Research

Analog compute-in-memory can make matrix operations efficient, but scaling to large language models introduces communication bottlenecks across many interconnected dies. I developed Isobar, a simulation framework that connects physical weight mapping with inference performance by modeling mapping-dependent execution and contention on shared communication resources. The framework reveals when mapping choices determine system bottlenecks and combines hierarchical mapping search with fast performance evaluation to explore deployments at thousands-of-dies scale.

Can pretrained attention blocks be replaced with simpler sequential modules while preserving model accuracy? My work explores BiLSTM- and Mamba-based attention replacement, using block-wise knowledge distillation to adapt vision transformers for compute-in-memory inference. Building on FAR, From Sparsity to Simplicity investigates why some attention blocks are easier to replace and uses explicit teacher sparsification to improve distillation and replacement fidelity.

Errors in an LLM agent’s actions can accumulate across interactions, making long-horizon tasks sensitive to both model compression and perturbations. We investigate perturbation-robust reinforcement learning to improve policy stability and trajectory-anchored recovery to preserve agent capabilities after structured pruning. By combining teacher-generated interaction histories with distillation on student-generated responses, this work addresses cross-turn error accumulation and evaluates agent robustness and efficiency on ALFWorld and WebShop.