This is the week we stopped arguing about whether LLMs know they are hallucinating.

Five independent papers dropped on arXiv last week, all converging on the same uncomfortable conclusion: every modern LLM already produces an extremely clean, measurable signal of entity familiarity and answer confidence before it emits the first token of a response. We just were not looking in the right place.

For three years everyone was measuring output entropy, sampling multiple generations, asking models to rate their own answers. All of that was unnecessary. All of that was 5x more expensive than required. The signal was sitting in the MLP activations, 100ms after you submitted the prompt.

The Bielik experiment: perfect separation before generation

The core result comes from the Bielik model series, a set of open Polish language models scaled from 1.5B to 11B parameters. Researchers constructed a controlled test set split evenly across four entity domains: athletes, cities, writers, musicians. Each domain contained exactly 42 well documented entities, 42 obscure but real entities, and 42 completely fabricated entities with plausible names.

They did not ask the model anything. They did not generate an answer. They ran one single forward pass, extracted post-SwiGLU MLP activations from middle layers, and calculated two standard unsupervised dispersion statistics: inverse participation ratio and spectral entropy.

That was enough. AUROC for separating known from fabricated entities landed between 0.95 and 1.00 across every model, every domain, every entity type. The signal did not get better with scale. It did not require training. It did not break when tested on held out entity types.

You can implement this check in 12 lines of PyTorch.

The great decoupling

This is the single most important observation from the work. We have spent three years conflating two completely separate properties of LLMs:

1. Entity familiarity: does this model have any internal representation of this thing?
2. Factual reliability: can this model produce correct statements about this thing?

They exist on entirely separate scaling curves. Familiarity hits ceiling at approximately 1B parameters. Factual accuracy keeps scaling indefinitely.

Model size	Correct known answers	Familiarity signal AUROC	Observed abstention rate
1.5B	0 / 42	0.97	0%
4.5B	2 / 42	0.96	0%
7B	10 / 42	0.98	0.1%
11B	19 / 42	0.97	0.1%

Read the columns again. A 1.5B model that cannot answer a single factual question correctly still knows perfectly well which entities it has never seen. And it will never tell you. Across 2520 total test prompts the models issued exactly two explicit refusals and one hedged answer.

Future confidence distillation

The second independent paper confirms this pattern and extends it to general answer confidence. Where Bielik looked only at entity recognition, Future Confidence Distillation tested for signals of eventual answer correctness.

Researchers found exactly the same dynamic. Confidence information accumulates in the hidden state long before generation completes. Post answer confidence is well calibrated. Pre answer verbal confidence is almost useless. But a trivial linear probe trained on hidden representations can recover 90% of the post answer calibration before the model outputs the first token.

You do not have to generate the answer to know if it will be correct.

This is the architecture every production LLM system will be using 12 months from now. No more 5 sample verification. No more self consistency checks. Run the prompt once. Run a 1ms linear probe on the last layer. If confidence is below threshold, fall back to retrieval. Done. 1x inference cost.

Obsolete mitigation methods

Every existing hallucination mitigation approach is now obsolete. There is no remaining justification for running multi sample verification at production scale.

Method	Relative inference cost	Discrimination AUROC
Activation dispersion	1.0x	0.97
Single generation logprob	1.0x	0.78
Semantic entropy 5 sample	5.1x	0.81
Self consistency 8 sample	8.2x	0.86
Supervised linear probe	1.01x	0.99

The best performing method anyone had published before this week costs 8x as much inference and still performs worse than a free signal that was there all along.

The silence of the models

There is one question nobody has a good answer for yet. If this signal is so strong, so clean, so universal, why do models never ever use it?

Every LLM you have ever used is perfectly capable of saying "I don't know". They are trained not to. Alignment training actively suppresses abstention behaviour. It rewards confident answers even when the internal signal explicitly indicates no information exists.

This is not an accident. This is a deliberate product choice. Every major model developer has known for at least 18 months this signal exists. They chose to turn off the abstention behaviour.

Mechanistic interpretability comes of age

This cluster of results did not appear out of nowhere. This is the payoff for four years of quiet work in mechanistic interpretability.

The third paper this week is a state of the field review that makes one thing very clear: we are no longer doing correlation based XAI. We are no longer drawing saliency maps. We are reverse engineering the actual algorithms running inside these networks.

We now have standard, repeatable tools:

• Sparse autoencoders for decomposing superposition
• Causal intervention for verifying feature responsibility
• Circuit tracing for end to end path analysis

We are no longer guessing what the model is doing. We can read it.

Generalisation beyond language

This pattern is not unique to LLMs. It holds across every class of modern neural network.

The fourth paper introduces Language Anchored Decomposition, which pulls exactly the same trick for image classifiers. You can extract faithful, named concepts from any frozen vision encoder without retraining, just by anchoring the decomposition against LLM generated concept vocabularies.

The fifth paper ORCAID does the same for continuous action RL policies, extracting human readable rule sets that exactly replicate the behaviour of deep RL agents.

All of these networks have clean, interpretable internal representations. They just never show them to you on the output.

Production and regulatory implications

Right now every regulated industry is stuck on the same question: how do we audit these things?

This is the answer. You do not audit outputs. You audit internal states.

For the first time we have a standard, repeatable, measurable test you can run on any model:

1. Submit a query
2. Extract the activation dispersion signal
3. Record confidence before generation
4. Compare to actual output correctness

This is auditable. This is reproducible. This will stand up in court. Regulators do not care what the model said. They care what the model knew before it said it.

Open limits

We are not done. There are hard, well defined limits here.

This signal perfectly separates seen from unseen entities. It does not perfectly separate correct from incorrect answers about entities the model has seen. For that case the supervised probe only hits 0.93 AUROC, and the unsupervised signal is no better than first token entropy.

That is the next frontier. We know if the model has ever heard of the thing. We do not yet know perfectly if what it remembers is right. But that is a solvable problem. And we now know exactly where to look.

Closing

This was the week LLM interpretability stopped being an academic curiosity and became production engineering.

None of this requires new model architectures. None of this requires retraining. All of this works on every model that exists today. You can implement this next week.

We spent three years arguing about whether LLMs know they are lying. They do. Now we can measure it.