Home / About

About

I am an undergraduate researcher at the Machine Learning Lab at IIIT Hyderabad, where I study world models, representation learning, graph learning, machine unlearning, and AI safety.

My research is centered on a recurring question: what assumptions are our learning systems treating as fixed, even when they should themselves be objects of inference or design?

In graph learning, this led me to investigate how much node-classification performance can be recovered using deterministic closed-form solvers, and what exact guarantees become possible when prediction is an explicit system rather than the result of iterative gradient training. The resulting preprint develops closed-form predictors with exact retrain-equivalent unlearning for graph-object modifications.

In a separate line of work, I study the distinction between a graph as evidence and a graph as an inference circuit. Most graph models assume that the observed graph should also determine the path along which computation flows. My work asks when that assumption is valid and what becomes possible when the two roles are separated; the resulting preprint separates them and tests the consequences.

My current world-model research studies representation collapse from the perspective of future consequences. Instead of requiring representations to remain generically diverse, I investigate whether they can compress aggressively while preserving precisely the distinctions needed for prediction, action, and control.

I also work on AI-safety evaluation, particularly the measurement of emergent misalignment and the degree to which reported effects survive changes in model family, intervention, behavioral construction, evaluator, and human framing.

Across these projects, I emphasize preregistration, reproducibility, negative results, and experiments designed to falsify the central claim. I have also built private research infrastructure for maintaining long-running autonomous research programs across hypotheses, experiments, evidence, and manuscript development.

Alongside academic research, I founded Varaigen, where I work on production AI systems and applied research.

IIIT Hyderabad · Machine Learning Lab Hyderabad, India

CV Scholar GitHub

Now

What I am working on now

Collapse-contained world models Testing whether future-consequence equivalence can replace generic anti-collapse objectives in JEPA-style learning.

Inference structure in graph learning Studying when the computational circuit should be separated from the observed evidence graph.

Human-grounded AI evaluation Auditing which model-behavior effects survive evaluator changes and blinded human judgment.

Autonomous research infrastructure Developing tools for persistent, falsification-oriented research campaigns.

For organisers and applications

Third-person bio

copy and paste

Aditya Gaur is an undergraduate researcher at the Machine Learning Lab at IIIT Hyderabad and the founder of Varaigen. His work spans world models, representation learning, graph learning, exact machine unlearning, AI-safety evaluation, and autonomous research infrastructure. He is the first author of “Closed-Form Node Classification with Exact Graph Unlearning,” which develops deterministic graph predictors with exact retrain-equivalent unlearning. His current research studies decision-grounded representation collapse, the separation of graphs as evidence from graphs as inference circuits, and human-grounded evaluation of emergent misalignment.

Closed-Form Node Classification with Exact Graph Unlearning

Graph predictors fitted by regularized linear solves rather than gradient training — above or within one point of the strongest reproduced tuned GNN on all sixteen benchmarks, and unlearnable exactly, model selection included.

the pdf here is the october 2026 revision. arxiv still shows the may 2026 version until its update posts.

Research / Evidence, not circuits

Public preprint, 2026

Graphs Are Evidence, Not Circuits

Graph learning has mistaken its evidence for its machinery. The evidence says what is observed; the query says what must be inferred; the circuit specifies how observations are combined. Message passing usually inherits its circuit from the graph — this paper makes that choice explicit, and tests it.

Read the paper Code — on request

+4.28AUC points from recompiling the circuit — same GCN update family, matched depth and parameters, identical evidence

.5075AUC after an isomorphic scramble: the whole circuit topology kept, its alignment with the evidence broken

9.6AUC points a linear readout stays below Bayes on one of two matched queries — after every observation is in reach

52%Of a strong GNN/ensemble baseline's remaining ranking error removed by TSGL on Minesweeper, without label inputs

The gist — one board, two questions

The paper opens with two Minesweeper clues. Step through the deduction, then ask about a different cell: the evidence never changes, but the computation the question needs does.

The question

Is the observed graph the right circuit for inference, or only the evidence about it?

Evidence, query, circuit

Message-passing GNNs use their input graph twice: once as the carrier of evidence, and once as the circuit that assembles an answer. Those are different objects. Observed edges describe relationships in the data; computational dependencies specify how to combine observations for a particular question.

The paper treats evidence, query and inference circuit as separately variable parts of a graph-learning problem. Reusing the observed graph as the circuit is still a legitimate choice — but it is an inductive bias, not a requirement, and it can be tested.

Keep the question; change the circuit

On Minesweeper, each revealed clue is compiled into a factor connected to the cells it constrains, and the same residual GCN update family runs over that circuit instead of the grid. With eight layers and parameters matched within 0.7%, AUC rises from .9021 to .9449, positive in all 25 split–seed pairs.

