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

Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 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.

Trains a probe-based confidence predictor on truth-relevant attention-head activations to score each CoT step and guides beam search with it, improving reliability over CoT, self-consistency and PRM baselines on LLM and MLLM reasoning benchmarks.

Problem

CoT reasoning reliability is undermined by error accumulation across intermediate steps, and surface token probabilities are poor indicators of step correctness.

Approach

Probe attention-head activations for truthfulness on a binary CoT step dataset, select the top heads, concatenate their last-token activations, train a sigmoid confidence predictor with an ECE-style soft-target loss, and combine its score with generation probability inside beam search.

Key results

• On LLaMA2-13B-Chat calibration over WikiQA, the predictor reaches ECE 0.037 and AUC 0.934, versus 0.058 ECE and 0.868 AUC for ACTCAB.
• With LLaMA2-70B, the method averages 71.5 across five unimodal reasoning datasets, above SC (70.6) and PRM (69.9).
• On multimodal benchmarks with Qwen2.5-VL-7B, the method averages 65.6 versus 65.5 for PRM and 65.4 for SC, a small margin.

Surprising findings

• On SciQ, Sequence Likelihood achieves higher AUC (0.949) than the proposed predictor (0.937).
• Random selection among candidates in the ablation sometimes performs below the few-shot CoT baseline.

Limitations

• Gains over strongest baselines are often small or mixed (e.g., Qwen2.5-VL-7B multimodal average 65.6 vs 65.5 for PRM; some benchmarks where baselines win).
• Predictor is trained on a non-CoT dataset for calibration evaluation, and a fixed lambda=0.5 was chosen by grid search on several datasets.

Experimental setup

Datasets: GSM8K SVAMP StrategyQA BoolQ Boolean_Expressions CLEVR-Math ScienceQA RealWorldQA MMStar CommonsenseQA SciQ WikiQA

Benchmarks: GSM8K SVAMP StrategyQA BoolQ Boolean_Expressions CLEVR-Math ScienceQA RealWorldQA MMStar CommonsenseQA SciQ WikiQA

Metrics: accuracy ECE Brier score AUC

Baselines: Few-Shot CoT Self-Consistency Self-Evaluation Guided Beam Search Process Reward Model Sequence Likelihood Is True probability Verbalization ACTCAB Revise prompting

Builds on: Truthfulness probing of attention heads (Li et al. style probes) ACTCAB Self-Evaluation Guided Beam Search Self-Consistency Process Reward Models

Closest papers by meaning

• FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.92)
• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.91)
• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.91)
• How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning (ICML 2026 · similarity 0.91)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.91)