[Paper Review] Full-Stack Domain Enhancement for Combustion LLMs: Construction and Optimization

Quanjia Xiao, Weimin Ouyang+5 more|arXiv (Cornell University)|Feb 27, 2026

Machine Learning in Materials Science0 citations

TL;DR

The paper presents a full-stack workflow that constructs a combustion-specific corpus, applies multi-stage model adaptation (CPT, SFT, RLVR), and introduces FlameBench to achieve state-of-the-art combustion reasoning beyond general LLMs and RAG baselines.

ABSTRACT

Large language models (LLMs) in the direction of task adaptation and capability enhancement for professional fields demonstrate significant application potential. Nevertheless, for complex physical systems such as combustion science, general-purpose LLMs often generate severe hallucinations due to insufficient domain knowledge and the inability to adhere to physical conservation laws. To address this issue, we propose the first full-stack domain-enhanced LLM workflow tailored for the field of combustion science, which integrates automated domain corpus construction, incremental pre-training, instruction fine-tuning, and verifiable reward-based reinforcement learning. This workflow ensures that the model truly internalizes physical laws rather than merely learning textual statistical patterns. We also release FlameBench, a standardized evaluation benchmark specifically designed for complex reasoning tasks in combustion science. Experimental results demonstrate that the model developed in this work significantly outperforms state-of-the-art general-purpose closed-source models and traditional retrieval-augmented generation methods on combustion science reasoning tasks. This work lays a solid technical and resource foundation for the subsequent development of domain-specific scientific research agents with reliable scientific reasoning capabilities.

Motivation & Objective

• Motivate the need for domain-specific LLMs in combustion science due to domain knowledge gaps and physics constraints.
• Propose a full-stack pipeline integrating corpus construction, incremental pre-training, supervised fine-tuning, and verifiable reinforcement learning to enforce physical consistency.
• Introduce FlameBench as a standardized benchmark for evaluating combustion-domain reasoning.
• Demonstrate that the proposed workflow yields superior domain reasoning and compares favorably to RAG and general LLM baselines.

Proposed method

• Construct a large-scale combustion-specific corpus (~5B domain tokens within ~30B total tokens) from English/Chinese publications and physics/chemistry resources.
• Perform continue pre-training (CPT) on the mixed corpus to inject domain knowledge while preserving general language ability.
• Apply two-phase supervised fine-tuning (SFT-General then SFT-Combustion) to align instructions and domain-specific reasoning patterns.
• Use reinforcement learning with verifiable rewards (RLVR) under a KL constraint to improve physically consistent multi-parameter reasoning.
• Develop FlameBench with 436 domain-specific questions to evaluate domain knowledge and constrained reasoning.
• Compare end-to-end training (CPT–SFT–RLVR) against baselines including CPT, SFT, RAG baselines, and general LLMs.

Figure 1: Dataset token distribution by category.

Experimental results

Research questions

• RQ1 Can a full-stack domain-enhancement workflow improve combustion-specific knowledge retention in LLMs beyond general-domain pre-training?
• RQ2 Does multi-stage adaptation (CPT, SFT, RLVR) enforce physical consistency and improve multi-physics reasoning in combustion tasks?
• RQ3 How does an end-to-end domain-adapted model perform relative to RAG-based approaches and other closed-source models on a combustion-specific benchmark?

Key findings

• CPT yields a substantial improvement over the baseline, increasing FlameBench accuracy from 26.8% to 33.3%.
• SFT-General provides marginal gains over CPT, while SFT-Combustion raises accuracy to 35.1%.
• RLVR optimization substantially improves accuracy to 43.8%, with stabilized output length and higher mean rewards under low-entropy policy.
• RLVR-Opt outperforms the best RAG baseline (RAG + GLM-4) by 11.71 percentage points on FlameBench.
• SFT-Combustion and RLVR enable competitive performance against closed-source models and approach GLM-4 in some subfields.
• Compared to RAG methods, RLVR-Opt achieves higher accuracy and removes retrieval overhead, indicating stronger internalized domain knowledge and reasoning.

Figure 2: Data processing pipeline for the combustion-specific pre-training corpus.