the lower threshold is flexible enough to capture unsolvable cases without excessive premature stopping.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Conformal risk control with upper and lower thresholds lets LLMs adaptively stop reasoning while guaranteeing a maximum error rate and minimizing token use.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Distribution-free risk control sets upper and lower thresholds so LLMs stop reasoning early while keeping error rates below a user target.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"eb61cdec9e6ced04830e638a1f0c94d35e96bcfa5561e38448309840e575fb9e"},"source":{"id":"2602.03814","kind":"arxiv","version":2},"verdict":{"id":"5cbab2e9-02e3-4482-aab5-7ab1af81fa57","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T07:43:16.978409Z","strongest_claim":"Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. For scenarios with multiple budget controlling criteria, we incorporate an efficiency loss to select the most computationally efficient exiting mechanism.","one_line_summary":"Conformal risk control with upper and lower thresholds lets LLMs adaptively stop reasoning while guaranteeing a maximum error rate and minimizing token use.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The validation set is representative of future test instances so that the distribution-free risk guarantees transfer; the parametric form chosen for the lower threshold is flexible enough to capture unsolvable cases without excessive premature stopping.","pith_extraction_headline":"Distribution-free risk control sets upper and lower thresholds so LLMs stop reasoning early while keeping error rates below a user target."},"references":{"count":13,"sample":[{"doi":"","year":null,"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","ref_index":1,"cited_arxiv_id":"2501.12948","is_internal_anchor":true},{"doi":"","year":null,"title":"Efficiently serving llm reasoning programs with certaindex","work_id":"57494d13-0d57-400a-8e38-67a61f1677d1","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"URL https: //aclanthology.org/2025.acl-short.50/. Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000)","work_id":"1ec88e82-b25f-41d9-bd00-eb3bdab29850","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Fractured chain-of-thought reasoning","work_id":"bd0660ba-832d-46f0-90d4-ba4ae2efb956","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Xin Liu and Lu Wang","work_id":"125b44ab-fe4c-401f-9212-9a320bdf3efd","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":13,"snapshot_sha256":"80749694133f90a9d50e2d789320e93f7e436ffca6d52d3ec737e0f953a24c4d","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"487b269d6cc23194c947d055142fe5f8bd21392509f0a6156b82033ad5006ee0"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}