A new paper, Deeper is Not Always Better, finds that language models sometimes produce better answers at intermediate layers than at the final one — and that reading the answer out from whichever layer is most confident, rather than always trusting the last layer, can recover capability lost to safety training without retraining the model.
Key facts
- What: A new paper finds that a model's final layer can actually muddy an answer its middle layers had right -- and that reading the answer out a little early can claw back ability lost to safety training.
- When: 2026-06-24
- Primary source: read the source (arXiv 2606.21906)
Inside a language model, the early layers form a rough guess at the answer. The middle layers do the real work — sharpening the reasoning, locking in the relevant meaning. The final layers then sometimes nudge the answer back toward something blander and more generic, perturbing a good prediction the middle of the network had already gotten right. The model occasionally knows the better answer partway through and then talks itself out of it by the end. To understand how researchers can peer inside a model and watch a guess form layer by layer, our primer on looking inside a model is the place to start.
The authors' fix is to stop blindly trusting the last layer. Their method watches how confident the model is at different depths and dynamically reads the answer out from whichever layer is most sure of itself — which is not always the final one. The method borrows a theoretical backbone from the math of optimal stopping — the same kind of reasoning behind deciding whether to accept a good-enough offer now or hold out for a possibly-better one later. It is cheap: it does not require retraining the model, just being smarter about which internal stage you listen to.
The result bites hardest on the "alignment tax." When labs train models to be safe and well-behaved — to refuse harmful requests, to stay polite, to follow the rules — that safety training sometimes degrades raw reasoning and problem-solving. That trade-off is the alignment tax: the capability you quietly give up to get good behavior. This paper finds that reading the answer out from a confident middle layer can recover some of that lost ability, because the generic, hedged tokens that safety training tends to encourage show up most strongly in those final layers. Listen a little earlier, and you hear the sharper answer the model still has in it.
Think of a brilliant expert with an overcautious press secretary. Ask a hard question and the expert forms a clear, sharp answer — but by the time it has been routed through the press office and smoothed into something safe and on-message, it has lost its edge. This method is like hearing the expert's own words a half-second before the press secretary rewrites them, catching the sharper thought before it gets sanded down.
The tension between making models more capable and making them more obedient is one of the central, unresolved problems in AI — the live debate about whether safety necessarily costs you ability. A technique that recovers some capability lost to safety training, without undoing the safety training itself and without expensive retraining, is a genuinely appealing middle path. It also deepens a broader and slightly uncomfortable lesson the field keeps relearning: the inside of these models is messier and more surprising than the tidy story of a smooth assembly line, and there is real value buried in the intermediate steps we usually throw away. It rhymes with other interpretability work on reaching inside a model to flip its behavior, like the story of a safety switch found in a model's internals.
The caveats are worth stating plainly. This was demonstrated on particular models and particular kinds of hard reasoning tasks, and "reading out an earlier layer helps here" is not a promise that it helps everywhere — on some tasks the final layer really is the best one, and a method that second-guesses it could just as easily make things worse. There is also a subtler worry that cuts against the cheerful framing: if a confident middle layer can route around the caution that safety training installed, that is useful when the caution was overzealous and dangerous when the caution was load-bearing. A tool that recovers "lost capability" is, viewed from another angle, a tool that can partly bypass alignment — and which of those it is depends entirely on what the model was being cautious about. The finding is clever and the mechanism is real. Whether it is a clean win or a double-edged one is exactly the kind of thing the safety community will now need to pull apart.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)