gni.works
applied research at the boundary of attention, signal,
and visual reasoning
agni.works investigates the behavior of attention in vision-language models during extended reasoning — how it decays, how it fails silently, and how it can be intervened on without retraining.
The framing: attention is not a fixed architectural property. It is a process that can be conditioned at inference time on signals from the reasoning trace, on uncertainty in the model's own outputs, and on exogenous predictive models running alongside it. The load-bearing claim of the current work is that this conditioning recovers visual grounding without modifying the base model.
REVEAL →
A framework for modulating visual attention in vision-language models during extended chain-of-thought reasoning, via learned phase displacement, patch salience bias, and exogenous conditioning. Provisional patent filed with USPTO, April 2026.
Research →
Active investigations across conversational sentiment analysis, equine biomechanics from the PFERD motion-capture dataset, and evaluation patterns for vision-language models in real-world deployments. Several of these are upstream of REVEAL methodologically; all are independent surfaces.
Engineering →
The applied technical work that supports the research — vision-language model pipelines with prompt engineering and multi-pass validation, edge inference on Jetson platforms, and a hybrid edge-cloud architecture for contextual interpretation of affect signals.
Writing
Notes and observations from active research are published in /writing as they become available.
Contact
For research collaboration, dataset access discussion, or correspondence: Arjun.Joshi@Agni.works