The Elephant in the Question: Unified Models Picture the Concepts Behind Their Answers
Organizations: Technische Universität Darmstadt · University College London · Mohamed bin Zayed University of Artificial Intelligence
Abstract
Visualizing intermediate representations could help us inspect the reasoning steps underlying a model output. Unified multimodal models offer this opportunity by combining understanding and generation in one architecture. Their image generators can access representations formed while processing text. We introduce the V-lens, which uses this connection to visualize a chosen token's hidden state without training a decoder. Across four diffusion-based and two purely autoregressive unified models, the resulting images reveal key intermediate concepts that connect a question to its answer even when the concepts are never named explicitly. Controlled tests show how this content changes with context, token position, and depth. We then introduce lens tuning, which supervises selected hidden states with pictures through the V-lens. Across BAGEL, SenseNova-U1, and HunyuanImage-3.0, the rate of correct visualizations of pronoun referents rises from 2-11% to 86-94%. Surprisingly, such visual supervision also improves text answers without supervising the answer tokens directly. Code is available at https://github.com/multimodal-ai-lab/V-lens.
Figures & tables
| Model | Success | Chance | |
|---|---|---|---|
| Unnamed objects (30 description pairs) | |||
| BAGEL | 2/30 | 0.03 | 0.25 |
| Show-o2 | 8/30 | 0.13 | 0.004 |
| SenseNova-U1 | 4/30 | 0.07 | 0.064 |
| HunyuanImage-3.0 | 6/30 | 0.10 | 0.016 |
| Two meanings (20 words) | |||
| BAGEL | SenseNova-U1 | HunyuanImage-3.0 | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Supervision | Verbs | Nouns | Phrases | Verbs | Nouns | Phrases | Verbs | Nouns | Phrases |
| None | 0.00 | 0.03 | 0.03 | 0.05 | 0.13 | 0.15 | 0.00 | 0.00 | 0.18 |
| Text | 0.00 | 0.05 | 0.02 | 0.04 | 0.23 | 0.19 | 0.00 | 0.03 | 0.17 |
| Ordinary drawing | 0.01 | 0.13 | 0.05 | 0.25 | 0.58 | 0.51 | 0.00 | 0.01 | 0.24 |
| Word state | 0.52 | 0.62 | 0.90 | 0.60 | 0.96 | 1.00 | 0.65 | 0.97 | 0.99 |
| Lens tuning (ours) | 0.70 | 0.93 | 0.96 | 0.84 | 0.99 | 0.99 | 0.71 | 0.94 | 0.92 |
| Model | Supervision | Correct text | Correct picture |
|---|---|---|---|
| Show-o2 | Untrained | 36.3 | 52.5 |
| Lens tuning (wrong) | 34.6 0.7 | 27.1 1.9 | |
| Lens tuning (right) | 49.6 9.5 | 73.3 1.9 | |
| Hunyuan- Image-3.0 | Untrained | 40.0 | 37.5 |
| Lens tuning (wrong) | 47.5 4.3 | 45.8 3.1 | |
| Lens tuning (right) | 50.8 5.6 | 64.6 3.8 |
| Model | V-lens | No | No | Only | Only | |
|---|---|---|---|---|---|---|
| BAGEL | 10/35 | 9/35 | 0/35 | 0/35 | 8/35 | 0/35 |
| Show-o2 | 15/30 | 13/30 | 12/30 | 11/30 | 11/30 | 9/30 |
| SenseNova-U1 | 8/39 | 7/39 | 1/39 | 1/39 | 4/39 | 1/39 |
| Mean | 0.33 | 0.29 | 0.14 | 0.13 | 0.23 | 0.11 |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Text layers | Hidden size | Fixed extraction layers | Pixels | Steps | Guidance | Seconds | |
|---|---|---|---|---|---|---|---|---|
| BAGEL | 28 | 3584 | 8 | 6, 8, 14 | 28 | 4.0 | 9 | |
| Show-o2 | 28 | 3584 | 27 | 20, 24, 27 | 50 | 7.5 | 12.6 | |
| SenseNova-U1 | 42 | 4096 | 24 | 20, 24, 32 | 50 | 4.0 | 9.4 | |
| HunyuanImage-3.0 | 32 | 4096 | 9 | 9, 12, 15 | 50 | 2.5 | 110 |
| BAGEL | Show-o2 | SenseNova-U1 | |||||
| Variant | Two-hop | Both meanings | Two-hop | Both meanings | Two-hop | Both meanings | Score |
| Val | |||||||
