{"paper":{"title":"Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Language models self-improve under RL when they already use reasoning behaviors like verification and backtracking, even if answers start wrong.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Anikait Singh, Ayush Chakravarthy, Kanishk Gandhi, Nathan Lile, Noah D. Goodman","submitted_at":"2025-03-03T08:46:22Z","abstract_excerpt":"Test-time inference has emerged as a powerful paradigm for enabling language models to ``think'' longer and more carefully about complex challenges, much like skilled human experts. While reinforcement learning (RL) can drive self-improvement in language models on verifiable tasks, some models exhibit substantial gains while others quickly plateau. For instance, we find that Qwen-2.5-3B far exceeds Llama-3.2-3B under identical RL training for the game of Countdown. This discrepancy raises a critical question: what intrinsic properties enable effective self-improvement? We introduce a framework"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"the presence of reasoning behaviors, rather than correctness of answers, proves to be the critical factor -- models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the four identified cognitive behaviors are the primary causal drivers of self-improvement differences, and that the controlled priming experiments isolate their effect without confounding influences from model architecture, training history, or unmeasured variables.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Language models that naturally exhibit verification, backtracking, subgoal setting, and backward chaining improve substantially during RL on verifiable tasks, and these behaviors can be instilled via priming with reasoning-focused examples or filtered pretraining to enable self-improvement.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Language models self-improve under RL when they already use reasoning behaviors like verification and backtracking, even if answers start wrong.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b3b7f7ed2aee6989940632b64d448131598ff90e729b43e131a4114ab3d29fe4"},"source":{"id":"2503.01307","kind":"arxiv","version":2},"verdict":{"id":"d32bc172-cd96-443c-8f1d-e4c9ed08c525","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T11:35:31.116346Z","strongest_claim":"the presence of reasoning behaviors, rather than correctness of answers, proves to be the critical factor -- models primed with incorrect solutions containing proper reasoning patterns achieve comparable performance to those trained on correct solutions.","one_line_summary":"Language models that naturally exhibit verification, backtracking, subgoal setting, and backward chaining improve substantially during RL on verifiable tasks, and these behaviors can be instilled via priming with reasoning-focused examples or filtered pretraining to enable self-improvement.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the four identified cognitive behaviors are the primary causal drivers of self-improvement differences, and that the controlled priming experiments isolate their effect without confounding influences from model architecture, training history, or unmeasured variables.","pith_extraction_headline":"Language models self-improve under RL when they already use reasoning behaviors like verification and backtracking, even if answers start wrong."},"references":{"count":13,"sample":[{"doi":"","year":2024,"title":"REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization","work_id":"557f9e99-cb00-4dd2-92fd-67ddcddbb35d","ref_index":1,"cited_arxiv_id":"2501.03262","is_internal_anchor":true},{"doi":"10.1007/bf00992696","year":1971,"title":"Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters","work_id":"a8d50b24-bdf5-46ed-bc4f-2927dfd81f1d","ref_index":2,"cited_arxiv_id":"2408.03314","is_internal_anchor":true},{"doi":"","year":null,"title":"Backtracking Only: This dataset focuses exclusively on the backtracking strategy, where the model explores solution paths and retreats when encountering dead ends","work_id":"f16db8fe-855e-4c6c-a8fa-f1b9b2d94974","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Backtracking with Answer Verification: In addition to backtracking, this dataset incorporates answer verification, where the model checks its intermediate solutions with the target number","work_id":"21872aad-715f-42df-b5f6-c07cc9b3f044","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Backtracking with Subgoal Setting: This dataset combines backtracking with explicit subgoal setting, where the model breaks down complex problems into manageable intermediate steps","work_id":"6ed3f964-44b3-49d8-9995-34339ece07bc","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":13,"snapshot_sha256":"d2d6933031d2432c2eabe21412e254ed5ecfd3aff4e4e56a5ba08b58f7538823","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"854cd3056e3b064c23e435fb4997486e49c6b91dacf8af2db1acc702d7ea385b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}