UniCo Framework Generates Data to Train Universal Causal Reasoners in LLMs
Reactions from ranked influencers
2 posts
New project on data generation for improving LLM causal reasoning!
Introducing UniCo馃, a data generation framework for training Universal Causal Reasoners. We show causality-centered training develops a causal mindset towards faithful and consistent reasoning in the real world. With continued SFT on 66.6K UniCo data, Qwen3 and Olmo-3-Instruct show substantial gains on 18 causal query types (+22.9%), 7 established causal tasks (+8.1% over SOTA data generation frameworks), and faithfulness in real-world general reasoning (+20.2%). [1/n]
We began this project with a sense of curiosity: different causal reasoning benchmarks evaluate very different capabilities. Can we train a causal reasoner that performs well across all of them? The answer is yes, and it even improves reasoning faithfulness in general domains!
Introducing UniCo馃, a data generation framework for training Universal Causal Reasoners. We show causality-centered training develops a causal mindset towards faithful and consistent reasoning in the real world. With continued SFT on 66.6K UniCo data, Qwen3 and Olmo-3-Instruct show substantial gains on 18 causal query types (+22.9%), 7 established causal tasks (+8.1% over SOTA data generation frameworks), and faithfulness in real-world general reasoning (+20.2%). [1/n]