Suryaansh Jain

Introduces SimLoss, a reference-free objective that trains vision-language models for single-pass fine-grained captioning by aligning their hidden states with a frozen image embedding, without needing fine-grained caption data. Its full fine-tuning variant achieves the highest precision and nearly matches the F1 of a multi-stage pipeline while running about 20× faster.

Introduces reward-based φ-function state abstractions for the STAR estimator, enabling low-variance, low-bias off-policy evaluation that outperforms importance-sampling and model-based baselines across RL environments. Manuscript under submission; available upon request.

Shows that LLM judges are strongly biased toward accepting outputs, with a 96% true positive rate but under 25% true negative rate. Introduces a minority-veto ensemble that mitigates this bias, and a regression-based calibration that reduces the maximum absolute error to 1.2% on 366 high-school Python programs, a 2× improvement over the best ensemble of 14 LLMs.

Provides a bound on the cop number of graphs in terms of their 2-component order connectivity.

Shows that FO model checking on induced subgraphs of hypercubes is as hard as on general graphs, so fixed-parameter tractability cannot be extended to all FO-definable problems on this class.