Post-Training Boosts LLMs' Causal Reasoning, 14B Model Surpasses GPT-4
A recent study published on arXiv (2602.06337) explores the potential of post-training to enhance large language models (LLMs) as causal reasoners. The researchers present CauGym, an extensive dataset featuring seven essential causal tasks for training and five varied test sets. They thoroughly assess five different post-training methods: SFT, DPO, KTO, PPO, and GRPO. Results from experiments across five in-domain tasks and four established benchmarks indicate that with suitable post-training, smaller LLMs can effectively perform causal inference, frequently outperforming larger models. Remarkably, their model with 14 billion parameters achieves an accuracy of 93.5% on the CaLM benchmark, in contrast to OpenAI's o1 model, which scores 55.4%. This research sheds light on the often-overlooked effects of post-training on causal estimation abilities, suggesting its capacity to make advanced AI reasoning more accessible.
Key facts
- Paper on arXiv: 2602.06337
- Introduces CauGym dataset with seven core causal tasks
- Evaluates five post-training methods: SFT, DPO, KTO, PPO, GRPO
- 14B parameter model achieves 93.5% accuracy on CaLM benchmark
- OpenAI o1 achieves 55.4% accuracy on same benchmark
- Post-training enables smaller LLMs to surpass larger models in causal inference
- Study covers five in-domain and four existing benchmarks
- Research addresses limited exploration of post-training effects on causal reasoning
Entities
Institutions
- arXiv
- OpenAI