You Can Do It: Mitigating RL Rollout Distress with Psychological Guidance
Jack Thompson, Anastasia Wei
Large language models (LLMs) often exhibit "functional distress"—manifested as increased activation in frustration and desperation vectors—during reinforcement learning (RL) rollouts on difficult tasks. Drawing on human psychological research, we investigate whether interventions such as growth mindset, resilience, and self-compassion prompts can mitigate this distress. Evaluating Gemma 3 12B, Qwen 3 14B, and Ministral 3 14B Reasoning, we find that while psychological interventions have minimal impact on functional distress—with verbalized distress reduction often driven by evaluation scaffolding rather than prompt content—inoculation prompting reliably increases positive emotional valence during code generation. Our results highlight the challenges of mitigating functional distress in models and suggest that training techniques like inoculation may be more effective than psychological framing for managing emotional states during RL.
The best result is quite underplayed! The test setup not the psychology prompts caused the distress reductions, and the experiment proving that is well done.
Thoughts :
1. Says Figure X in some places without the figure number, no conclusions section.
2. The 2 ways of measuring distress ran on different models and tasks so difficult to compare.
3. How often the models cheated is not mentioned, that seems like the primary downstream outcome of distress.
4. The measured difference are very small, much smaller than prior work you built on.
The idea of testing out human-inspired methodologies to mitigate distress in LLMs is creative, interesting, and seems potentially quite important. This sprint seems like a good start towards a more thorough and robust investigation into the research idea.
Cite this work
@misc {
title={
(HckPrj) You Can Do It: Mitigating RL Rollout Distress with Psychological Guidance
},
author={
Jack Thompson, Anastasia Wei
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
