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

The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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.

Introduces DC-CoT, a data-centric benchmark that systematically compares augmentation, selection, and mixing strategies for chain-of-thought distillation into 3B-8B student LLMs and VLMs across textual, agentic, and visual reasoning tasks.

Problem

Data-centric CoT distillation (augmentation, selection, mixing) lacks a systematic benchmark assessing the effect of each technique on small student models across tasks, teachers, and students.

Approach

A unified benchmark that evaluates reverse-thinking, answer/question augmentation, rephrasing, teacher/student/LLM-judge filtering, and length/teacher mixing under one LoRA distillation pipeline, with IID, OOD, teacher/student-size, and data-volume analyses.

Key results

• Reverse Thinking augmentation on Llama-3.1-8B reaches 66.45% textual average versus 41.81% for Vanilla CoT, the largest gain among tested strategies.
• Filtering with Teacher correctness gives 44.73% textual average vs 42.80% No Selection on Llama-3.1-8B; Teacher Mixing and Length Mixing both slightly underperform No Mixing on textual average.
• Reported that DC-CoT Reverse Thinking reaches 69.2% on ARC-Challenge with Llama-3.1-70B teacher versus 64.8% for logit-based KD (single comparison).

Surprising findings

• Vanilla CoT distillation sometimes underperforms No CoT fine-tuning; increasing seed data for Vanilla CoT can degrade performance, violating the usual scaling trend.
• Teacher Mixing and Length Mixing do not improve average textual performance over No Mixing, and the strongest teacher is not always best for small or specialized students.

Limitations

• Due to budget constraints, GPT-o4 and other teachers/benchmarks were not fully explored; agentic evaluation limited to WebArena.
• Limited hardware restricted extension to larger and more student models.
• Some inconsistencies in reported numbers and claims (e.g., speedup figures, teacher assignments per task) reduce clarity of conclusions.

Experimental setup

Datasets: StrategyQA CommonsenseQA ARC-Challenge GSM8K MATH ANLI Date Understanding WebArena Visual-CoT OK-VQA CLEVR BoolQ

Benchmarks: WebArena Visual-CoT OK-VQA CLEVR StrategyQA CommonsenseQA ARC-Challenge GSM8K MATH ANLI Date Understanding BoolQ

Metrics: accuracy success rate 95% confidence intervals

Baselines: Zero Shot Zero Shot CoT No CoT Vanilla CoT Logit-based KD

Builds on: MetaMath RevThink I-SHEEP Chain-of-Thought prompting LoRA

Closest papers by meaning

• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.91)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.91)
• Scaling Code-Assisted Chain-of-Thoughts and Instructions for Model Reasoning (NeurIPS 2025 · similarity 0.90)
• CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs (ICLR 2026 · similarity 0.90)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)