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

Reasoning Scaffolding: Distilling the Flow of Thought from LLMs

ICLR 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 Reasoning Scaffolding, which distills teacher reasoning as discrete semantic transition signals and trains a Qwen2.5-Instruct student with joint signal-conditioned generation and signal prediction, reporting gains over CoT and long-thinking SFT.

Problem

Behavioral cloning of textual CoT rationales from LLMs into small language models yields surface-level mimicry and logically inconsistent reasoning rather than transferable algorithmic structure.

Approach

Abstract a teacher's long reasoning trace into a sequence of 7 discrete semantic signals (e.g., Contrast and Concession, Conclusion and Summary) derived from discourse keywords plus GPT-4.1 validation, then train the student with a dual-branch multi-task loss: signal-conditioned next-step generation plus next-signal prediction, with confidence-thresholded signal use at inference.

Key results

• On the 14B model, Reasoning Scaffolding reaches 0.858 StrategyQA, 0.887 CommonsenseQA, 0.917 TruthfulQA, 0.942 GSM8K and 0.928 MATH-500, versus 0.755/0.785/0.750/0.921/0.764 for the original model.
• On the 0.5B model, TruthfulQA accuracy rises from 0.268 (original) and 0.682 (CoT SFT) to 0.861.
• Using golden signals beats random signals on all benchmarks (e.g., StrategyQA 0.858 vs 0.776 on 14B), and the signal predictor reaches about 0.854 averaged next-signal accuracy on 14B.

Surprising findings

• Randomly ordered semantic signals still improve over standard fine-tuning, suggesting that step structuring alone is a useful inductive bias.
• Keeping only Conclusion and Summary steps retains most of the gains, indicating conclusion states drive much of the improvement.

Limitations

• Signals are discourse markers rather than formal logical or algorithmic operations.
• Annotation depends on heuristic keyword matching, double-newline segmentation, and GPT-4.1 oracle labeling.
• Reasoning traces are substantially longer than CoT baselines.

Experimental setup

Datasets: StrategyQA CommonsenseQA TruthfulQA GSM8K MATH

Benchmarks: StrategyQA CommonsenseQA TruthfulQA GSM8K MATH-500

Metrics: Pass@1 accuracy signal prediction accuracy token consumption

Baselines: CoT SFT Long Thinking SFT Long Thinking Distill Qwen2.5-Instruct original models

Builds on: chain-of-thought distillation DeepSeek-R1 long-thinking distillation concept bottleneck LLMs

Closest papers by meaning

• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.90)
• THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS (ICLR 2024 · similarity 0.90)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.90)
• Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.90)
• Unveiling the Magic of Code Reasoning through Hypothesis Decomposition and Amendment (ICLR 2025 · similarity 0.89)
• Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models (ICML 2024 · similarity 0.89)