🚨New work: A SOTA prompt optimization framework + a novel unsupervised measure of prompt performance. Written rubrics and constitutions are crucially important for AI improvement, safety following, pluralism and adaptation. But some of the things we might want a rubric to do are difficult to put into words; how would you even describe “comic voice”?

Let me break this down. Step 1: optimizing the message. How do you know a prompt is as good as it can get? We introduce an algorithm, ε-estimation, that treats prompt optimization as a structured search over a search space of prompt variations. We improve via a novel submodular (roughly) optimization algorithm and can estimate how much missing search is left, using a concept from ecology, capture-recapture (CR). This approach ends up being remarkably effective: ε-estimation matches or beats SOTA optimizers (GEPA, MIPROv2) by up to 14% at equal budget on 6 benchmarks, and its mid-run forecasts are within ~5% of where the search actually ends up.

🚨New work: A SOTA prompt optimization framework + a novel unsupervised measure of prompt performance. Written rubrics and constitutions are crucially important for AI task improvement, safety following, pluralism and adaptation. But some of the things we might want a rubric to do are difficult to put into words; how would you even describe “comic voice”? I’m human, albeit not that funny, and I'm totally not sure what that means. What kinds of written constructs can you expect your judge to do well on and how good can you expect it to be?

This has important implications for rubric design and learning strategies moving forward: the concepts models can't be told are exactly the ones we most want them to have for frontier tasks like: taste, pluralistic alignment and discovery. Thank you @Diyi_Yang , @sanmikoyejo , @noahdgoodman and Dan Ho! Please check out the full paper!! Articulation Upper Bounds paper link: arxiv.org/pdf/2610.19721

So excited for this research! I've always been annoyed at the supremacy of verified rewards, and now we can finally show it. In this work, we develop ways of measuring: how much of a reward signal is verifiable, how much can we put into words, and how much is shared but tacit, or impossible to articulate?

Tacit Knowledge: Skills, Apprenticeship and Meaning Paul Meehl recounts how, in 1970s, psychologists were wondering: why did rats at UC Berkeley exhibit latent learning while rats at University of Iowa did not, and rats at the University of Minnesota only sometimes did? All groups of psychologists seemed to be running, on paper, the same experiments governed by the same procedures and undergirded by the same theories. Were Berkeley’s experiments irreplicable because of bad science? (If you don’t know what latent learning is, it’s a fascinating learning mechanism we don’t think of much in AI, it is kind of similar to pre-training but without any explicit self-supervised signal: it involves open-ended exploration without reward signals of, say, a maze; it seems to speed up maze-learning and is not measurable until a reward signal is introduced.) It turns out, the Berkeley experimenters were playing with the mice throughout the week and being nice to the mice. The mice were thus less resistant and more playful, and more likely to explore the maze. On some level, this is an example of Collin’s Relational Tacit Knowledge: knowledge that is, in theory, articulable it just hasn’t been articulated yet. In other skills and domains, it truly is inarticulable: what does it actually mean to play with the mouse? What is the right way to play and the wrong way? There is a level below which we cannot articulate this skill; someone must be apprenticed into it. Only after someone is apprenticed, can I name the skill: “play with the mouse” and they will know what to do.