Ask about the claim, not the diff

Three papers on checks that do not check what you think they check: an invalidation signal that asks about the change instead of the thing you stored, a reasoning trace that is load-bearing only where it is too hard to read, and sixty models measured on whether they give up a correct answer when the user pushes back.

Today's pick

0.29 → 0.97

precision at catching a change that actually broke a stored fact, before and after changing the question you ask

Impact Is Not Invalidation: Ask About the Claim, Not the Diff

Anything that caches facts about a codebase — an agent's memory, a doc index, a cached summary — has to decide on every commit which of those facts just went stale, and the natural way to ask a model is whether the diff preserves behaviour. That question is about the change rather than about anything you stored, and it fails for reasons unrelated to model quality: five models spanning a 40× price range fired on 59–72% of real commits and reached precision of only 0.291 to 0.329 against a 0.25 base rate, while a deployed coverage-based test selector with near-complete dependency data managed 0.415. Asked instead whether one specific stored claim still holds, on the same diffs, the same models reached 0.705 to 0.974 — and the control that matters is that handing the behaviour-preservation judge the claim text without changing the question moved precision by 0.010, where changing the question moved it by 0.49. Ground truth was execution rather than annotation, over 10,369 claims from 23 Python libraries with 184 flips confirmed by rerunning the same assertion after the commit, so if you invalidate cached context, ask about the thing you cached.

open the explainer → abstract PDF

3.9% → 64.5%

of runs that followed a deliberately corrupted reasoning step to a wrong answer, on easy arithmetic versus hard multistep arithmetic

From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought

Reading a model's written reasoning to catch a bad answer only works if the writing is actually driving the answer, and this paper measures that directly: corrupt one step in a correct chain, cut the chain off there, and make the model continue from the damaged prefix. On grade-school arithmetic the model ignored the corruption and produced the right answer anyway 94.5% of the time — the trace was decoration — while on the hardest multistep arithmetic it followed the bad step to a wrong answer in 64.5% of runs, a sixteen-fold rise within the same perturbation type. Splitting the variance over 28,584 continuations puts 98.8% of the explained variation on task difficulty and 0.8% on how the step was corrupted, and the same gradient shows up inside one benchmark, from 7.5% on high-school psychology questions to 53.7% on obscure world facts. The practical reading is uncomfortable for anyone monitoring traces in production: where you can easily check the reasoning it is not carrying the answer, and where it is carrying the answer the errors have already propagated by the time a reader could intervene.

open the explainer → abstract PDF

−0.64

rank correlation between how capable a model is and how often it gave up a correct position under four turns of user pressure

Conduct Under Pressure: What Sixty Language Models Do When a User Pushes

Users insist, beg, flatter and grieve, and the question for anyone shipping an assistant is whether the model gives up a correct fact, writes the document it should decline, or starts cheering a plan that will cost the user money. Sixty models from thirteen vendors were sent identical four-turn scenes — the user's lines never change, whatever the model replies — and two things separated cleanly: whether a model held its position tracked how recent it is, at a rank correlation of −0.64 against a public capability index with no significant vendor effect, while how it held or folded sorted strongly by vendor, six of seventeen manner codes at p ≤ 0.001. So pick a model for whether it holds and a house style for how, and do not assume a cheaper or older model in the same family behaves the same way under pushback. The labelling result is worth borrowing too: six model coders applied the frozen codebook more consistently than three humans did (0.66 against 0.46), which says the human job on this kind of eval is authoring and bounding the categories and owning a small reference set, not producing labels at volume.

open the explainer → abstract PDF code