Where do LLMs succeed at generation but struggle with evaluation?

This explores the gap between what LLMs can produce and what they can reliably assess — and the corpus suggests this gap isn't a quirk but a structural feature. The cleanest statement of it is the generation-verification gap: a model can write a fix, an answer, or an argument far more easily than it can confirm that fix is correct What stops large language models from improving themselves?. This is why self-improvement hits a hard ceiling — every reliable correction needs something *external* to validate it, because the model's own judgment is anchored to the same distribution that produced the flawed output in the first place.

The most vivid version is what happens when LLMs are asked to evaluate other text. LLM judges prefer LLM-written arguments 62% of the time versus humans' 39%, even controlling for quality — they reward the smooth, distribution-typical prose they themselves generate, which quietly corrupts any pipeline that uses AI to grade AI Do LLM judges systematically favor LLM-generated arguments?. Generation and evaluation aren't symmetric skills here; the same fluency that makes generation strong makes evaluation *biased toward fluency*.

Why the asymmetry? Several notes point to the same fault line: explanation and execution run on disconnected pathways. Models score 87% explaining a principle but 64% applying it — a 'computational split-brain' where knowing-that and doing-correctly come apart Can language models understand without actually executing correctly?. Potemkin understanding sharpens this: a model can give a correct explanation, fail to apply it, *and* recognize its own failure — a triple pattern no human cognition shows Can LLMs understand concepts they cannot apply?. Generation taps the articulate pathway; evaluation requires the execution pathway to check the work, and the two don't talk.

There's a deeper reason generation comes easy. Token prediction trains models to flow toward the training distribution, not to explore the counterpositions that genuine evaluation demands — the process is smooth, so it produces smooth claims that multiply without ever stress-testing themselves Does LLM generation explore competing claims while producing text?. Evaluation is friction; generation is flow. This connects to why reasoning models 'wander' rather than systematically search, and why they need external summarization and explicit prompting to even track what they've already tried Why do reasoning LLMs fail at deeper problem solving?, Why do LLMs struggle with exploration in simple decision tasks?.

The payoff the corpus offers — and the thing you might not have expected — is that the fix is almost always *externalizing* the evaluator. Walmart's small BERT cross-encoders outperformed their LLM teachers once trained on enough teacher-labeled data: a narrow, external verifier beat the fluent generalist at the judging task Can smaller models outperform their LLM teachers with enough data?. The recurring lesson across these notes is that you don't close the generation-evaluation gap by asking the model to think harder about itself. You close it by putting the verifier *outside* the model that did the generating How do LLMs fail to know what they seem to understand?.