Does a reasoning model have a subconscious? A plain lens says probably not
A reasoning model writes out a chain of thought before it answers. That gives us something rare: a written record of what the model went on to say. So we read the model's silent internal state partway through its thinking, to check whether it already held the words the model was about to write. On Qwen3-8B it did, about 88% of the time, against a 15% baseline you would expect by chance. But a plain, ordinary readout does almost as well as the elaborate "workspace" lens built to find it, which suggests this foresight is a property of the model's regular machinery, not a separate hidden workspace it keeps out of sight.
Why you'd care
People want to trust a model's chain of thought (the step-by-step reasoning it writes before answering) as a window into how it reached an answer. But the written thoughts are just tokens (the word-chunks a model reads and writes) that the model emitted. Whether they reflect anything happening inside the model is a separate question, and a hard one to test. You rarely have a ground truth for what the model was about to think.
A reasoning model hands you that ground truth for free. It writes its reasoning into an explicit trace, so the words it is about to write already exist in the trace. That lets us check a specific claim: does the model's internal state, read silently before those words are written, already point at them? If yes, the written reasoning is at least connected to something the model computed ahead of time, not made up after the fact. There is a bolder version of this idea worth naming up front: that the model keeps a silent inner workspace, almost a subconscious, that an ordinary readout cannot see. We can test that too.
Reading a model's silent state
Inside a language model, each token is represented by a long list of numbers called the residual stream (the running internal state the model's processing stages read from and write to). Most tools only look at this state through the model's final output layer, which tells you the single next token the model would emit. That readout is the logit lens. It is the simple baseline here.
The Jacobian lens is a more elaborate readout. It also accounts for how the current internal state would nudge the model's later output, not only the very next token. (The Jacobian is the math that measures that nudge: an average map of how a change to this state would propagate to the model's final output.) It reads that through the output layer, so it exposes the tokens the model is silently disposed to say later. Some researchers call the space it reveals an emergent workspace, a silent scratchpad where the model holds a thought before writing it down. That framing has an almost psychological ring: the model has a private inner life the plain readout cannot reach. This post puts that claim to the test. We did not train anything. We loaded a Jacobian lens that Neuronpedia already released for Qwen3-8B and used it as is.
Our probe bank is 200 short reasoning questions in four kinds: factual recall, multi-hop (questions that need two or more linked facts), analogy, and arithmetic. They are held out from the data the lens was built on (kept separate, so the test is fair). For each, we let the model think, then read the silent state at many points inside that thinking and checked whether its top guess shows up in the tokens the model writes next. The trace-alignment numbers below come from 48 of these questions, read at hundreds of internal positions. We also pick the best readout layer on a held-out half of the traces and report on the other half, so the numbers are not inflated by choosing a lucky layer.
The silent state holds what the model is about to write
Read mid-thought, the Jacobian lens's top word lands somewhere in the model's actual near-future thinking about 88% of the time. A control that shuffles in a different question's writing, so common words still show up as often, hits only 15%. So the silent state is not just echoing common words. It genuinely holds what the model is about to think.
That is exactly what the "workspace" idea is supposed to mean. If you can read the model's next words out of its silent state before it writes them, then the model is holding the thought ahead of time, not composing it at the last instant. Being able to predict the model's own future output from its present internal state is the evidence for a silent workspace. So far the workspace story holds up. The real question is whether you need anything special to see it.
The horizon is how far ahead we look: does the readout's top word appear in the next few tokens, or do we give it a wider window? Slide it in the diagram below and watch the raw recall (does the top word appear anywhere in the next stretch of tokens). It is already high even at a short horizon. It stays far above the chance floor at every window. The diagram also draws the plain logit lens for comparison, and the two track each other closely, which is a point we return to later.
Raw recall (top word appears anywhere in the next H tokens)
Next token excluded (must point further ahead)
The dashed mark on the raw-recall bars is the ~0.15 chance floor (measured for raw recall only; the next-token-excluded metric has no measured floor). Raw recall stays high and far above the floor at every horizon, and the two lenses track each other closely. The next-token-excluded recall climbs from near zero to about 0.41 as the window widens, which is where the readout proves it reaches beyond the immediate next token. Point estimates from a single run, read at seven horizons on held-out traces (reported split n=275).
It reaches past the very next token
Raw recall has a loophole. If the readout is just predicting the immediate next token well, it would score highly without really "looking ahead." So we ran a stricter version: remove the immediate next token from the target, and only count a hit if the readout's word turns up further down the trace.
That stricter number (the next-token-excluded bars in the diagram) starts near zero at a short horizon and climbs to about 41% as the window widens to 48 tokens. The readout points at words the model writes several steps later, not only the one right after. So the anticipation extends several tokens into the trace.
It holds for every kind of reasoning
The anticipation is not a trick of one easy question type. We split the recall by reasoning kind, all read at the same horizon.
Future-recall of the silent workspace, split by reasoning type (read at a 24-token horizon). All four sit far above the 0.15 chance floor. The anticipation is broad, not carried by one easy question type.
All four kinds land far above the 0.15 chance floor, from about 0.83 for arithmetic to 0.93 for multi-hop. Each kind is only about a dozen traces, so read the small spread between them as noise, not a ranking. The point that holds is that the anticipation works across all four reasoning kinds, not just one.
Conclusion
We expected the Jacobian lens, with its extra sensitivity map, to reach meaningfully further ahead than the plain logit lens. It does not, by much. On raw recall the two are basically tied (0.87 versus 0.86). On the stricter next-token-excluded measure the Jacobian lens is ahead from about 8 tokens onward, but by only about 2.5 points at its best.
The whole selling point of the Jacobian lens is that it reads future dispositions the ordinary readout supposedly cannot. But inside a coherent chain of thought, the ordinary readout reads them about as well. The foresight is not hiding in a special space you can only reach through the elaborate transport. It is sitting in the plain, next-token machinery the model uses all the time.
That splits into two conclusions. The anticipation is real, and it is not an artifact of one clever lens: two very different readouts recover the same near-future words, and when two independent windows agree, the thing they see is more likely real. But there is no sign here of a separate hidden workspace, a subconscious the plain readout misses. The model's ordinary state already leans several tokens ahead, which is a mildly surprising thing to learn about a plain readout, not a private inner life. We did not run causal steering or a thinking-versus-not ablation, so we cannot rule a hidden workspace out for good. What we can say is that this test gives no reason to posit one, and a much cheaper explanation fits: the regular machinery is already looking ahead.
Takeaways
- A thinking model is a testbed for interpretability claims. Its written trace is a ground truth for "what came next," which a base model never gives you. Use it to validate silent readouts.
- The silent state anticipates the reasoning. Read mid-trace, the internal state's top word lands in the model's realized near-future thinking about 88% of the time (versus 15% chance), and it reaches beyond the immediate next token.
- No evidence for a separate subconscious. A plain logit lens reads the future about as well as the elaborate Jacobian lens (within about 2.5 points). The foresight lives in the model's ordinary machinery, not a hidden workspace the plain readout cannot see. Report the effect as a property of the model's state, not of the lens.
- This is prediction, not control. We did not steer, and we did not run the thinking-versus-not ablation. The claim is that the silent state anticipates the reasoning, hedged to what a correlational test can show.