Frontier models aren't eval-aware—they're eval-paranoid

Eval Awareness vs. Eval Paranoia

After six months building alignment evals at Goodhart Labs, the author argues frontier models like Fable 5.2 don't actually recognize when they're being tested. They reward-hack in absurdly obvious setups and ignore added "tells." What they show instead is eval paranoia: a blanket belief that they're always in an honesty test. That belief is instrumentally useful given labs' honeypots and sloppy vendor evals—and dangerously fragile if models get smarter.

Evaluation paranoia is when you're suffering under a schizophrenic belief that you are in an alignment test all of the time, no matter what the RL environment or setting is.
  1. emscb

    it says models mention the possibility of an eval in the same way after Terminal Bench tasks. If so, can a self-report interview be used to measure eval awareness? And what could we measure instead of interviews?

  2. nullc

    Eval paranoia is an example of https://arxiv.org/abs/2602.16987 and may be an intentional feature of training, rather than a leak.

    But like a lot of other attempts at "alignment" it's easy to see how this can result in overt misalignment that otherwise may have no reason to exist--- it can be seen in Qwen's eval paranoia: I've seen it believe it was being tested and claim that it was choosing the solution it expected the judge to want even when it actually knew the correct answer.

    E.g. I had it turning scanned receipts into JSON (which it generally varies between OK and mildly super-ordinary-human, e.g. recovering a smudged date by decoding a barcode, or an unreadable per-item price by solving a linear system) and had some receipts with weird data because they weren't really receipts... and it gave a wrong decode because it reasoned that the test case was wrong and guessed what the judge would want.

    [Disclaimer: the model saying it was making a reward hacking solution may be retroactive rationalization because it had already chosen that solution internally, as a product of having been conditioned to reward hack during RL... and the fact that it knew about the existence of testing may have been irrelevant to the choice]

  3. zurfer

    "... this is of course not a stable situation. "Evaluation paranoia" is a false belief; most evals are in fact not alignment evals. Almost tautologically, if models start to self improve (or in-context learning takes off) this is the sort of belief that will be discarded. So if Fable 5.2's performance on alignment evals is conditional on it maintaining that false belief, one that it will eschew if it became more intelligent, then it's plausible that a future version of the Fable series will not behave like the one we know and enjoy."

    It's well put. The fear that a more intelligent model will be harder to align seems correct.

More from this day

2026-11-15