New study on AI feedback: student ability matters more than teacher quality

A new paper from a team of Polish researchers (Cupiał et al., accepted at RLxF Workshop @ ICML 2026) tackles a fundamental question in AI: when a language agent improves over multiple interactions, is it actually learning from feedback, or just benefiting from resampling, format correction, or extra test-time computation? To separate these effects, the authors introduce a controlled student-teacher protocol evaluated on four challenging benchmarks (Omni-MATH, Codeforces, BBEH Linguini, ARC-AGI1) and thirteen open-weight models used in both student and teacher roles. They compare external feedback, self-feedback, and unguided self-refinement, while varying interaction history, task difficulty, and teacher access to privileged information. The findings are striking: across settings, multi-turn improvement is often not evidence of feedback use. Self-generated feedback performs no better than simply retrying the task without any guidance, while the best external teachers yield substantial feedback-specific gains. Dense student-teacher interaction matrices reveal that gains are driven more by the student's ability to use feedback than by the teacher's identity, although teacher choice still matters for a fixed student.

These results carry clear practical implications for anyone building feedback-based AI agents: evaluations must include repeated-attempt baselines to avoid over-attributing improvement to feedback. The real bottleneck isn't having a great teacher—it's having a student that can effectively act on the guidance provided. The paper also releases its controlled student-teacher evaluation framework, enabling other researchers to replicate and extend the analysis. For the AI community, this work underscores the importance of designing agents that can learn from instruction, not just produce better outputs through trial and error. It challenges the assumption that more sophisticated feedback automatically yields better results, and shifts focus toward improving the mechanism by which agents incorporate feedback into their next attempts.