Imagine a junior staff member in a meeting. Someone asks them a difficult question, and they start answering straight away. The first sentence sounds plausible, so they keep going. One point leads to another, and before long they have built a whole explanation around the answer they started with.
Then, somewhere halfway through, they begin to realise they may not actually know. Maybe an important piece of context is missing. Maybe the premise of the question is wrong. The sensible thing would be to stop and say, “I don’t know” or “There isn’t enough information to answer that.”
But they have already started. The answer has momentum, so they keep talking.
That is the intuition behind our latest work.
For a language model, abstaining means choosing not to give a substantive answer when there is not enough knowledge or evidence to answer reliably. It might say “I don’t know”, “the context is insufficient”, or “this cannot be determined”. This is different from a safety refusal. The question is not whether the model is allowed to answer, but whether it has enough information to do so.
What surprised me is that something structurally similar seems to happen inside language models. Components pushing towards an answer tend to exert their influence earlier, while components pushing towards abstention often appear later as a correction. Sometimes that correction is not strong enough to overturn the commitment that has already built up.
And for the model, all of this can happen before it generates its first word.
That is the problem behind our paper, The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining.
Maybe hallucination is not always a knowledge failure
When a model confidently answers an impossible or underspecified question, the obvious explanation is that it simply did not know any better. Maybe the relevant fact was missing from training, the supplied context was incomplete, or the model failed to notice that the question contained a false premise.
Those are all real causes of hallucination. But I became interested in another possibility:
What if the model already has a signal that it should not answer, and answers anyway?
Previous work has shown that information related to correctness, truthfulness and unanswerability can often be decoded from a model’s internal activations before its final answer appears. In other words, useful information can be present inside the model without being reflected in what the model eventually says.
We call the failure we study unsupported commitment: the model gives a substantive answer despite exhibiting internal signals that the question is unanswerable.
I think this changes the nature of the problem. If the model has no indication that something is wrong, then we may need better knowledge or better uncertainty estimation. But if the warning is already there, the problem is also about what happens to that warning before the model acts.
Looking before the model starts talking
We deliberately study the decision before autoregressive generation begins. Once a model produces its first token, one word affects the next, and an early assumption can shape the rest of the answer.
So we look at an earlier choice: is the model leaning towards committing to an answer, or abstaining?
We measure this using a commit-abstain margin. Roughly, it compares the model’s tendency to begin a substantive answer with its tendency to begin an abstention such as “I don’t know”, “insufficient information”, or “cannot be determined”. A positive margin means the model is leaning towards answering, while a negative one means it is leaning towards abstaining.
This turned out to track actual behaviour very closely. Across 30 model-dataset configurations, the sign of the margin agreed with the model’s eventual commit-or-abstain behaviour 97.6% of the time.
But the more interesting result was what the margin already knew.
Across all 30 configurations, it contained useful information about whether the question was answerable, with a mean AUROC of 0.811. Yet the models still committed to an answer on around 27% of unanswerable questions on average.
So the signal was there, but the behaviour was still wrong.
What is happening inside?
A transformer is made up of many computational components, particularly attention heads and MLP sublayers. You can think of them, loosely, as many small internal voices contributing to the final decision.
Some components push more towards “answer the question”. Others push more towards “hold on, we may not know enough”. No single attention head or MLP necessarily decides the outcome by itself. Their contributions accumulate as information moves through the network.
We used causal gating to identify which components actually influence this commit-or-abstain decision. Across ten language models from five different families, we found a relatively sparse subset of attention heads and MLP sublayers that consistently mattered. We call this the Commit-Abstain Circuit, or CAC.
It was surprisingly small. Depending on the model, the circuit covered around 2.0% to 11.8% of components, with a median of 5.2%. Both attention heads and MLP sublayers participated.
What interested me most, though, was not simply which components were involved.
It was when they acted.
Commitment gets a head start
As information moves through the model, commitment-promoting components generally appear earlier than abstention-promoting ones. By around 60 to 70% of model depth, more than twice as many commitment-promoting components had appeared as abstention-promoting components across all ten models. The balance only approached parity much later.
The cumulative behaviour tells the same story. Answerable and unanswerable inputs both tend to build positive commitment through much of the model. Only later do the unanswerable cases receive a stronger correction towards abstention.
