IC-384Llama-3-8B-Instruct shows larger absolute gains from prompt optimization than the stronger Gemma-2-9B-IT, indicating that prompt-optimization benefit is inversely related to base model capability

Sarkar Snigdha Sarathi Das, Ryo Kamoi, Bo Pang, Yusen Zhang, Caiming Xiong, Rui Zhang

SourceGReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers

The paper measures the improvement from zero-shot chain-of-thought to GREATER-optimized prompts on GSM8K, BBH, and FOLIO for two released models. Llama-3-8B-Instruct gains +3.0 on GSM8K (79.6 to 82.6), +6.5 on BBH (62.2 to 68.7), and +4.0 on FOLIO (58.6 to 62.6). Gemma-2-9B-IT, the stronger model, gains only +0.8 on GSM8K (88.6 to 89.4), +4.9 on BBH (71.7 to 76.6), and +4.1 on FOLIO (65.0 to 69.1). The authors explicitly state that 'stronger models get lesser benefit from prompt optimization, as good prompts only offset for lower capabilities of weaker models.' This pattern is consistent across all three benchmarks and both models.

Evidence
correlational
Key metric
Llama-3-8B-Instruct: GSM8K 79.6→82.6, BBH 62.2→68.7, FOLIO 58.6→62.6; Gemma-2-9B-IT: GSM8K 88.6→89.4, BBH 71.7→76.6, FOLIO 65.0→69.1 (zero-shot CoT to GREATER)
Caveat
The observation is made across only two models from different families at similar parameter counts; the paper does not test multiple sizes within a single family to isolate the size effect from the architecture effect.
Model
Llama 3 8B Instruct, Gemma 2 Gemma-2-9B-IT
Datasets
GSM8K [eval], FOLIO [eval]
Methods
Zero-shot Chain-of-Thought / Wei et al. 2022 (Chain of Thought) / Zero-shot chain-of-thought prompting [compared-to]
Related work
APE [compared-to], APO [compared-to], PE2 [compared-to], TextGrad [compared-to]
Extraction
automatic-extraction