The paper measures how much CRITIC (with code interpreter feedback) improves over Program-of-Thought on TabMWP across three Llama-2 variants. The absolute improvement grows monotonically with model size: +4.7 for 7B, +9.4 for 13B, and +16.0 for 70B. The authors state that 'the benefits derived from critic are more pronounced when paired with larger language models,' indicating that larger models can better leverage external execution feedback to correct their programs.
Evidence
correlational
Key metric
TabMWP improvement over PoT: +4.7 (Llama-2-7B), +9.4 (Llama-2-13B), +16.0 (Llama-2-70B)
Caveat
Measured on a single dataset (TabMWP) with a specific tool (Python interpreter); the scaling trend is reported for one task family (mathematical program synthesis) and one model family (Llama-2).