https://dhyeymavani.com/tags/probing/ProbingHugo Blox Builder (https://hugoblox.com)en-usSun, 04 Oct 2026 00:00:00 +0000https://dhyeymavani.com/media/icon_hu7433143094870913649.pnghttps://dhyeymavani.com/tags/probing/https://dhyeymavani.com/publication/observability-foundation-model/Sun, 04 Oct 2026 00:00:00 +0000https://dhyeymavani.com/publication/observability-foundation-model/<p>Accepted to the <strong>NeurIPS 2026 main conference</strong> (25 pages). Subject classes <code>cs.LG</code>, <code>cs.AI</code>.</p> <p>The short version: a linear probe showing a label is recoverable from a model&rsquo;s hidden states tells you neither that the information goes beyond what the raw input already exposes, nor that the model actually uses it. So we audit against both — the strongest raw-window baseline, and the same architecture stripped of its trained configuration — and then test use directly by swapping residuals from high-burst donors. Recoverability and use come apart.</p>