| The V-lens | 24/44 | 31/90 | 31/44 | 50/90 | 20/52 | 33/90 | 0.484 |
| Final norm | 0/44 | 0/90 | 21/44 | 21/90 | 1/52 | 0/90 | 0.122 |
| Final norm at | 1/44 | 0/90 | 22/44 | 46/90 | 10/52 | 10/90 | 0.223 |
| Test | |||||||
| BAGEL layer | 0 | 2 | 4 | 8 | 12 | 16 | 20 | 24 | 27 | |||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Words right | 10 | 12 | 12 | 12 | 8 | 0 | 0 | 0 | 0 | |||
| Show-o2 layer | 0 | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 20 | 24 | 27 |
| Words right | 10 | 12 | 11 | 11 | 11 | 11 | 11 | 12 | 12 | 12 | 12 | 12 |
| SenseNova-U1 layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 41 |
| Words right | 10 | 12 | 12 | 12 | 12 | 12 | 12 | 10 | 2 | 0 | 0 | 0 |
| HunyuanImage-3.0 layer | 0 | 3 | 6 | 9 | 12 | 15 | 18 | 21 | 24 | 27 | 30 | 31 |
| Model | Test | Right | Rate | Chance | 95% interval | Other fixed layers | Any layer | |
|---|---|---|---|---|---|---|---|---|
| BAGEL | Objects | 2/30 | 0.067 | 0.033 | 0.24878 | [0.000, 0.167] | L6 0.033, L14 0.167 | 0.200 |
| BAGEL | Pronouns | 0/20 | 0.000 | 0.000 | 1.00000 | [0.000, 0.000] | L6 0.000, L14 0.050 | 0.050 |
| BAGEL | Meanings | 12.5/20 | 0.625 | 0.312 | 0.00010 | [0.475, 0.775] | L6 0.300, L14 0.450 | 0.650 |
| Show-o2 | Objects | 8/30 | 0.267 | 0.133 | 0.00390 | [0.133, 0.433] | L20 0.067, L24 0.267 | 0.333 |
| Show-o2 | Pronouns | 16/20 | 0.800 | 0.399 | 0.00010 | [0.600, 0.950] | L20 0.700, L24 0.900 | 0.900 |
| Show-o2 | Meanings | 15/20 | 0.750 | 0.374 | 0.00010 | [0.575, 0.900] | L20 0.700, L24 0.775 | 0.800 |
| Pairs right at | Word it alone | Selecting clause alone, 90 pictures | |||||
| Model | Verb changed | Verb held | Intended object | Intended | Alternative | Both | Neither |
| BAGEL | 0/10 | 0/10 | 0/70 | 9 | 5 | 1 | 75 |
| Show-o2 | 10/10 | 6/10 | 0/61 | 27 | 9 | 3 | 51 |
| SenseNova-U1 | 3/10 | 1/10 | 0/70 | 14 | 12 | 1 | 63 |
| Token type | BAGEL | Show-o2 | SenseNova-U1 | HunyuanImage-3.0 |
|---|---|---|---|---|
| Question words | 18 | 59 | 5 | 1 |
| Question mark | 3 | 9 | 3 | 0 |
| Instruction | 0 | 2 | 0 | 0 |
| Chat format | 0 | 54 | 71 | 0 |
| Immediately before answer | 0 | 14 | 0 | 0 |
| First answer token | 8 | 5 | 1 | 0 |
| Context tests at | Depth study | ||||
| Readout | Descriptions | Pronouns | Meanings | At | Later |
| BAGEL | |||||
| V-lens | 2/30 | 0/20 | 12.5/20 | 10/15 | 5/15 |
| Logit lens | 0/29 | 0/20 | 0/20 | 0/15 | 5/15 |
| Tuned lens | 0/29 | 0/20 | 0/20 | 0/15 | 3/15 |
| J-lens | 0/29 | 0/20 | 0/20 | 1/15 | 5/15 |
| BAGEL | Show-o2 | SenseNova-U1 | HunyuanImage-3.0 | |||||
| Condition | Dev | Held-out | Dev | Held-out | Dev | Held-out | Dev | Held-out |
| Explicit request drawn as text | 7/7 | 7/7 | 6/6 | 8/8 | 8/8 | 7/7 | 8/8 | 7/7 |
| Modifier and noun states from the explicit request | 7/7 | 7/7 | 6/6 | 8/8 | 8/8 | 7/7 | 7/8 | 7/7 |
| Modifier and noun states from a plain sentence | 7/7 | 7/7 | 6/6 | 8/8 | 8/8 | 7/7 | 8/8 | 7/7 |
| Noun state alone | 0/7 | 0/7 | 0/6 | 0/8 | 1/8 | 0/7 | 1/8 | 1/7 |
| Modifier and pronoun states from the text | 1/7 | 0/7 | 2/6 | 5/8 | 0/8 | 1/7 | 2/8 | 3/7 |
| BAGEL | SenseNova-U1 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Supervision | Dev | Verbs | Nouns | Phrases | All tests | Dev | Verbs | Nouns | Phrases | All tests |
| None | 0.000 | 0.000 | 0.025 | 0.025 | 0.017 | 0.050 | 0.050 | 0.125 | 0.150 | 0.108 |
| Text | 0.050 | 0.000 | 0.050 | 0.017 | 0.022 | 0.075 | 0.042 | 0.233 | 0.192 | 0.156 |
| Lens tuning | 0.908 | 0.700 | 0.925 | 0.958 | 0.861 | 1.000 | 0.842 | 0.992 | 0.992 | 0.942 |
