Research map / Large language models / Reasoning and chain of thought

Deep Think with Confidence

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DeepConf, which uses local token-confidence signals (lowest/bottom-10%/tail group confidence) to weight, filter, and early-stop reasoning traces in parallel test-time scaling, improving accuracy and cutting tokens versus self-consistency.

Problem

Parallel test-time scaling via self-consistency majority voting has diminishing returns and high token cost, and global trace-level confidence cannot detect local reasoning breakdowns or stop low-quality traces early.

Approach

Use sliding-window (local) group confidence from model token distributions to weight and filter traces offline, and to terminate traces online once their lowest group confidence falls below a warmup-derived threshold, combined with consensus-based adaptive sampling.

Key results

• Offline, GPT-OSS-120B reaches 99.9% on AIME25 with DeepConf@512 (top-10% tail confidence) vs 97.0% for majority voting at 512 traces.
• Online DeepConf-low reduces generated tokens by up to 84.7% on GPT-OSS-120B AIME25 while reaching 97.9% accuracy versus 97.1% for Cons@512.
• DeepSeek-8B on AIME25 improves from 82.3% (cons@512) to 87.4% with top-10% filtering in offline setting.

Surprising findings

• Aggressive filtering (top 10%) can hurt on some settings due to model overconfidence on incorrect problems, e.g., GPT-OSS-120B.
• Head-based confidence (first tokens) has weak correlation with correctness and typically matches plain majority voting.

Limitations

• Models can be highly confident on incorrect reasoning paths, limiting confidence-based filtering.
• Evaluated only on math and GPQA reasoning benchmarks with open-source reasoning models.
• Online method requires a per-problem warmup of 16 complete traces and tuned consensus threshold tau.

Experimental setup

Datasets: AIME24 AIME25 BRUMO25 HMMT25 GPQA-Diamond

Benchmarks: AIME24 AIME25 BRUMO25 HMMT25 GPQA-Diamond

Metrics: accuracy generated tokens pass@1 cons@512

Baselines: Self-consistency (majority voting) Pass@1 Mean trace confidence (self-certainty) Early-Stopping Self-Consistency Dynasor-CoT

Builds on: self-consistency self-certainty Dynasor Early-Stopping Self-Consistency

Closest papers by meaning

• Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.92)
• Consilience for Verifier-Free Test-Time Scaling (NeurIPS 2026 · similarity 0.89)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.87)
• Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · similarity 0.87)
• A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.87)
• Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · similarity 0.87)