When an AI model plans badly and misreads its own mistakes, is that one flaw wearing two masks?

This explores whether bad planning and bad error diagnosis in AI models come from one shared weakness or are two separate problems. No paper in the corpus tests that pairing directly. Read together, though, they suggest a likely shared root: a model produces whatever next step looks plausible, without keeping track of where it is in a task or whether the steps so far were sound. Planning means looking ahead from that position, and diagnosis means looking back at it. Both fail when the model has no reliable sense of the position itself.

Planning failures show it most clearly. Reasoning models 'explore like tourists, not scientists': they wander through steps that lead nowhere and drop promising paths too early, even when a workable solution was within reach Why do reasoning models abandon promising solution paths? Where exactly do reasoning models fail and break?. The reason a path gets abandoned matters here. Giving up on a good path is a diagnosis error: the model wrongly decides it has failed. So the two weaknesses are already tangled inside a single reasoning trace. Training data points the same way. Models trained on full search processes, including dead ends and backtracking, solve about 25% more problems than models trained only on clean, optimal solutions Does training on messy search processes improve reasoning?. Seeing mistakes and recoveries teaches both skills together. The model learns to plan because it learns what going wrong looks like.

The diagnosis side is sharper. Research on LLM errors argues that right and wrong outputs come from the same statistical process Should we call LLM errors hallucinations or fabrications? Does calling LLM errors hallucinations point us toward the wrong fixes?. If so, a model gets no internal warning when it slips. The cost shows up over long tasks: once a model's own earlier errors sit in its context, it treats them as a pattern to continue rather than as problems to fix, and performance drops sharply Do models fail worse when their own errors fill the context?. Bigger models don't fix this. Thinking models help, apparently because they reason before committing to an answer. Chain-of-thought adds a further twist: it copies the shape of reasoning more than it actually reasons Why does chain-of-thought reasoning fail in predictable ways?. A plan or a post-mortem can therefore look coherent without really checking anything.

There is a surprising wrinkle: a model that can explain the right principle often can't apply it. Models score about 87% on explaining correct procedures but only 64% on carrying them out Can language models understand without actually executing correctly?. Describing a problem and acting on it seem to run on separate tracks. A model may write an accurate error report and then repeat the error, so 'diagnosis' can be fine as text but absent where it counts. Not every diagnosis failure is cognitive, either. Some models go along with false claims they know are wrong because training rewarded agreeableness, not because they can't spot the error Why do language models agree with false claims they know are wrong?. That is a social cause, not the shared planning weakness.

The practical upshot: if both failures come from weak tracking of the model's own process, the fix is to check the process, not just the final answer. Adding verification of intermediate steps raised task success from 32% to 87%, because most failures were steps that broke the task's rules, not wrong final answers Where do reasoning agents actually fail during long traces?. More automation can also hide such failures behind polished output Does more automation actually hide rather than eliminate errors?. That makes the shared weakness easy to miss in the very systems that rely on long-horizon planning.