Reasoning with OmniThought: A Large CoT Dataset with Verbosity and Cognitive Difficulty Annotations

TL;DR

OmniThought addresses the lack of large-scale, richly annotated chain-of-thought datasets by introducing a self-contained pipeline that creates over 2 million CoT processes annotated with Reasoning Verbosity (RV) and Cognitive Difficulty (CD). Generated by two teacher LRMs and validated via LLM-based judging, the dataset enables targeted CoT sampling that matches problem difficulty and model capacity, improving LRM training efficacy. Empirical results with Qwen2.5 backbones show RV/CD-guided training yields stronger reasoning abilities and more efficient CoT outputs, with released models demonstrating superior performance on diverse reasoning tasks. The work delivers a scalable framework for aligning CoT generation with model capabilities, offering practical benefits for training LRMs to solve complex problems while controlling reasoning length.

Abstract

The emergence of large reasoning models (LRMs) has transformed Natural Language Processing by excelling in complex tasks such as mathematical problem-solving and code generation. These models leverage chain-of-thought (CoT) processes, enabling them to emulate human-like reasoning strategies. However, the advancement of LRMs is hindered by the lack of comprehensive CoT datasets. Current resources often fail to provide extensive reasoning problems with coherent CoT processes distilled from multiple teacher models and do not account for multifaceted properties describing the internal characteristics of CoTs. To address these challenges, we introduce OmniThought, a large-scale dataset featuring 2 million CoT processes generated and validated by two powerful LRMs as teacher models. Each CoT process in OmniThought is annotated with novel Reasoning Verbosity (RV) and Cognitive Difficulty (CD) scores, which describe the appropriateness of CoT verbosity and cognitive difficulty level for models to comprehend these reasoning processes. We further establish a self-reliant pipeline to curate this dataset. Extensive experiments using Qwen2.5 models of various sizes demonstrate the positive impact of our proposed scores on LRM training effectiveness. Based on the proposed OmniThought dataset, we further train and release a series of high-performing LRMs, specifically equipped with stronger reasoning abilities and optimal CoT output length and difficulty level. Our contributions significantly enhance the development and training of LRMs for solving complex tasks.

Figures (7)

• Figure 1: A motivation example of CoTs with different Reasoning Verbosity (RV) and Cognitive Difficulty (CD) levels. For simplicity, we only present key steps in these CoTs.
• Figure 2: Our framework, including the dataset construction pipeline and the training procedure. Note: We leverage DeepSeek-R1 and QwQ-32B as the teacher models in all modules of our work. Other sufficiently strong models can also be leveraged as teacher models.
• Figure 3: The distribution of CoT processes per problem.
• Figure 4: Comparison using CoTs from varying RV levels as training sets.
• Figure 5: The distributions of CD scores of CoTs in OmniThought.