R-GSM

anchor

Findings

• IC-391Llama-3 and Qwen-1.5 models exhibit position bias in LM-as-a-judge, retrieval-augmented QA, and math reasoning, with larger models showing less bias [eval]