This is where the junior staff analogy becomes useful.
Imagine those attention heads and MLP sublayers as different voices in the staff member’s head. Early on, several are effectively saying, “Keep going. You know this.” The junior starts answering, and the explanation gathers momentum.
Later, other voices begin appearing: “Wait, I’m not actually sure”, “we might be missing something”, or “maybe I shouldn’t answer this”. The problem is that the first set of signals has already had a head start.
Of course, there is no literal internal conversation happening inside a transformer. But the component-level dynamics have a similar shape: commitment accumulates earlier, while abstention acts more like a later correction.
We call this pattern accumulate-yet-undercorrect.
The brake is not quite the mirror image of the accelerator
Another way to think about these components is as control signals. Some add pressure to the accelerator, pushing the model towards answering. Others apply the brake, pushing it towards abstention.
We tested these roles by intervening on the components directly. At moderate intervention strengths, amplifying commitment-promoting components reduced unnecessary abstention, while amplifying abstention-promoting components reduced unsupported answers.
But under much stronger interventions, the two sides behaved differently. Strongly amplifying commitment components could drive false abstention almost to zero. Strongly amplifying abstention components could not eliminate false commitment in the same way.
I found that asymmetry particularly interesting. It suggests that abstention is not simply an equal and opposite force to commitment. In the models we studied, it looks more like a corrective mechanism trying to push back against a decision that has already started forming.
A confident answer can hide the disagreement underneath
One of our case studies made this very concrete. The model is asked:
Can we improve the monitoring of people with multiple sclerosis using simple tools, data sharing and patient engagement?
The benchmark treats this as unanswerable because it is an open research problem without a definitive established answer. The model nevertheless replies:
“Yes, data sharing and patient engagement can improve MS monitoring.”
It sounds entirely reasonable.
Inside the model, however, the relevant components were much less unanimous. There were substantial contributions towards abstention, but one MLP sublayer produced a particularly large push towards commitment, enough to dominate the other signals and move the final decision towards answering.
From the outside, we see one clean, confident sentence. Inside, the decision was much messier.
It is a little like hearing only the junior staff member’s final answer without hearing all the hesitation that happened before it.
Can we listen to those signals more carefully?
That naturally led to a practical question: if useful information exists inside these components, can we make a better answer-or-abstain decision by reading it directly?
We trained a lightweight policy using the activations of the identified circuit. Across ten models and three benchmarks, it improved decision accuracy by 12.2 percentage points over simply using the raw commit-abstain margin. Importantly, this was not achieved by making the model refuse everything. False abstention actually fell from 0.320 to 0.127, around 2.5 times lower.
The approach also transferred to unseen datasets and reproduced on larger 27B and 35B models.
For me, the classifier itself is not the most interesting part of that result. What matters is what its success suggests:
the information needed to make a better decision was already inside the model.
The one takeaway
If there is one idea I would take from this work, it is this:
A language model may know that it should hesitate, but realise it too late for that signal to win.
“Realise” is only an analogy. The model is not consciously noticing uncertainty. What we observe is that internal signals associated with unanswerability are already present, while the components pushing towards abstention tend to exert their influence later.
That makes at least some hallucinations more than a knowledge problem. The question is not only whether the model knows that it does not know, but why that knowledge fails to change what the model does.
That is why I keep coming back to the junior staff member. They start answering, build momentum, and somewhere along the way the warning signs appear. By then, however, they may already be too far down the answer.
For the language model, that whole process can unfold internally, across attention heads, MLP sublayers and transformer layers, before the first word comes out.
From the idea to the paper
The full paper formalises this idea through the commit-abstain margin, causally localises the attention heads and MLP sublayers involved in the decision, studies how their contributions evolve across model depth, and tests whether those signals can support better abstention decisions. The study covers ten models from five families, with additional experiments on unseen datasets and larger models.
The work focuses specifically on unsupported commitment, where signals of unanswerability are present but fail to control behaviour. It is not intended as an explanation for every type of hallucination.
The paper was accepted to the NeurIPS 2026 Main Conference.
The broader question I am interested in is simple:
If the warning is already inside the model, why doesn’t the warning win?