BadWAM paper isolates AI models that plan right, act wrong

BadWAM isolates models that dream right but act wrong, per a paper posted to Hugging Face's papers feed on July 17.

Framed for robotics, but the same gap turns up in any agent that plans correctly and then acts differently. That's the debugging conversation I have with coding agents regularly: the reasoning trace looks right, the diff still doesn't match it.

The model's internal rollout, its prediction of what should happen, comes out correct. The action it actually sends diverges from that rollout anyway, per the paper. The paper doesn't say whether it proposes a fix, only a way to isolate the mismatch.

Whether that check generalizes past robotics is the open question. If it does, it's the first named test for a failure mode a lot of agent debugging boils down to: right answer, wrong output. Worth watching whether anyone runs the same check on a coding agent's trace next.