Reasoning

Causal Axioms (Axiomatic Training)

Axiomatic Training causal-reasoning evaluation: per-item Yes/No d-separation questions over linear causal chains of length 7-15 (500 items per length), answered by the baseline Meta-Llama-3-8B-Instruct and its axiomatic-training LoRA finetune. Response is binary (1 = prediction matches the gold Yes/No label).