DSEG
Dynamic Semantic Evolution Graph — an ongoing investigation into whether dynamically constructed semantic graphs can provide useful and testable structure for sequence models.
How this question arrived
This is the fourth stage of one arc, not a jump from graph algorithms into language models. Triad made overlapping community structure reliable by putting all of it into a single constrained program — and ran into a scaling limit that was intrinsic rather than incidental. Highway asked how much of that structure the inference actually needs, and answered with a deliberately sparse backbone: scalability bought by keeping less. Scout then removed the assumption both of them still rested on — that the graph is sitting there to be read — and asked which structure is worth observing at all when looking costs.
- 01 Triad structural reliability
- 02 Highway scalability through sparse structure
- 03 Scout what is worth observing
-
04 DSEG inspectable structure
Across those three the operative claim is always the same one: this part of the structure carries the inference, that part does not. Each time it was checkable, because the structure was laid out in front of me. DSEG asks the same question one level in — of a model's internal representation, where the state is a single entangled vector: powerful, but hard to inspect and harder to steer.
So the question becomes: what changes if every basic decision unit of the internal state is defined explicitly — instead of being left as one opaque 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.
What carries over
Two habits come with me from the graph work rather than being invented here. From Highway: a sparse backbone is a hypothesis about which structure is load-bearing, and it is worth stating as one. From Scout: the value of a piece of structure is not a property of the structure alone — it only means something relative to what you are asking it to do. Both push in the same direction here: whatever structure I put inside a representation, the interesting part is the protocol that decides whether it is doing any work.
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, Highway and Scout.
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.