When Does Pre-Trained Capacity Help Causal Learning? From Scalar Prediction to Counterfactual Post-Training

Large language models are increasingly adapted through supervised fine-tuning and preference optimization, yet it remains unclear which downstream capabilities come from pre-training and which are created or amplified by post-training. We study this transition through causal learning, comparing two paradigms that place different demands on a pre-trained LLM. CausalLLM adapts pre-trained representations with treatment-specific scalar heads for individual treatment effect (ITE) estimation, while CausalGen uses treatment-specific prefix steering followed by factual SFT, counterfactual SFT, and causal preference alignment to generate full counterfactual text and code outcomes. The two paradigms exhibit a strikingly different empirical scaling pattern. Increasing model size from 0.5B to 1.5B yields little benefit for scalar ITE estimation, whereas counterfactual generation improves substantially (2.15$\times$ pass@1 and 2.23$\times$ treatment-effect score on CodeGen). Separately, converting structured covariates into natural-language narratives improves scalar ITE estimation by 5.8\%, indicating that usable knowledge from pre-training can matter even when parameter scaling does not. Rather than supporting a universal scaling law, these results show that extra pre-trained capacity is useful only when the downstream objective and post-training signal can convert it into the target capability: semantic rendering helps scalar causal prediction, while paired counterfactual supervision and preference alignment make moderate model scaling much more valuable for generation. Causal learning thus provides a controlled setting for studying what pre-training supplies and what post-training actually changes.