A common way to build a specialized model is to start with a general-purpose foundation model and adapt it to a domain or expert community. We usually describe this adaptation through its mechanics, such as how the model's parameters are updated, or what feedback it learns from. That tells us how the model changes, but not how specialization constrains the model.
While reading recent work on fine-tuning, I realized that I didn't have a good answer to that basic question. What exactly changes? When we adapt a model to an expert community, are we primarily teaching it the community's consensus? What happens to the disagreement and uncertainty within it? If the model moves closer to the judgment of that community, what does it move away from? And in science, where discoveries sometimes come from ‘fortunate accidents' that most experts would have avoided, could fine-tuning make those paths harder to find?
What does fine-tuning learn from experts?
Thinking Machines and Bridgewater recently used fine-tuning to reproduce expert judgment in finance. The foundation model could already read the documents; what it lacked was an investor's sense of what mattered. Fine-tuning rests on the idea that much of the needed capability is already there, and that specialization can be learned without rebuilding the foundation.
Their use of the word judgment is revealing. To keep things simple, I will narrow judgment to the question: given the available evidence, how much weight should we assign to each possible state of the world? In machine-learning terms, that judgment becomes visible in the conditional distribution the model assigns over those possibilities. This framing does not explain where expert judgment comes from. It only gives us one way to see its effect. Faced with the same evidence, an expert and a non-expert may reach very different views of the world.
Fine-tuning is valuable because it can change that distribution for a particular domain while leaving much of the foundation intact. Many teams across the industry are already working to reproduce expert behavior in specialized domains. The Thinking Machines work is one recent example in which fine-tuning produced a more efficient model that outperformed the frontier models they tested on their financial tasks. Fine-tuning turned difficult-to-articulate expert decisions into repeatable model behavior.
But can expert decisions alone teach the judgment behind them? That is not straightforward to assume.
Are we learning judgment or just the decision?
Fine-tuning on expert answers carries an implicit assumption that, by learning the answers, the model will also learn the judgment that produced them. Expert supervision may take the form of an answer, a ranking, a preference, or a reward. In each case, the model is pushed toward some responses and away from others. That may be enough to reproduce expert choices without reproducing the judgment behind them.
Consider the sequence: 4, 5, 14, 15, …
One natural continuation is 24. The increments alternate between one and nine. Another is 40, because 4, 5, 14, 15, and 40 are the first positive integers whose English names begin with F.
Both explanations fit the observations. For a model specialized in linguistic patterns, 40 may seem like the answer favored by expert consensus. But that consensus does not tell us how individual experts viewed 24. Some may have considered it wrong, others unlikely, and others simply less relevant in that context. The label 40 does not reveal which of those judgments led to the choice.
An expert label tells us which answer was chosen, but not how strongly it was preferred or which alternatives remained plausible. Even when the training process samples many rollouts and receives dense feedback along them, that feedback does not by itself reveal how the expert would judge the paths not taken. The challenge is to learn the expert's choice without erasing uncertainty that still matters.
Can specialization suppress useful uncertainty?
The authors of Hindsight Merging found that fine-tuning can suppress beneficial variability inherited from pretraining, and that restoring some of it improved performance. That variability is not necessarily the same thing as uncertainty. But it shows that fine-tuning can remove useful alternatives along with noise.
This gives the concern a concrete form. A model may reproduce the answers experts give without learning how their judgment changes. That difference is easy to miss while the environment resembles the fine-tuning data. It appears when new evidence supports an explanation the model has learned to ignore: can it recover that explanation, or has specialization made the dominant one feel inevitable?
When expert judgment misses the exception
Science has an awkward relationship with bad judgment. Most unlikely ideas are wrong, and expertise is valuable because it filters them. Yet some discoveries begin with an observation the prevailing expert view has learned to dismiss.
When Robin Warren and Barry Marshall began studying Helicobacter pylori, stress and lifestyle were still considered the major causes of peptic ulcers. Warren gave the bacteria he saw in stomach biopsies more weight than the prevailing medical view did. Their work changed the accepted explanation of ulcer disease and earned them the 2005 Nobel Prize in Physiology or Medicine.
The point is not to favor unlikely ideas. A judgment can look mistaken under a field's current beliefs and still motivate the experiment that changes those beliefs. Scientific judgment depends on expert filters, but also on their ability to revise.
A model fine-tuned on current scientific judgments may become better at normal scientific work. It may recognize established patterns and avoid familiar dead ends. But if it learns only the field's final answers, it may inherit the consensus without inheriting the process by which that consensus can change. Some discoveries begin with what current experts would regard as bad judgment. Training those deviations away could produce an excellent practitioner of present science, but one less prepared for scientific change.
What should survive fine-tuning?
So, what are we losing in fine-tuning? Not necessarily knowledge. The foundation model may still be able to produce an alternative when it is explicitly requested. What may be lost is the model's tendency to consider that alternative on its own.
This brings us back to the questions at the beginning. When we fine-tune toward an expert community, we may learn where that community tends to agree without preserving the disagreement within it. As the model moves closer to the expert consensus, it may move away from explanations that experts consider less likely, less relevant, or simply outside the usual frame.
Some of that narrowing is exactly what specialization is for. But not every suppressed alternative is noise. Some may become important when the evidence changes. In science, those are often the paths that lead away from the accepted explanation.
The loss, then, may be one of recoverability. An idea that once arose naturally may now appear only after the user names it. The model may still contain the knowledge, but the path from new evidence to that possibility has become harder to follow.
This is not an argument against fine-tuning. It is an argument for asking what kind of judgment we want to preserve. Matching expert answers is not enough. A model may agree with experts today and still fail to revise for the reasons an expert would tomorrow. Expert judgment is not only where belief settles. It is also how belief moves when the world stops behaving as expected.