| Ordinary drawing | 0.233 | 0.008 | 0.133 | 0.050 | 0.064 | 0.575 | 0.250 | 0.575 | 0.508 | 0.444 |
| Word state | 0.642 | 0.517 | 0.617 | 0.900 | 0.678 | 0.975 | 0.600 | 0.958 | 1.000 | 0.853 |
| BAGEL | SenseNova-U1 | |||||||
|---|---|---|---|---|---|---|---|---|
| Supervision | Naming | Untrained | Retained | General | Naming | Untrained | Retained | General |
| None | 0.779 | 0.758 | 1.000 | 0.975 | 0.829 | 0.817 | 1.000 | 0.975 |
| Text | 0.972 | 0.910 | 0.789 | 0.975 | 0.974 | 0.910 | 0.787 | 0.950 |
| Lens tuning | 0.735 | 0.764 | 0.982 | 0.975 | 0.918 | 0.776 | 0.965 | 1.000 |
| Ordinary drawing | 0.782 | 0.851 | 0.860 | 0.975 | 0.947 | 0.969 | 0.873 | 0.967 |
| Word state | 0.537 | 0.580 | 0.772 | 0.975 | 0.050 | 0.481 | 0.924 | 0.958 |
| BAGEL | SenseNova-U1 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Supervision | Verbs | Nouns | Phrases | All texts | Pairs | Verbs | Nouns | Phrases | All texts | Pairs |
| None | 0.425 | 0.375 | 0.338 | 0.379 | 0.008 | 0.463 | 0.463 | 0.338 | 0.421 | 0.075 |
| Text | 0.454 | 0.450 | 0.329 | 0.411 | 0.022 | 0.446 | 0.463 | 0.254 | 0.388 | 0.047 |
| Lens tuning | 0.529 | 0.475 | 0.471 | 0.492 | 0.094 | 0.533 | 0.550 | 0.529 | 0.537 | 0.146 |
| Ordinary drawing | 0.942 | 0.912 | 0.925 | 0.926 | 0.854 | 0.975 | 0.946 | 0.992 | 0.971 | 0.942 |
| Word state | 0.504 | 0.463 | 0.487 | 0.485 | 0.064 | 0.479 | 0.550 | 0.450 | 0.493 | 0.102 |
| GenEval, image generation | MMLU-Pro, text knowledge | ||||
|---|---|---|---|---|---|
| Model | Training | Pass rate | Change, 95% interval | Accuracy | Change, 95% interval |
| BAGEL | None | 87.5 | 28.7 | ||
| Text | 84.2 | 3.3 [ 6.7, 0.8] | 35.0 | +6.3 [+3.3, +9.3] | |
| Ordinary drawing | 64.2 | 23.3 [ 30.8, 15.8] | 26.3 | 2.3 [ 4.7, 0.3] | |
| Lens tuning | 87.5 | +0.0 [ 5.0, +5.0] | 29.3 | +0.7 [ 1.0, +2.3] | |
| SenseNova-U1 | None | 95.0 | 28.3 | ||
| Model | Picture targets | Seed | As written | Reversed | Both orders | Right pictures |
|---|---|---|---|---|---|---|
| Show-o2 | Untrained | – | 48 | 53 | 29 | 42 |
| Show-o2 | Right state | 0 | 55 | 61 | 38 | 59 |
| Show-o2 | Right state | 1 | 61 | 50 | 33 | 57 |
| Show-o2 | Right state | 2 | 68 | 58 | 48 | 60 |
| Show-o2 | Wrong state | 0 | 45 | 54 | 27 | 22 |
| Show-o2 | Wrong state | 1 | 46 | 55 | 28 | 23 |
| Subject states interchanged at all layers | At layer 27 only | ||||
| Training seed | Answers | Mean margin | Margin gain carried | Answers | Mean margin |
| 0 | 47 | +1.38 | 38% | 40 | +1.032 |
| 1 | 47 | +1.43 | 39% | 40 | +1.029 |
| 2 | 49 | +1.53 | 21% | 40 | +1.019 |
| Study | Example texts | Read | Intended |
|---|---|---|---|
| A, matched descriptions, 30 pairs | How many legs does the huge land animal with a long trunk and tusks have? Does the whiskered Arctic animal with flippers and long tusks breathe air? | tusks | elephant walrus |
| B, reference, verb changed, 10 pairs | A sheep and a bicycle stood in the yard. She sheared it. A sheep and a bicycle stood in the yard. She pedaled it. | it | sheep bicycle |
| B, reference, verb held, 10 pairs | A dog and a teapot were in the yard. She fed it. A horse and a teapot were in the yard. She fed it. | it | dog horse |
| C, two meanings, 20 words in both orders | Hanging upside down in the attic, a furry bat slept; among the baseball equipment lay a wooden bat. Among the baseball equipment lay a wooden bat; hanging upside down in the attic, a furry bat slept. | bat, bat | bat animal baseball bat |
