Source-linked AI summary
CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning
Xinyu Zhu, Yihao Feng, Yanchao Sun, Xianzhi Du, Pingzhi Li, Olli Saarikivi, Yun Zhu, Yu Meng
TL;DR
Open reasoning-model development is limited by scarce long-CoT seed data, narrow domain coverage, and costly annotation. CHIMERA addresses these gaps with a compact, structured synthetic dataset and automated validation, then uses it to post-train a 4B Qwen3 model. The resulting model performs competitively across challenging reasoning benchmarks against substantially larger models.
Problem
Open reasoning post-training lacks detailed long-CoT seed data, broad scientific coverage, and scalable alternatives to costly human annotation.
Method
CHIMERA combines 9K long-CoT synthetic samples, a hierarchical taxonomy spanning 8 disciplines and over 1K topics, and automated model-based validation.
Results
Post-training a 4B Qwen3 model on CHIMERA yields strong performance across diverse reasoning benchmarks and reaches parity with substantially larger models.
Takeaways & Limitations
Structured, high-quality synthetic data rather than scale alone can support effective reasoning capabilities in LLMs.
Takeaways & Limitations
Detailed thinking processes from proprietary frontier models are often inaccessible, so they cannot directly synthesize the detailed solutions used for training.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have recently exhibited remarkable reasoning capabilities, largely enabled by supervised fine-tuning (SFT)- and reinforcement learning (RL)-based post-training on high-quality reasoning data. However, reproducing and extending these capabilities in open and scalable settings is hindered by three fundamental data-centric challenges: (1) the cold-start problem, arising from the lack of seed datasets with detailed, long Chain-of-Thought (CoT) trajectories needed to initialize reasoning policies; (2) limited domain coverage, as most existing open-source reasoning datasets are concentrated in mathematics, with limited coverage of broader scientific disciplines; and (3) the annotation bottleneck, where the difficulty of frontier-level reasoning tasks makes reliable human annotation prohibitively expensive or infeasible. To address these challenges, we introduce CHIMERA, a compact synthetic reasoning dataset comprising 9K samples for generalizable cross-domain reasoning. CHIMERA is constructed with three key properties: (1) it provides rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models; (2) it has broad and structured coverage, spanning 8 major scientific disciplines and over 1K fine-grained topics organized via a model-generated hierarchical taxonomy; and (3) it employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness. We use CHIMERA to post-train a 4B Qwen3 model. Despite the dataset's modest size, the resulting model achieves strong performance on a suite of challenging reasoning benchmarks, including GPQA-Diamond, AIME 24/25/26, HMMT 25, and Humanity's Last Exam, approaching or matching the reasoning performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.
1. Introduction
CHIMERA addresses data-centric barriers to open reasoning-model development with a compact synthetic dataset emphasizing long reasoning traces, broad coverage, and automated quality control. Post-training a 4B Qwen3 model on CHIMERA yields competitive performance across challenging reasoning benchmarks.
- Motivation: Three data-centric challenges constrain scalable reasoning post-training: scarce long CoT seed data, narrow domain coverage, and costly or infeasible annotation.These challenges particularly affect smaller and mid-sized models that rely more heavily on supervision quality and structure.
- Design principles: CHIMERA prioritizes long intermediate reasoning, structured domain diversity, and scalable quality control rather than maximizing dataset size.These principles target narrow coverage, error accumulation, and uncontrolled quality risks in naive synthetic generation.
- CHIMERA: 9K synthetic samples provide long CoT trajectories, coverage of 8 scientific disciplines and over 1K topics, and automated cross-validation.Topics follow a model-generated hierarchical taxonomy, while strong reasoning models validate problem validity and answer correctness.
- Results: A 4B Qwen3 model post-trained with SFT and RL on CHIMERA performs strongly across GPQA-Diamond, AIME24/25/26, HMMT25, and Humanity’s Last Exam.The model reaches performance competitive with substantially larger models including DeepSeek-R1 and Qwen3-235B.
- Contributions: The contributions formalize data-centric challenges, introduce CHIMERA, and show that a 4B model can approach larger models across difficult reasoning benchmarks.The reported comparison spans a range of challenging reasoning tasks rather than a single benchmark.
