Research Statement

Abhinav Madahar · अभिनव ਮਦਾਹਰ
April 10, 2026 CE

My research begins from taking seriously the possibility that transformer-based models are candidate general intelligences. If that is right, then they should not be studied only instrumentally—as systems for producing outputs, improving benchmarks, or supporting new techniques—but also descriptively, as systems in their own right. I am interested in one route toward that shift.

At present, much of AI research still analyses transformer-based models through unstable implementation-level proxies: weights, architectures, prompting methods, benchmarks, and local interventions. These are useful handles, but they are not yet a stable object of theory. The result is a field that is empirically productive but descriptively limited. Systems with increasingly general cognitive capacity are studied mainly as engineering artefacts rather than as objects whose structure should itself be characterised. My work is motivated by the view that this is an important scientific gap.

I do not take my own programme to be the only way such a shift could occur. A more descriptive science of AI could emerge through many kinds of work—behavioural, mechanistic, theoretical, and methodological. My contribution is aimed at one part of that transition: developing a level of description at which transformer-based models can be analysed as stable objects rather than as collections of contingent implementations. The broader aim is to help move AI toward a setting in which its central systems can be described, compared, and understood on principled terms.

The central idea guiding my current work is that trained models should be studied up to a behaviour-preserving equivalence. On this view, the relevant object is not an individual model instance, but an equivalence class of systems that share a common underlying structure. I am interested in identifying a mathematical domain in which these equivalence classes can be represented and compared, so that properties that appear fragile at the level of parameters become stable and intelligible at the level of structure. This is the sense in which I approach transformer-based models descriptively: not by cataloguing every implementation detail, but by trying to locate the level at which the system itself becomes scientifically legible.

Within this framework, I treat reasoning as a structural phenomenon rather than merely a procedure. Existing work often studies reasoning through the techniques used to elicit it—prompting, search, decomposition, or verification. Those methods can be effective, but they do not by themselves answer a deeper question: what is it about the underlying system that makes such interventions work at all? My research therefore asks not only how reasoning can be improved, but what reasoning consists in at the level of the system’s organisation. More broadly, I am interested in how generalisation, control, and interpretability look once the underlying object of study is formulated structurally rather than instrumentally.

A related part of this programme concerns architecture and asymptotic behaviour. Rather than treating architectures as interchangeable implementations, I am interested in how architectural choices constrain the kinds of structures a model can realise, especially in the large-scale limit. This provides a way to think about architectural comparison that is not exhausted by finite-scale benchmark differences or engineering contingencies. It also supports a broader descriptive ambition: not only asking what a given model does, but what kinds of systems different architectural families are.

The broader ambition of this research is to contribute to a more descriptive science of AI. If transformer-based systems are plausible candidate general intelligences, then the field will ultimately need ways of studying them as systems rather than only as tools. My work explores one path toward that goal by asking what kind of object a transformer-based model is, at what level it should be described, and how reasoning and related capacities arise from that level of organisation. I see this not as a complete solution, but as one part of a larger effort to make modern AI more intelligible, more comparable, and more amenable to cumulative scientific understanding.