CMU Paper: Most Wrong Actions by System One Models Follow Correct Reading of State

CarnegieMellonU · hf · 2026-10-14

CMU's "Reading Right, Acting Wrong" studies how System One models (scoring action options in a single pass) fail in agent loops: did the model misread the state, or read it right and still act wrongly?

The authors pair every next-action question on a frozen benchmark (simulated environments, customer-service policies, tool use, web navigation, coding) with independent factual probes about the same state.

Key findings across 17 fast systems and one reasoning model:

Most wrong actions follow correct reading; the more accurate the system, the higher this error share;
Errors cluster on the same states across systems of every kind;
Where facts provably determine the action, models that read every fact still fail to combine them — composing their own answers via code repairs the failure;
On real states, what's missing is often not a fact but the next sub-goal.

Original post →

CMU Paper: Most Wrong Actions by System One Models Follow Correct Reading of State

More from coding & agent