The control keeps everything but the alignment: variables are permuted within equal-degree classes, so the whole typed topology survives up to isomorphism while the circuit no longer matches the evidence. Performance falls to .5075 — near chance. Correct observations can even amplify a wrong circuit: under random membership, clamping correct training anchors costs 19.2 AUC points.

FigureAlignment, not graph shape alone, carries the gain

Left: the same GCN update family run over the observed adjacency or over compiled clue membership. Right: Minesweeper AUC at matched budget — evidence adjacency .9021, aligned circuit .9449, isomorphic scramble .5075, degree swaps .5058, and a sixteen-layer aligned reference at .9550. — One shared GCN update family, run on the observed adjacency or on compiled clue membership; tie-aware AUC over five splits and five seeds. The isomorphic scramble keeps the whole typed topology, features, architecture, depth and initialization, but changes which variables feed each factor.

Reach is not interaction

Holding the evidence, the class prior and the Bayes AUC fixed, the paper builds two queries on the same graph. For one, a linear readout is optimal. For the other, every linear score is capped 9.6 AUC points below Bayes at any radius: more propagation brings the second observation into reach, but cannot create the product the query needs. A product or MLP readout reaches Bayes the moment that observation arrives.

The theorem concerns fixed linear propagation and affine readouts. Nonlinear GNNs can express richer interactions; whether they learn the useful ones from the access they are given is the empirical question the rest of the paper takes up.

FigureSeeing the second check helps only a readout that can combine it

Two plots of population AUC against propagation radius. Leaf query: the linear readout stays at .800 at every radius while product and MLP readouts reach the Bayes .896 once the second check arrives at radius three. Center query, same evidence: the linear readout reaches .896 from radius one. — Fitted population AUC over propagation radius. Leaf query: the distant observation first arrives at radius three, after which product and MLP readouts attain Bayes while the linear score stays at its exact cap. Center query, same evidence: a linear readout suffices from radius one.

Two-System Graph Learning

TSGL puts the distinction to work. It preserves a capable GNN, learns explicit relationships within a declared library — affine, count and parity tables over relational memberships — and mixes them in only when a locked proposal earns acceptance on held-out nodes. A refusal returns the saved baseline exactly; the refused specialist is never executed.

FigurePreserved perception, supplemented inference

System diagram: raw evidence feeds a GNN perceiver, a typed-inference branch with learned affine, count and parity tables, and generic heads. A mixture of the preserved prediction and a specialist is selected on one validation split, then accepted or refused once on another; refusal keeps the baseline. — Candidates are fitted on training targets; the checkpoint, candidate and mixture weight are chosen on a selection set. The proposal is locked before a separate acceptance set decides once: deploy the mixture, or keep the baseline pointwise.

Results

Minesweeper. Against the best validation-selected GNN or GNN ensemble, strict TSGL raises AUC from 98.19 to 99.13 without label inputs, accepted on all nine splits — removing 52% of the remaining ranking error. Generic correction and an independent-model ensemble earn acceptance on none. With training-label anchors, in a separate matched panel, TSGL reaches 99.97.

Citation graphs. On arxiv-year, typed supplementation beats both generic correction and a second-GNN ensemble. The same recipe, unchanged and refitted on the training labels of a previously unused patent-citation graph, beats both controls in all three predeclared views, improving on the baseline by 1.85 points on average under strict access.

Reuse. A separate count-family compiler freezes a learned relationship and reuses it on 16 independent boards, answering new joint queries with no target fitting.

In these tests, every accepted deployment improved the test score, and every refusal returned the baseline exactly.

Where it does not win

On three fresh social graphs — Penn94, Twitch-gamers and Genius — the typed library beat the stronger generic or ensemble control in none of 18 evaluations. The full system deployed those controls instead: 14 improvements, 4 exact baseline returns, no observed regression. That is a boundary of the tested library and recipe, reported in full, not evidence that those graphs hold no useful relationships.

Selection costs something even when nothing is deployed: on Genius, the typed candidates took 28.75 of 33.93 summed fitting hours and were selected in none of six evaluations.

Reproducibility

Preregistered endpoints and selection rules for the interventions and fresh-data tests
Frozen predictions, scored misses and post-hoc controls kept distinct
Negative results and the social-graph boundary reported in full
Code, configurations, masks and retained predictions available on request

Citation

Graphs Are Evidence, Not Circuits

Aditya Gaur, Charu Sharma

Preprint2026

PDF

What comes next

Learning circuit membership and building query-specific circuits automatically are the next architectural steps. The paper shows why those objects matter; it does not yet supply an unrestricted compiler.

Next project All research

Research

Five areas

How I approach research

Publications

Public preprints

Ongoing research and manuscripts

Engineering projects

Experience

About

What I am working on now

Third-person bio

Closed-Form Node Classification with Exact Graph Unlearning

Decision-Grounded Representations for Predictive World Models

Graphs Are Evidence, Not Circuits

Human-Grounded Audits of Emergent Misalignment

Vyasa: Autonomous Research Harness