2. Method
CHIMERA uses a modular LLM-driven pipeline to expand broad subjects into organized topics, generate verifiable problems, synthesize detailed reasoning, and filter examples. The resulting dataset emphasizes structured cross-domain coverage and long-form supervision, including 9,225 problems across 8 subjects and 1,179 topics.
- Overview: A modular pipeline expands coarse seed subjects into fine-grained topic hierarchies before generating problems and solutions.Its stages are decoupled and intermediate artifacts are saved for later filtering and curation.
- Subject coverage: Mathematics accounts for 48.3% of CHIMERA, while the remaining distribution includes computer science, chemistry, physics, and other disciplines.The mathematics subset is further decomposed into fine-grained subfields.
- Problem generation: Generated problems must be solvable, self-contained, unambiguous, and easy to verify, with LLM-based filtering enforcing these criteria.The generation process targets expert-level problems that are not open research questions.
- Quality control: Two independent verifiers retain a problem only when both judge its validity and answer correctness positively.Using independent models is intended to reduce systematic bias and hallucinated solutions from a single evaluator.
- Solution synthesis: A strong open reasoning model regenerates detailed trajectories and labels each trajectory according to whether its final answer matches the original answer.Correct trajectories can be used for supervised fine-tuning.
- Dataset statistics: 9,225 problems span 8 subjects and 1,179 topics, with solutions reaching 11K words and substantially longer prompts than prior datasets.CHIMERA emphasizes explicit subject structure and long-form reasoning over sheer dataset scale.
3. Experiments
Experiments compare Qwen3-4B models trained with CHIMERA against the base model and public synthetic baselines across diverse reasoning benchmarks. CHIMERA fine-tuning produces broad gains, including under inference-time scaling, while SFT supplies most of the improvement before RL adds further gains.
- Experimental Setup: Models are evaluated on scientific reasoning, mathematical problem solving, and knowledge-intensive benchmarks, comparing the base model with OpenScience- and CHIMERA-fine-tuned settings.All experiments use Qwen3-4B-Thinking-2507 as the base model, with HLE restricted to text-only problems.
- Main Results: CHIMERA fine-tuning yields gains of +4.3 on GPQA-Diamond, +5.3 on AIME24, +6.5 on HMMT Feb 25, +9.7 on HMMT Nov 25, and +1.7 on HLE.The 4B model matches or surpasses 8B–70B baselines on almost all benchmarks and approaches much larger models.
- Main Results: OpenScience fine-tuning performs worse than the base model across benchmarks, plausibly because its multiple-choice format encourages shortcut elimination over explicit multi-step reasoning.The authors contrast this format with free-form reasoning tasks.
- Inference-Time Scaling Performance: CHIMERA improves pass@k over the base model across all tested sampling budgets on GPQA-Diamond and HLE.On GPQA-Diamond, pass@8 reaches 90.7% versus 81.5%; on HLE, pass@1 rises from 7.3% to 9.0% and pass@8 from 19.5% to 24.0%.
- SFT-Only Performance on CHIMERA: SFT alone provides most performance gains across benchmarks, while RL contributes additional improvements.SFT produces gains including +3.0 on GPQA-Diamond, +4.9 on AIME24, +3.9 on HMMT Feb 25, +9.0 on HMMT Nov 25, and +1.7 on HLE.
4. Analysis
CHIMERA is evaluated for dataset difficulty, synthetic-problem quality, and contamination risk. It is substantially harder than prior synthetic datasets, receives higher quality rankings under the reported protocol, and shows near-zero benchmark overlap.
- 4.1. Data Difficulty Analysis: CHIMERA remains substantially harder than prior synthetic datasets for Qwen3-4B-Thinking-2507.The base model achieves near-saturation performance on prior datasets but only 37.5% accuracy on CHIMERA.
- 4.1. Data Difficulty Analysis: 88% accuracy on DAPO-Math-17K, DeepMath-103K, and OpenScience contrasts with 76% on OpenR1-Math-220K and 37.5% on CHIMERA.
- 4.2. Data Quality Analysis: LLM-generated problems receive higher average quality scores than human-curated HLE problems under both independent evaluators.The ranking protocol scores clarity, well-posedness, and reasoning depth.
