Ask a model which continent the biggest land animal lives on and it answers Africa. But it never says the word elephant. The elephant concept exists only in some intermediate state along the way. Now there is a way to draw that state, without training any extra decoder.
How the drawing works
The paper is called V-lens (arXiv:2611.24793, https://arxiv.org/abs/2611.24793, titled The Elephant in the Question, by Tobias Braun, Marcus Rohrbach and Lorenzo Cavallaro of Technische Universität Darmstadt, University College London and Mohamed bin Zayed University of Artificial Intelligence). Three steps.
First it fixes an insertion layer R, the layer whose states the image generator can still read. That layer is calibrated, not guessed: the authors pick 12 visually distinct common nouns and require every model to recover all 12 of 12. The chosen layers are BAGEL 8/28, Show-o2 27/28, SenseNova-U1 24/42, HunyuanImage-3.0 9/32, Liquid 6/28 and Janus-Pro 8/30, where the first number is the layer and the second the total layer count.
Second it carries the hidden state of a chosen token over. The state is taken from an extraction layer, centred on the target layer mean and rescaled by the spread of the two layers, so its scale lands where the generator recognises it.
Third it drops the state into a slot of an ordinary drawing request. The state goes into the thing slot of a request like a photo of a thing, and the model's own image generator does the rest.
That third step matters. On the two-hop test set, averaged over three models, giving only the word and no state scores 0.03; adding just the final normalisation gives 0.11; copying the state unchanged gives 0.25; a Jacobian map gives 0.27; the full V-lens gives 0.33.
How accurate the pictures are
The first test deliberately leaves the object unnamed, and the two descriptions even share the same last word: an animal with a long trunk and a tall plant with a thick trunk both end in trunk. Reading that hidden state should give an elephant in one case and a tree in the other. Over 30 description pairs, BAGEL gets 2/30 right against a chance level of 0.03 with p=0.25; Show-o2 gets 8/30 with p=0.004; SenseNova-U1 4/30 with p=0.064; HunyuanImage-3.0 6/30 with p=0.016.
The second test is easier: one word with two meanings, such as seal as the animal or as the stamp, and the picture has to follow the context. Over 20 words BAGEL reaches 12.5/20 and the other three models all reach 15/20, with p=0.0001 throughout.
Scoring runs through three multimodal judges, Qwen3-VL, Gemma-3 and LLaVA-OneVision-2. A picture counts when at least two of them identify the subject, and a state counts when at least two of three drawing seeds succeed. The judges were calibrated too: on a labelled set the automatic judgement is right on 146/156 pictures, including 67/72 with background distractors, and three human raters each agree with the automatic judgement on at least 88% of 600 sampled pictures.
The depth result is the sharpest. In Show-o2 the concept is still recoverable on 11 of 15 questions at layer 24, then drops to 0/15 at layer 27, right before the answer is emitted. The prespecified difference is 0.73 with a 95% interval of [0.53, 0.93]. The concept is wiped out at the moment the answer lands.
An even harsher test asks the model to think about one thing while copying an unrelated sentence verbatim. In Show-o2 the word the inside the copied sentence draws a bicycle at layer 25 in all three seeds when the model is told to think about a bicycle, and in no seed when it is told to think about a rose or given no such instruction. In HunyuanImage-3.0 the final period token of the copy reveals a bread roll at layer 25.
Lens tuning: the pictures can train the model
If a state can be drawn, it can also serve as a supervision signal. The authors attach rank-8 LoRA adapters to the text layers and supervise a hidden state with pictures the frozen model draws itself. They call this lens tuning.
On pronoun reference, the untrained models are mostly blind: BAGEL 0/20 pairs, SenseNova-U1 4/20, HunyuanImage-3.0 0/20, while Show-o2 already sits at 16/20. After tuning, over 120 test pairs and three seeds, BAGEL goes from 0.02 to 0.86, SenseNova-U1 from 0.11 to 0.94 and HunyuanImage-3.0 from 0.06 to 0.86, which is the rise from 2-11% to 86-94% the paper reports.
More surprising is that purely visual supervision also improves text answers. Over 80 training stories, Show-o2 correct text answers rise from 36.3 to 49.6 with paired correct pictures rising from 52.5 to 73.3; HunyuanImage-3.0 goes from 40.0 to 50.8 on text and from 37.5 to 64.6 on pictures.
Swapping the trained subject state back into the original model lifts correct answers from 41/56 to 47-49/56, carrying 21% to 39% of the margin gain. The control arm is there too: ordinary drawing supervision lowers GenEval by 23.3 points in BAGEL, so the gain from lens tuning is not just extra training. On general capability, GenEval and MMLU-Pro show no significant loss, and all six intervals include zero.
The cold water
This is not a real-time tool yet: every reading requires generating a full image, and the paper measures 9 to 110 seconds per reading.
The item sets are small, mostly hand written: 16 two-hop questions, 30 description pairs, 20 two-meaning words, 20 pronoun pairs, 16 two-state pairs, 40 state scenarios and 28 interchange scenarios. The limitations section says plainly that recovery of unnamed objects is modest and is not significantly above chance in BAGEL or SenseNova-U1.
There is also a self-referential problem: lens tuning trains hidden states through the same lens used to evaluate them. And word state supervision without pictures beats lens tuning on unseen phrases in SenseNova-U1 and on unseen nouns and phrases in HunyuanImage-3.0. One more detail: in the wrong-picture control arm for HunyuanImage-3.0, correct text answers also rise, from 40.0 to 47.5, so the text improvement does not depend entirely on the picture being right.
In terms of actual payoff, the text answer gain is about ten extra stories out of 80, measured only on the supervised stories. The interchange test moves whole hidden states and has no control donor trained on different picture targets, so it does not isolate what the pictures contributed. The general capability checks use a single training seed.
What comes next
The mechanism is not new from scratch. Diffusion Lens, published at ACL 2024 (see https://aclanthology.org/2024.acl-long.524/), already read text-to-image text encoders by generating images from intermediate representations, and found that uncommon concepts require further computation compared to common ones. The difference is that it needs an external text encoder and a separate diffusion model, while V-lens uses the generator that the unified model already carries.
There is also prior art warning about the same failure mode. Faithful-Patchscopes (see https://arxiv.org/abs/2602.00300) found that when a language model decodes its own hidden representations, inherent linguistic patterns can override the contextual information encoded in them: a hidden state that encodes the attribute purple for broccoli still produces green in the explanation, and faithfulness drops by 18.84% on average. V-lens swaps the decoding channel, but the ceiling stays: the generator can only draw what it knows how to draw, and a missing picture does not establish that the hidden state lacks the concept.
So what is the real prize here? Not that the model can draw. It is that a unified model hands you a magnifying glass for its own internals. The catch is that the glass still has to take a photograph, at 9 to 110 seconds per shot, which is fine for offline auditing and useless for live debugging. Once that becomes cheap, the question of what the model is actually thinking stops being a metaphor.
The code is at https://github.com/multimodal-ai-lab/V-lens. The repository ships adapters for the models, the judge panel and the lens tuning training scripts, but as the README states it carries no pictures, results or weights.