I'm a 3rd year PhD student at Princeton University advised by Jia Deng, and an Anthropic Fellow. I received my B.S in Computer Science from Columbia University in 2024, where I was fortunate enough to do research with Professors Mark Santolucito, Kathleen McKeown, and Baishakhi Ray.
I love hiking, horror movies, peanut butter, and reading. Good reasons to contact me: research, book recs, hiking recs.
I study perception, memory, and reasoning in foundation models, with a particular interest in the visual and embodied capabilities required for robust interaction with the physical world. More broadly, I care about how these systems perceive, remember, and act in non-textual environments, including robotic settings.
Shmuel Berman, Jia Deng
ECCBench evaluates memory beyond raw accuracy along three axes: efficiency (the FLOPs needed to answer from memory), compression (whether compressible inputs are remembered better), and calibration (whether a system abstains when it is uncertain). Pretrained VLMs compress over text but not video, and are poorly calibrated on both.
Shmuel Berman, Michael Ilie, Jia Deng, Daniel Freeman
An evaluation of whether frontier language models can perceive a scene, track a robot’s state, and issue actions that reliably change the physical world, across classic control tasks, simulated quadrupeds and humanoids, a robotic arm, and a real Unitree Go2.
Shmuel Berman, Kathleen McKeown, Baishakhi Ray
A multi-agent framework that combines language models with a theorem prover to translate natural-language clues into structured constraints and solve zebra puzzles more reliably.
