{"paper":{"title":"Learning to Reason without External Rewards","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Large language models can improve at reasoning by using only their own internal confidence as the reward signal.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Aosong Feng, Dawn Song, Sergey Levine, Xuandong Zhao, Zhewei Kang","submitted_at":"2025-05-26T07:01:06Z","abstract_excerpt":"Training large language models (LLMs) for complex reasoning via Reinforcement Learning with Verifiable Rewards (RLVR) is effective but limited by reliance on costly, domain-specific supervision. We explore Reinforcement Learning from Internal Feedback (RLIF), a framework that enables LLMs to learn from intrinsic signals without external rewards or labeled data. We propose Intuitor, an RLIF method that uses a model's own confidence-termed self-certainty-as its sole reward signal. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, enabling"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks while achieving better generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the model's self-certainty score reliably indicates correct reasoning and does not encourage reward hacking or overconfidence in incorrect outputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Intuitor enables LLMs to learn complex reasoning from self-certainty signals alone, matching supervised RL performance on math benchmarks while generalizing better to code generation without gold solutions or test cases.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models can improve at reasoning by using only their own internal confidence as the reward signal.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"413e65f2c0095f14fbe0af388adc5b7212507fe4807fe16163e3629ff6355cda"},"source":{"id":"2505.19590","kind":"arxiv","version":4},"verdict":{"id":"7135c8df-9dc2-4fae-9f8f-bbd33c2ee816","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:13:34.177291Z","strongest_claim":"Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks while achieving better generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases.","one_line_summary":"Intuitor enables LLMs to learn complex reasoning from self-certainty signals alone, matching supervised RL performance on math benchmarks while generalizing better to code generation without gold solutions or test cases.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the model's self-certainty score reliably indicates correct reasoning and does not encourage reward hacking or overconfidence in incorrect outputs.","pith_extraction_headline":"Large language models can improve at reasoning by using only their own internal confidence as the reward signal."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"bea6854910e03788df088a3ea5ad244a858f860c0aacd3932bfb93c950a30740"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}