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

Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 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.

Proposes DynaSteer, a training-free inference-time representation editing framework that steers LLM reasoning trajectories at high-entropy sentence forks using pattern clustering, Fisher-LDA truth directions, and rollback, improving MATH and coding accuracy over ITI, ACT, and DRESS.

Problem

Existing prompting methods like CoT and "Wait" make LLMs think longer without steering them toward correct conclusions, and representation editing has rarely been applied to dynamic, unfolding reasoning trajectories where the final answer is not yet visible.

Approach

Localize truth-steering interventions to sentence-level high-entropy reasoning forks and early steps, disentangle reasoning patterns by clustering difference vectors, derive purified truth directions with Fisher-LDA instead of mean-difference, and roll back committed erroneous sentences.

Key results

• On Qwen3-1.7B, DynaSteer reaches 74.80 on MATH500 and 47.50 on AMC23 versus 69.20 and 40.00 for plain generation and 72.80 and 42.50 for the best RepE baseline (DRESS).
• On Llama3.2-3B-Instruct, DynaSteer reaches 52.40 on MATH500 vs 48.80 for the best baseline (ACT), a 7.38% relative gain.
• On out-of-domain coding, DynaSteer scores 67.68 on HumanEval+ and 54.50 on MBPP+ with Qwen3-1.7B, above all RepE baselines.

Surprising findings

• Removing entropy monitoring or temporal decay does not significantly hurt accuracy, though each reduces inference token consumption (about 56% and 52%).
• Fine-tuned GRPO reaches 84.6% vs DynaSteer's 82.3% on a comparison setting, reported in the limitations section.

Limitations

• Assumes truth is linearly encoded; authors hypothesize non-linear truth manifolds explain the gap to GRPO.
• Gains in ablation are small and several components do not show clear accuracy benefits.
• Evaluated only on three small-to-mid Qwen3/Llama backbones (1.7B to 8B).

Experimental setup

Datasets: GSM8K MATH HotpotQA 2WikiQA

Benchmarks: GSM8K MATH500 AMC23 HumanEval+ MBPP+ HotpotQA 2WikiQA

Metrics: pass@1 accuracy

Baselines: Plain Wait prompt ITI ACT DRESS TruthX TruthFlow ROME MEMIT

Builds on: Inference-Time Intervention (ITI) Adaptive Activation Steering (ACT) DRESS Linear Representation Hypothesis Fisher Linear Discriminant Analysis

Closest papers by meaning

• Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization (AAAI 2026 · similarity 0.88)
• DynaAct: Large Language Model Reasoning with Dynamic Action Spaces (NeurIPS 2025 · similarity 0.88)
• DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching (ICML 2026 · similarity 0.88)
• Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.88)
• DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search (ICLR 2025 · similarity 0.88)
• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.87)