Putting more metacognitive skills into LLMs is always fun.
I wonder if you can train something like that for other perturbation types—like deployment misconfiguration, quantization or noise injection. Recalling how poor GPT-OSS reception was in part caused by avoidable inference issues, how Claude had a genuine “they made the LLM dumber!” inference bug, and so it goes.
Putting more metacognitive skills into LLMs is always fun.
I wonder if you can train something like that for other perturbation types—like deployment misconfiguration, quantization or noise injection. Recalling how poor GPT-OSS reception was in part caused by avoidable inference issues, how Claude had a genuine “they made the LLM dumber!” inference bug, and so it goes.