Title: Beyond capability: safe & reliable foundation models at scale

Date: Tuesday, November 17, 2026

Time: 12:00 PM to 2:00 PM Eastern Time (US)

Join in Person: CODA 114 (first floor conference room; just walk in, no special access needed)

Join Virtually: Zoom

ShengYun Anthony Peng

shengyun-peng.github.io 

CS Ph.D. Candidate

School of Computational Science and Engineering

College of Computing

Georgia Institute of Technology

Committee:

Dr. Duen Horng (Polo) Chau — Advisor and Committee Chair, School of Computational Science and Engineering, Georgia Institute of Technology

Dr. Wei Xu — School of Computer Science, Georgia Institute of Technology

Dr. Alan Ritter — School of Interactive Computing, Georgia Institute of Technology

Dr. Jianfeng Chi — Member of Technical Staff, OpenAI

Dr. Pin-Yu Chen — Principal Research Scientist, IBM Research

Abstract:

Foundation models achieve remarkable benchmark performance, but growing capabilities do not guarantee safety, reliability, or efficiency in real-world deployment. My dissertation empowers foundation models to operate safely and reliably at scale through three complementary thrusts:

(1) Aligning for Safety. We discover a universal safety basin phenomenon across popular open-source LLMs: a local parameter region preserving safety. This discovery, featured in the International AI Safety Report, reveals why finetuning erodes alignment and inspires STAR-DSS, our novel dynamic safety shaping method to mitigate this erosion. For safer reasoning, we introduce RECAP, a principled reinforcement learning (RL) method for post-training that teaches models to recover from flawed thinking.

(2) Designing for Reliability. Safe behavior must be paired with reliable predictions. In Robust Principles, we synthesize a suite of three generalizable robust architectural design principles for CNNs, achieving state-of-the-art (SOTA) adversarial robustness on RobustBench. This work earned the Best Poster (Paper) Award at BMVC and guided DARPA’s research on adversarially robust architectures. These insights also enable accurate, efficient business table recognition with lightweight encoders.

(3) Scaling Model Capabilities. Complementing architectural design, we unlock accurate table recognition with simpler encoders through self-supervised pretraining. Our UniTable training framework unifies both the training paradigm and training objective of table recognition, achieving SOTA performance. UniTable (530+ GitHub stars) powers table extraction in Open Parse (3.2K GitHub stars) and is used by startups for retrieval-augmented generation (RAG). This research has been selected for oral presentations and presented in invited talks. To scale visual understanding efficiently, we establish a principled framework revealing optimal inference compute allocations for video VLMs.

My dissertation research significantly impacts academia, industry, and society, advancing foundation models through collaborations with NVIDIA, Meta, IBM, Intel, and ADP. It has contributed to the AI foundations of industry research grants totaling over $1.7 million.