DSEG
Dynamic Semantic Evolution Graph — an ongoing investigation into whether dynamically constructed semantic graphs can provide useful and testable structure for sequence models.
The arc
My research has moved from structural reliability (Triad) to scalability (Highway). Building on that work, I am now investigating a third question — interpretability and control. In today's large models the internal state is a single, entangled vector: powerful, but hard to inspect and harder to steer.
DSEG asks what changes if we explicitly define every basic decision unit of the internal state — instead of leaving it as one opaque black-box function.
The premise
Give the representation a structure you can read — a graph of interpretable units and the relations between them.
Decision units
Treat the internal state not as one entangled vector, but as a set of discrete, interpretable decision units — each a small, nameable carrier of one part of the computation.
Relations
Model the structure between those units, so the graph of relations — not a single opaque activation — is what carries the reasoning forward.
Control by construction?
The open question: if the structure were explicit, could the representation become something you can inspect, audit, and edit by construction — rather than probe after the fact? Testing that is what this work is currently about.
Why it matters
If representations were structured this way by construction, interpretability and control would stop being post-hoc add-ons and become properties of the model itself. I see DSEG as a graph-structured semantic control layer — one whose value I currently locate in inspection and editing rather than in perplexity or speed — and as the natural continuation of the structural thinking behind Triad and Highway.
Early findings
Early experiments showed that coarse structural signals were usable by the model and distinguishable from shuffled controls. However, several stronger architectural hypotheses — including unconstrained per-edge parameterization — did not improve predictive performance. I report both directions here, because the negative ones are what shaped the question I am asking now.
Status
This is early, exploratory work that I am developing hands-on. A preprint is in preparation; I will link concrete findings here once they are public.