External Feedback Outperforms Self-Refinement for LLM Improvement

In multi-turn language agent settings, improvements in final accuracy are often observed, but it's frequently unclear whether these gains are due to genuinely useful feedback or simply other factors like resampling, format correction, or additional computation during testing. This research introduces a controlled student-teacher protocol to isolate and study when natural-language feedback truly drives improvement beyond what repeated attempts alone can achieve. The study evaluated thirteen open-weight models across various tasks, examining external feedback, self-feedback, and unguided self-refinement. The findings reveal that multi-turn improvement is often not strong evidence of effective feedback use; self-generated feedback, for instance, offered minimal gains beyond unguided self-refinement. Instead, the most substantial feedback-specific improvements were observed when strong external teachers provided guidance that went beyond generic retry instructions. The research also highlights that the student model's inherent ability to effectively utilize feedback is a more critical driver of interactive gains than the specific identity of the teacher, though teacher quality remains important. These results underscore the need to evaluate feedback-based agents against repeated-attempt baselines and suggest that enhancing an agent's capacity to act on feedback is a central bottleneck for interactive improvement.