- 4.2. Data Quality Analysis: Under o4-mini, HLE, gpt-5, and gemini-3-pro problems score 1.14, 2.67, and 2.20, respectively.Under gemini-2.5-pro, the corresponding scores are 1.30, 2.53, and 2.18.
- 4.3. Data Contamination Analysis: 8-gram and 13-gram overlap with GPQA-Diamond is exactly 0, while HLE shows 7.89 × 10^-6 for 8-grams and 0 for 13-grams.Lower overlap indicates weaker lexical overlap and lower contamination risk.
- 4.4. Dataset Format: Each CHIMERA sample contains hierarchical labels, a multi-step problem, a verifiable answer, a detailed solution, and a correctness label.
5. Related Work
Related work spans reasoning benchmarks and scalable synthetic-data pipelines. CHIMERA addresses the narrower gap of training datasets designed for difficult, long-horizon reasoning while providing evidence that generated problems can match or exceed human-curated quality under blind evaluation.
- 5.1. Reasoning Benchmarks: Existing benchmarks evaluate mathematical, scientific, graduate-level, frontier, and contamination-aware reasoning capabilities.
- 5.1. Reasoning Benchmarks: Many existing training datasets exhibit near-saturation performance for modern LLMs, motivating compact datasets tailored to complex, long-horizon post-training.
- 5.2. LLMs for Data Generation: Synthetic-data methods use LLMs to bootstrap instruction–response pairs and scalable reasoning supervision with reduced reliance on human annotation.
- 5.2. LLMs for Data Generation: Blind evaluations rate LLM-generated problems on par with human-curated benchmarks in clarity and reasoning depth, while the reported experiments show downstream reasoning improvement.
6. Conclusion
The paper identifies cold-start supervision, limited domain coverage, and costly annotation as barriers to scalable reasoning post-training. CHIMERA mitigates these barriers with structured synthetic data and enables strong cross-benchmark performance from a 4B model.
- 6. Conclusion: Cold-start supervision, limited domain coverage, and costly human annotation are identified as core barriers to scalable reasoning post-training.
- 6. Conclusion: CHIMERA combines long Chain-of-Thought trajectories, broad scientific coverage, and fully automated quality control.
- 6. Conclusion: A 4B model post-trained on CHIMERA reaches parity with substantially larger models such as DeepSeek-R1 and Qwen3-235B across diverse reasoning benchmarks.The conclusion attributes this result to the dataset’s compact, structured, and high-quality synthetic-data design.
A. Prompts
The appendix documents the prompts used for the data synthesis pipeline.
- A. Prompts: Table 6 lists the prompts used for the data synthesis pipeline.
B. Subjects and Topics
CHIMERA organizes its subject coverage through expanded topic sets, spanning a wide range of scientific and mathematical subfields. The listed topics include foundational areas, specialized theories, and interdisciplinary applications.
- CHIMERA uses expanded topic sets to organize its subjects and topics.
- It also covers advanced mathematical fields such as algebraic geometry, geometric Langlands, model theory, operator algebras, and symplectic dynamics.
- The topic inventory spans foundational areas including number theory, graph theory, geometry, topology, probability, statistics, and analysis.
- The topics extend beyond pure mathematics to information theory, machine learning, networks, epidemiology, climate risk, and environmental modeling.
C. Examples of CHIMERA
The examples illustrate CHIMERA’s broad collection of reasoning problems across mathematics, computer science, chemistry, and other technical domains. Samples pair formal tasks with detailed solution processes, numerical calculations, or proof-oriented reasoning.
- CHIMERA examples include mathematical, computer-science, and chemistry problems organized by subject.
- Mathematical examples include stochastic-process analysis, sieve theory, recurrence solving, asymptotics, and probability calculations.
- Detailed solutions are presented as step-by-step reasoning, including substitutions, derivations, asymptotic arguments, and verification of final results.
- Other examples formulate optimization problems with constraints, including proving uniqueness, deriving an optimizer, and reporting numerical quantities.
- The examples also include proof-oriented computer-science tasks, such as establishing a bijection between projective dependency trees and head-annotated phrase-structure trees.