Two compute-skipping policies for diffusion LLMs (Fast-dLLM v2) that reuse stable hidden states across denoising steps to cut FLOPs: a layer-level cache-reuse policy and a stability-aware token-level policy that recomputes only the least-similar tokens.
Report (PDF)
ACM CCS ‘25 (AR: 13.9%) — Distinguished Paper Award (1.39%)
wiretap.fail
This project explores efficient transformer attention. I implemented a PQ‑tree‑based token subsampler to reduce complexity from O(L2) to O(L log L), and reproduced some of previously proposed attention approximation/subsampling methods.
We produce the first English Louvre Dataset and investigate how effectively NLP models understand art-related text by curating and annotating a cultural-heritage corpus (Van Dyke’s painting textbook, translated Louvre records, etc.), defining a BRAT schema (entities, relations, events), and fine‑tuning DistilBERT for NER. Results show strong performance and qualitative reasoning on museum-style queries.
Paper (PDF)
From time-series air measurements in classroom environments, this project extracts features and predicts occupancy and related information using machine learning techniques.
SGX-Step is a practical attack framework for precise enclave execution control — I was fortunate enough to maintain the codebase in 2023 for newer Linux kernel patches under the supervision of Prof. Jo Van Bulck.