research
Research positions, ongoing work, and the software I have built.
My work sits at the intersection of language model evaluation, efficient deep learning, and interpretability — three angles on the same question of whether a reported number reflects a capability you can depend on.
Research positions
Extending Asymptotic Probability Decoding (Chang et al., 2024) from extrapolation across model scale to extrapolation across context length.
What the work involves
- Removing the method’s reliance on a family of same-recipe language models at multiple sizes — the requirement that makes it impractical for most model families in the wild.
- Auditing the training pipeline against the original formulation to isolate sources of training instability: the curve-fitting objective, top-_k_ candidate selection, and the numerical conditioning of the parametric decay fit.
Three lines of work across pruning, evaluation, and benchmark construction.
Project detail
- Sparse subnetworks in hyperspectral image classification under the Lottery Ticket Hypothesis. Iterative magnitude pruning across 3D CNN, hybrid 3D–2D CNN, and Mamba backbones, evaluated on three benchmark scenes.
- Evaluation protocols for post-training language model assessment, separating instruction-following compliance from answer correctness across arithmetic, logical, and code-generation benchmarks.
- Benchmark construction and automated evaluation pipelines for formal-language tasks spanning the Chomsky hierarchy.
The compliance–correctness split
Score a small language model on a reasoning task and a single accuracy number has to answer two questions at once: did the model reason correctly, and did it return the answer in the format the grader demanded? When a model reasons well but ignores the format, strict accuracy marks it wrong — indistinguishable from a model that genuinely failed. Drag the grader’s strictness and watch the two views disagree.
Strict accuracy reads 40%, but the model actually reasons correctly 70% of the time. The gap is six responses that were right on the merits and marked wrong only for ignoring the format.
Selected software
A multi-role platform — farmers, buyers, coordinators, transporters — implementing a staged workflow from harvest listing through demand matching to fulfillment, with role-based access enforced by PostgreSQL row-level security. Integrates the 1,051-record Bhutanese administrative hierarchy (Dzongkhag–Gewog–Chiwog) into a cascading selector for standardized geographic entry.
Expense classification, budget forecasting, and spending-pattern analysis over user transaction histories.