{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:5CB62F5S426VGG2FZUHOBOMXII","short_pith_number":"pith:5CB62F5S","schema_version":"1.0","canonical_sha256":"e883ed17b2e6bd531b45cd0ee0b9974214301e7121f8e8ad21dab941de352c22","source":{"kind":"arxiv","id":"2508.15260","version":1},"attestation_state":"computed","paper":{"title":"Deep Think with Confidence","license":"http://creativecommons.org/licenses/by/4.0/","headline":"DeepConf uses model-internal confidence to filter weak reasoning traces, reaching 99.9% accuracy on AIME 2025 while cutting tokens by 84.7%.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jiawei Zhao, Xuewei Wang, Yichao Fu, Yuandong Tian","submitted_at":"2025-08-21T05:48:38Z","abstract_excerpt":"Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional mo"},"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":"2508.15260","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-08-21T05:48:38Z","cross_cats_sorted":[],"title_canon_sha256":"d51da53a66a1207a671bc95f9f5b03e63ebb297d7d46f939aea43897eac3ed8c","abstract_canon_sha256":"dece95c2b7e1c506da16d2c512b9e0bf6c3b4fd1ef430486a5d981067387d028"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:48.026780Z","signature_b64":"1U1+34h5nugLI0uJRSixwlRnquhLcD1K26XjxdvmSyncamd3xjVWHbArvPSlUzy6XS0n/kfR69qglerqUJMTAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e883ed17b2e6bd531b45cd0ee0b9974214301e7121f8e8ad21dab941de352c22","last_reissued_at":"2026-05-17T23:38:48.026210Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:48.026210Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Deep Think with Confidence","license":"http://creativecommons.org/licenses/by/4.0/","headline":"DeepConf uses model-internal confidence to filter weak reasoning traces, reaching 99.9% accuracy on AIME 2025 while cutting tokens by 84.7%.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jiawei Zhao, Xuewei Wang, Yichao Fu, Yuandong Tian","submitted_at":"2025-08-21T05:48:38Z","abstract_excerpt":"Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional mo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That model-internal confidence signals reliably indicate the quality or correctness of individual reasoning traces, allowing effective dynamic filtering without additional training or tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DeepConf filters low-quality reasoning traces using model-internal confidence signals, reaching up to 99.9% accuracy on AIME 2025 while reducing generated tokens by up to 84.7% versus full parallel sampling.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DeepConf uses model-internal confidence to filter weak reasoning traces, reaching 99.9% accuracy on AIME 2025 while cutting tokens by 84.7%.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5db56bd422fe8c58ed9b577bcb5bc786e62c63c983f183ad154af04a9e6a406d"},"source":{"id":"2508.15260","kind":"arxiv","version":1},"verdict":{"id":"7485b2b8-cca9-4e0f-ab75-e669a96c37ed","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T11:26:17.715523Z","strongest_claim":"on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.","one_line_summary":"DeepConf filters low-quality reasoning traces using model-internal confidence signals, reaching up to 99.9% accuracy on AIME 2025 while reducing generated tokens by up to 84.7% versus full parallel sampling.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That model-internal confidence signals reliably indicate the quality or correctness of individual reasoning traces, allowing effective dynamic filtering without additional training or tuning.","pith_extraction_headline":"DeepConf uses model-internal confidence to filter weak reasoning traces, reaching 99.9% accuracy on AIME 2025 while cutting tokens by 84.7%."},"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":"2508.15260","created_at":"2026-05-17T23:38:48.026295+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.15260v1","created_at":"2026-05-17T23:38:48.026295+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.15260","created_at":"2026-05-17T23:38:48.026295+00:00"},{"alias_kind":"pith_short_12","alias_value":"5CB62F5S426V","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"5CB62F5S426VGG2F","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"5CB62F5S","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":27,"internal_anchor_count":27,"sample":[{"citing_arxiv_id":"2505.11737","citing_title":"TokUR: Token-Level Uncertainty Estimation for Large Language Model Reasoning","ref_index":12,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22389","citing_title":"Unified Data Selection for LLM Reasoning","ref_index":39,"is_internal_anchor":true},{"citing_arxiv_id":"2602.01970","citing_title":"Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2509.09438","citing_title":"GrACE: A Generative Approach to Better Confidence Elicitation and Efficient Test-Time Scaling in Large Language Models","ref_index":41,"is_internal_anchor":true},{"citing_arxiv_id":"2510.07286","citing_title":"Evolutionary Profiles for Protein Fitness Prediction","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2510.14232","citing_title":"Scaling Test-Time Compute to Achieve IOI Gold Medal with Open-Weight Models","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2510.24941","citing_title":"Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2509.26645","citing_title":"TTT3R: 3D Reconstruction as Test-Time Training","ref_index":32,"is_internal_anchor":true},{"citing_arxiv_id":"2601.20255","citing_title":"HE-SNR: Uncovering Latent Logic via Entropy for Guiding Mid-Training on SWE-bench","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2601.21619","citing_title":"On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2602.06019","citing_title":"Multi-Token Prediction via Self-Distillation","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2605.14062","citing_title":"Know When To Fold 'Em: Token-Efficient LLM Synthetic Data Generation via Multi-Stage In-Flight Rejection","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2604.02543","citing_title":"Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08083","citing_title":"LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2605.10318","citing_title":"Extending Confidence-Based Text2Cypher with Grammar and Schema Aware Filtering","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23505","citing_title":"Uncertainty Propagation in LLM-Based Systems","ref_index":57,"is_internal_anchor":true},{"citing_arxiv_id":"2604.23333","citing_title":"Process Supervision of Confidence Margin for Calibrated LLM Reasoning","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.01133","citing_title":"When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems","ref_index":24,"is_internal_anchor":true},{"citing_arxiv_id":"2604.11025","citing_title":"Test-time Scaling over Perception: Resolving the Grounding Paradox in Thinking with Images","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2604.09850","citing_title":"Training-Free Object-Background Compositional T2I via Dynamic Spatial Guidance and Multi-Path Pruning","ref_index":17,"is_internal_anchor":true},{"citing_arxiv_id":"2604.07484","citing_title":"ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08083","citing_title":"LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15972","citing_title":"Weak-Link Optimization for Multi-Agent Reasoning and Collaboration","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2604.16029","citing_title":"Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15623","citing_title":"Overmind NSA: A Unified Neuro-Symbolic Computing Architecture with Approximate Nonlinear Activations and Preemptive Memory Bypass","ref_index":14,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII","json":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII.json","graph_json":"https://pith.science/api/pith-number/5CB62F5S426VGG2FZUHOBOMXII/graph.json","events_json":"https://pith.science/api/pith-number/5CB62F5S426VGG2FZUHOBOMXII/events.json","paper":"https://pith.science/paper/5CB62F5S"},"agent_actions":{"view_html":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII","download_json":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII.json","view_paper":"https://pith.science/paper/5CB62F5S","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.15260&json=true","fetch_graph":"https://pith.science/api/pith-number/5CB62F5S426VGG2FZUHOBOMXII/graph.json","fetch_events":"https://pith.science/api/pith-number/5CB62F5S426VGG2FZUHOBOMXII/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII/action/timestamp_anchor","attest_storage":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII/action/storage_attestation","attest_author":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII/action/author_attestation","sign_citation":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII/action/citation_signature","submit_replication":"https://pith.science/pith/5CB62F5S426VGG2FZUHOBOMXII/action/replication_record"}},"created_at":"2026-05-17T23:38:48.026295+00:00","updated_at":"2026-05-17T23:38:48.026295+00:00"}