Reasoning for Math and Optimization
Model training and evaluation for mathematical reasoning, automated optimization modeling, code-integrated thinking, and video reasoning.
The reasoning line studies how models solve problems when correctness can be checked: math, optimization modeling, medical reasoning, code-integrated thinking, critique, and multimodal video reasoning.
Research Storyline
MathScale and outcome-supervised verifiers use problems where correctness can be checked, giving reasoning models sharper feedback than preference alone.
ORLM asks models to formulate optimization problems with formal constraints, objectives, and solver-ready structure.
CoRT treats computation as part of reasoning, making intermediate steps more inspectable and less dependent on hidden mental arithmetic.
Video-R1 and HuatuoGPT-o1 show the same verifiable-reasoning idea in multimodal and medical settings.
Research Threads
Scaling instruction tuning and outcome-supervised verifiers for mathematical reasoning and planning.
Large models for automated optimization modeling, where outputs must match formal constraints and executable optimization structure.
Models reason with code inside the thinking process, making intermediate computation more explicit and testable.
Reinforcement-style training for video reasoning in multimodal LLMs, extending verifiable reasoning beyond text-only math.
Display Figures
Paper Trail
Studies scaling and verifier-style feedback for mathematical reasoning.
PaperTurns natural language problem descriptions into formal optimization models and solver-ready structure.
RepositoryUses executable computation inside the reasoning process to improve reliability and inspectability.
RepositoryTransfers reasoning-style training into video understanding and multimodal LLMs.
RepositoryWhy It Matters
- Reasoning models become more trustworthy when they can be trained and evaluated against verifiable constraints.
- Optimization modeling is a bridge from natural language to executable decision-making systems.
- Code and video reasoning show that "thinking" is not a single modality; it can involve computation, perception, and action.
Resource Map
Customizable framework for training large models for automated optimization modeling.
RepositoryCode-integrated reasoning within thinking for models that can use computation as part of the reasoning process.
RepositoryReinforcing video reasoning in multimodal LLMs.
Repository