REVEAL

Modern vision-language models exhibit a structural pathology during extended chain-of-thought reasoning: as the model generates longer reasoning sequences, attention to visual tokens decays. No specific model is to blame; the decay follows from how attention is computed across long contexts in current architectures. The visual tokens are a fixed prefix of the sequence; every generated token dilutes their share. For one open family of reasoning VLMs this has been proven formally: attention to visual tokens declines as reasoning progresses [1].

The failure mode is invisible. The model continues to produce confident outputs that read as grounded in the visual input, while the actual visual signal has dropped out of the reasoning process. Standard evaluation metrics do not surface this: the outputs look correct because they are linguistically coherent, but the model is no longer looking at what it claims to be looking at.

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.

REVEAL addresses the decay by modulating visual attention in response to reasoning-time signals, including model uncertainty and exogenous conditioning inputs, to recover visual grounding without retraining the base model.

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

The three mechanisms can be applied independently or composed. The framework is architecture-agnostic and designed for fine-tuning rather than pretraining, making it deployable on existing models without the cost of training from scratch.

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 third mechanism rests on a premise that can be tested without touching a model: that a lightweight, interpretable predictor running alongside the video can flag the moments worth looking at. The pipeline fits classical predictive baselines over pose kinematics extracted from video, and scores every frame by how far the observed motion departs from prediction. When parallel channels of evidence disagree, widening perceptual search is the statistically sound response, a result with deep roots 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.

REVEAL trains against a purpose-built corpus. Every clip passes through pose estimation, depth lifting, kinematic feature extraction, predictive baselining, and multi-pass chain-of-thought generation with cross-model review.

All six numbers describe the training scaffolding. Attention-level results will be reported when the training runs complete; this page makes no claims about them yet.

REVEAL is being validated across two application domains, both selected because they require sustained visual attention during multi-step reasoning over temporally rich signals:

Deception detection in video. Behavioral cues for deception are subtle, distributed across modalities (facial expression, gaze, gesture, vocal patterns), and require integration over time. This is a stress test for visual grounding during extended reasoning: the kind of task where attention decay pushes a model onto linguistic priors and away from the visual evidence.

Equestrian biomechanics analysis. Rider posture and horse gait analysis involve fine-grained spatial reasoning over multiple subjects (rider plus horse) with temporally evolving signals. The domain has well-developed quantitative biomechanics literature, providing reference data for empirical validation. The measured trace above comes from this domain.