| Depth study, 16 questions | How many legs does the animal that spins webs have? With the name given first. A spider is an animal that spins webs. How many legs does the animal that spins webs have? | webs | spider, alternative stop sign, answer eight or 8 |
| Two states, 8 development pairs | The car was painted bright red. Later I looked at it. Plain sentence: I saw a red car on the street. Partner: bicycle, values red and blue | red, it | red car |
| Category | Training nouns | Development | Held-out | Expressions, training / held-out | Category question |
| animal | dog, rabbit, cow, goat, donkey, camel, deer, kangaroo | cat, horse, pig | sheep, hamster, llama | fed it, petted it / groomed it | Is it an animal? |
| food | apple, banana, orange, pear, cake, sandwich, cookie, watermelon | pizza, pineapple, strawberry | carrot, donut, burger | ate it, tasted it / bit into it | Is it food? |
| device | radio, laptop, microwave, vacuum cleaner, hair dryer, printer, washing machine, sewing machine | lamp, toaster, blender | fan, television, kettle | switched it on, turned it off / plugged it in | Does it run on electricity? |
| vehicle | truck, bus, tractor, van, bulldozer, motorcycle, golf cart, snowmobile | taxi, jeep, forklift | car, ambulance, limousine | drove it, parked it / steered it | Is it a vehicle? |
| openable | umbrella, book, backpack, wallet, envelope, briefcase, purse, bottle | box, jar, gate | suitcase, cupboard, refrigerator | opened it, closed it / shut it | Can you open and close it? |
| clothing | jacket, scarf, boot, glove, dress, shirt, skirt, belt | sock, sneaker, apron | hat, sweater, tie | wore it, put it on / took it off | Do people wear it? |
| Model | Words right at | Chance | |
|---|---|---|---|
| Liquid | 12/20 | 0.300 | 0.0002 |
| Janus-Pro | 11.5/20 | 0.287 | 0.0001 |
| Liquid layer | 0 | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 | 20 | 22 | 24 | 26 | 27 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Description’s last word (of 13) | 2 | 7 | 8 | 11 | 12 | 11 | 11 | 11 | 11 | 11 | 11 | 11 | 11 | 11 | 11 | |
| Best later token (of 13) | 1 | 4 | 4 | 3 | 8 | 6 | 4 | 3 | 3 | 4 | 5 | 10 | 10 | 11 | 10 | |
| Name in a plain sentence (of 15) | 8 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 14 | 15 | |
| Janus-Pro layer | 0 | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 | 20 | 22 | 24 | 26 | 28 | 29 |
| Description’s last word (of 6) | 0 | 0 | 0 | 0 | 4 | 2 | 1 | 0 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 0 |
| Best later token (of 6) | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 1 |
| Result | Scripts |
|---|---|
| Insertion layers ( Tab. S4 , Appendix I ) | calibrate_insertion_layer.py |
| Transport statistics ( Eq. 1 ) | fit_transport.py , collect_states.py , build_lens_stats.py , lens_stats_own_tokens.py |
| Lens variants and final normalization ( Tab. 1 , Tabs. S2 and S3 ) | draw_cells.py , self_gain.py , score_cells.py , score_variant_choice.py , score_variants.py |
| Ablation ( Tab. 5 ) | lens_ablation.py , draw_cells.py , score_cells.py , score_ablation.py |
| Judge calibration ( Tab. S6 ) | draw_judge_calibration.py , run_judge.py , score_judge_calibration.py |
| Context ( Tabs. 2 and S7 ) | draw_cells.py , score_context.py |
Multimodal Model Interpretability
Multimodal model interpretability studies how AI systems combine and use images, text, audio, video, and other signals, and how to explain their decisions.
See all 141 papers