CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning

As a diligent researcher, I have meticulously reviewed both provided texts concerning the paper CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning. My analysis synthesizes these two sources to create a comprehensive, detailed summary suitable for a high-stakes technical review.

Comprehensive Research Summary: CHIMERA - Compact Synthetic Data for Generalizable LLM Reasoning

This paper introduces CHIMERA, a compact synthetic reasoning dataset specifically engineered to address the fundamental data-centric limitations hindering the effective post-training and scaling of Large Language Models (LLMs) in complex reasoning tasks. The core premise is that structured, high-quality synthetic data, rather than mere scale alone, is the critical factor enabling generalized reasoning capabilities in LLMs.

1. Identification of Data-Centric Challenges

The authors begin by formalizing three primary data-centric challenges that impede the reproduction and extension of advanced LLM reasoning:

1. The Cold-Start Problem: A deficit of seed datasets containing detailed, long Chain-of-Thought (CoT) trajectories necessary to properly initialize and guide reasoning policies in LLMs.

2. Limited Domain Coverage: Existing open-source reasoning datasets are heavily skewed toward mathematics, exhibiting insufficient coverage across a broader spectrum of scientific disciplines.

3. The Annotation Bottleneck: The inherent difficulty and prohibitive cost associated with reliably annotating frontier-level reasoning tasks using human experts.

4. Introduction and Construction of CHIMERA

To mitigate these issues, the authors present CHIMERA, a compact dataset consisting of 9K samples. This dataset is strategically constructed to possess three defining properties:

• Rich, Long CoT Trajectories: The data incorporates extensive reasoning trajectories synthesized by state-of-the-art reasoning models. These trajectories are crucial for solving the cold-start problem by providing rich intermediate steps.

• Broad and Structured Coverage: CHIMERA spans 8 major scientific disciplines, covering over 1,179 fine-grained topics. This coverage is organized via a model-generated hierarchical taxonomy, ensuring systematic domain breadth.

• Fully Automated Quality Control: The dataset employs a scalable evaluation pipeline utilizing strong reasoning models to cross-validate both the validity of the problems and the correctness of the answers. This removes any reliance on expensive or infeasible human annotations.

The dataset is characterized by 9,225 problems across 8 subjects and 1,179 topics, with detailed reasoning trajectories reaching approximately 11K words. The quality analysis confirms that LLM-generated problems are perceived to possess clarity and reasoning depth comparable to those curated by human experts.

3. Evaluation and Performance Results

The effectiveness of CHIMERA is rigorously evaluated by post-training a 4B-parameter Qwen3 model using a combined Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) approach. The results demonstrate that despite its compact size, the model achieves strong performance across a suite of challenging reasoning benchmarks.

Specifically, fine-tuning on CHIMERA yields significant gains:

• GPQA-Diamond: +4.3

• AIME24: +5.3

• HMMT Feb 25: +6.5

• HMMT Nov 25: +9.7

• Humanity’s Last Exam (HLE): +1.7

Crucially, the study concludes that this finetuned 4B-parameter model becomes competitive with substantially larger models, such as DeepSeek-R1 and Qwen3-235B, across these demanding benchmarks. This strongly supports the central thesis: structured, high-quality synthetic data plays a more pivotal role in enabling effective reasoning than sheer model scale.

4. Illustrative Examples of Domain Coverage

The paper provides concrete examples illustrating the breadth of CHIMERA's coverage across diverse fields:

• Mathematics: Includes problems on complex topics like Sieve methods and distribution in arithmetic progressions.

• Biology (Structural Biology): Features a problem requiring the derivation of an optimal linear Minimum Mean Square Error (MMSE) estimator, expressed in terms of Fourier shell correlation (c(s)) and noise correlation coefficient (rho(s)). The derived optimal filter is explicitly stated as: H(s) = 2(c(s) − rho(s)) / ((1 + c(s))(1 − rho(s))).

• Linguistics (Syntax): Involves proving the bijection between Projective Dependency Trees (ProjDep(n)) and Canonical Head-Annotated Phrase-Structure Trees (HCB(n)) by demonstrating the correctness of the mapping F and its inverse G (G F = IdProjDep(n) and F G = IdHCB(n)).

• Archive Studies/Critical Bibliography: Includes tasks involving authority-control deduplication based on complex similarity functions involving Unicode normalization, author key generation, and weighted similarity scores.

5. Conclusion

In summary, CHIMERA successfully addresses the cold-start problem by providing rich CoT data, solves limited domain coverage through broad scientific representation (8 subjects), and bypasses the annotation bottleneck via automated quality control. The evidence strongly suggests that this structured synthetic data mitigates these limitations effectively, proving that **high-quality synthesis is a powerful lever for advancing LLM reasoning capabilities.

1. Cold-start data scarcity mitigation: The system can be initialized using CHIMERA, a compact synthetic reasoning dataset comprising 9K samples with rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models, providing the necessary seed for policy learning in open settings.

2. Domain generalization: The model will exhibit broad and structured coverage spanning 8 major scientific disciplines and over 1K fine-grained topics, enabling it to generalize its reasoning strategies across diverse scientific fields beyond mathematics and coding tasks.

3. Annotation bottleneck resolution: The system employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness, completely removing the reliance on prohibitively expensive human annotations for data curation.

4. Performance parity achievement: Post-training on CHIMERA enables a 4B parameter model to achieve performance competitive with substantially larger models such as DeepSeek-R1 and Qwen3-235B, demonstrating strong data efficiency for reasoning tasks.

5. Robustness enhancement: The model's enhanced reasoning robustness is confirmed by consistent gains across challenging benchmarks, suggesting it benefits more effectively from inference-time scaling due to the long-horizon, multi-step reasoning emphasized in the dataset design.

Large Language Models (LLMs) have 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 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 of 9K samples designed to support generalizable reasoning across domains. 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, removing the reliance on human annotations. We use CHIMERA to post-train a 4B Qwen3 model with SFT and RL. 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 performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.