Projects
Research, projects, and writing.
-
AutoHarnessBenchA long-horizon SWE & self-improvement benchmark evaluating automatic harness optimization.
-
Privileged Self-DistillationA post-training method that distills training signal from failed rollouts corrected with privileged hints.
-
Meta-Reward: Reward Modeling as Harness OptimizationFraming reward modeling as a harness optimization problem.
-
→
67%→
80%
87%Meta-Agent: Continual Learning for AgentsAn open-source framework for agent continual learning.
-
-
-
Goldfish: Vision-Language Understanding of Arbitrarily Long VideosA SOTA vision-language model for question answering on hour-long videos. Harvard, contributor.
-
MiniGPT4-Video: Advancing Multimodal LLMs for Video UnderstandingA multimodal LLM that interleaves visual and textual tokens for video understanding. Harvard, contributor.
-
SlowFormer: Adversarial Robustness for Efficient Vision TransformersA universal patch attack that maximizes the compute cost of efficient vision transformers. UC Davis, with K L Navaneet et al.
-
Do Function Vectors Factor Task and Distribution?
An interpretability study of how LLMs represent tasks during in-context learning. Harvard & MIT, with Jacob Andreas.
-
Can LLMs Understand Quantifiers Like Humans?Uses parallel Rational Speech Act models to compare human and GPT-3.5/GPT-4 quantifier choices across two context conditions. Harvard & MIT, with guidance from Ced Zhang and Joshua Tenenbaum.
-
Continual Self-Supervised Learning with Knowledge DistillationA video embedding model for Twitch that learns new content without forgetting the old. Amazon Science, with Xiangbo Li and Saad Ali.
-
Mitigating Negative Transfer in Multi-Task Learning with EMA Loss Weighting
Student abstract on EMA loss weighting strategies that reduce negative transfer in multi-task learning. Stanford.
-
Deep Learning-Based Autism Spectrum Disorder Detection Using Emotion Features From Video
Detecting autism spectrum disorder from emotion features in video recordings. Stanford.
-
TikTok for Good: Creating a Diverse Emotion Expression DatabaseA diverse emotion expression database collected from TikTok. Stanford, contributor.
-
