Working at the edge of generative models and visual understanding.

Accepted at the ICLR 2026 Workshop on Deep Learning for Theory (DeLTa).

A reproducibility study and extension on the ineffectiveness of contrastive decoding at mitigating object hallucinations in multimodal LLMs.

A diffusion-distillation framework that supervises the student on the teacher's full denoising trajectory via K-fold branched output channels (K·C instead of C), recovering intermediate denoising steps that standard distillation discards. Cut FID against Progressive Distillation on CIFAR-10 (39.66 → 20.81 at 128 steps) and against SFD on ImageNet-64 at NFE 2 (10.25 → 9.57), for a ~0.01% FLOP increase and identical inference cost.

Led a reproducibility study stress-testing contrastive decoding (VCD, ICD, SID) across LLaVA-1.5 (7B/13B) and Qwen2.5-VL on POPE, MME, and CHAIR. Showed CD is statistically indistinguishable from spurious, non-visual controls on precision and hallucination rate, isolating its effect to a single recall shift (+12.21pp, p<0.001). A logit-lens analysis localized the internal layers where the intervention breaks down — establishing the gains as a distributional artifact, not corrected perception.

Probed the modality gap in LLaVA-1.5-7B by extracting text and image refusal directions from residual-stream activations and comparing them geometrically. Their cosine similarity never exceeds 0.3 while the text-side norm consistently dominates the image side — identifying the modality gap as a primary driver of weak vision-side refusal.

An end-to-end, resumable pipeline turning raw ChartQA samples into a supervised fine-tuning corpus of structured visual chain-of-thought traces (summary → extraction → computation → answer). Sampled 3 traces per item from a frontier VLM teacher, then filtered through five staged gates — schema, answer correctness, numeric grounding, arithmetic re-verification, and 2-of-3 self-consistency — keeping the highest-grounded survivor at 95% yield.