REVEAL
Reasoning-Entropy Visual Re-Attention via Learned Displacement and Salience
- provisional patent
- filed — USPTO, Apr 2026
- signal pipeline
- validated — 499 clips
- training corpus
- 10.9k CoT examples
- attention training
- Third Party Review of Methodology & Initial Findings
A vision-language model stops looking at the image partway through thinking about it. That is the whole problem. REVEAL is an attempt to do something about it from inside the attention mechanism, while the model reasons, rather than by handing it the picture again afterwards.
Three interventions, usable separately or together. None of them retrain the base model and none of them add a single token to the sequence. What they change is where attention lands, in response to signals that are already available while the model is generating.
On tasks that take several passes of looking, a model conditioned this way can beat its own base and models several times its size. Those are the tasks that cost everyone else the most: more tokens, more memory, more time on the clock, more on the invoice.
The framework and the first findings are with independent reviewers. Until that comes back, the mechanisms are described here at the level the patent filing permits and no further, and the numbers stay off the page.
Problem space
The visual tokens sit at the front of the sequence and they do not grow. Every token the model generates after them takes a share of the same attention budget. So the longer the reasoning runs, the less of it is pointed at the image — not because any particular model is badly built, but because that is what the arithmetic does. For one open family of reasoning VLMs it has been shown formally [1].
Nothing announces it. The model keeps producing answers that read as though they came from the image, in the same confident register, after the image has effectively left the conversation. Standard metrics score those answers well, because they are coherent, and coherence is most of what the metrics measure. The model has stopped looking at the thing it is describing and there is no line in the output where that happens.
The existing remedy is token re-injection: pasting copies of the visual tokens back into the sequence at reflection points. It works, but every re-injection lengthens the sequence and its memory footprint. REVEAL instead modulates the geometry of attention over the visual tokens that are already there, at zero additional sequence length.
Framework
The framework comprises three mechanisms:
- Phase displacement of position encodings [3], which geometrically broadens the attention pattern over visual tokens
- Patch salience bias, a per-patch content-aware modulation operating at the attention layer
- Exogenous conditioning signals, allowing domain-specific predictive models to drive attention adjustments based on prediction-vs-observation residuals
Any one of the three can be run on its own; they compose. The design assumption throughout is fine-tuning, not pretraining, because a method that only works if you train from scratch is a method almost nobody can use.
A provisional patent application covering the framework was filed with the USPTO on April 8, 2026. A nonprovisional application is planned for April 2027. Mechanism internals (the loss formulation, module architectures, and training procedure) are covered by the filing and not described here.
The conditioning signal, measured
The third mechanism rests on a claim that can be checked without going near a model: that something small and legible, running beside the video, can tell you which moments are worth a second look. So the pipeline fits classical predictors over pose kinematics and scores every frame by how far the motion departs from what was predicted. Surprise, in other words, measured rather than asserted. There is also a reason to trust the shape of that response — when channels of evidence disagree, widening the search is the statistically correct move, which is old ground in human cue integration [2].
A close reading of this trace, including why the late spikes are tracker noise and why that distinction shapes the signal design, is in the first research note.
Empirical base
There was no corpus for this, so one had to be built. Each clip goes through pose estimation and depth lifting, then kinematic features come off the skeletons, then a classical predictor runs over those to produce the residual channels. Chain-of-thought is generated on top in several passes, with one model reviewing another's work.
- 499
- video clips instrumented end-to-end
- 30,167
- frames with dual-skeleton pose — 39-keypoint quadruped + 17-keypoint rider
- 41
- kinematic features extracted per frame
- 7
- residual channels feeding the conditioning signal
- 10,931
- validated chain-of-thought examples in the current corpus
- 2
- validation domains — behavioral video analysis, equestrian biomechanics
Every number above is scaffolding — what was built, not what it proved. The attention-level results are a separate matter and this page does not claim them.
Validation domains
Two domains, picked because both punish a model for looking away:
Deception detection in video. The cues are small, spread across face, gaze, gesture and voice, and only mean anything once they are put together over time. It is close to a worst case for attention decay: when the visual signal thins out, a model has a rich set of linguistic priors about what deception sounds like waiting to fill the gap, and it will use them.
Equestrian biomechanics analysis. Two bodies moving together, each with its own skeleton, judged on spatial detail that changes frame to frame. The useful part is that equine biomechanics has a real quantitative literature behind it, so there is something to check the model against that is not another model. The trace above comes from this domain.
Status
- 2026-04Provisional patent application filed (USPTO).
- 2026-05Signal-extraction pipeline validated across 499 clips; chain-of-thought corpus assembly complete.
- 2026-06Training infrastructure validated on local dual-GPU hardware.
- 2026-07Attention-level training runs — in preparation.
- 2027-04Nonprovisional filing planned.
Notes from the work in progress are published in /writing as they become available.
References
- [1] Chu, X. et al. Qwen Look Again: Guiding Vision-Language Reasoning Models to Re-attention Visual Information (2025). arXiv:2505.23558.
- [2] Ernst, M. O. & Banks, M. S. Humans integrate visual and haptic information in a statistically optimal fashion. Nature 415, 429–433 (2002).
- [3] Su, J. et al. RoFormer: Enhanced Transformer with Rotary Position Embedding (2021). arXiv:2104.09864.
Contact
For research inquiries, dataset access discussions, or collaboration: Arjun.Joshi@Agni.works