←
Sherry Yang294 posts
Sherry Yang avatar

Sherry Yang

@sherryyangML

Assistant Professor @NYU_Courant; Staff Research Scientist @GoogleDeepMind. Previously Post-Doc @Stanford, PhD @UCBerkeley, M.Eng. / B.S. @MIT.

◷ Joined September 2015▧ 98 Photos and videos
Pika avatar
⟳ Sherry Yang retweeted
Pika@pika_labs·Oct 29

Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board. It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds. So, yeah. It’s kind of a flex 💅

1:03
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 22

This complements our Reward-Rate Policy Gradient work on MLE agents: RPG learns to maximize reward per unit time. MLE-World reuses experience to make experiment feedback cheaper. Two complementary directions for making better use of a limited research budget. x.com/sherryyangML/status/21…

Excited to share Reward-Rate Policy Gradient: RL that accounts for execution time. For autoresearch and ML engineering agents, actions can take seconds or hours. A fixed-step MDP misses this distinction. We optimize MLE agents' reward per unit time in a semi-MDP. arxiv.org/pdf/2609.36393 muhang-tian.com/reward-rate See details below:

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 22

Every experiment leaves behind useful experience, even when it fails. MLE-World turns those execution traces into a predictive model: code + task context → errors, training logs, and scores. Agents can use that feedback to revise ideas, then run real experiments when predictions are uncertain.

Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 22

On 17 MLE-bench Lite tasks, MLE-World achieves 19% higher mean leaderboard percentile than baseline using less than half the real-execution time. Mean percentile: 45.4 → 54.1 Execution time: 9.94 → 4.77 hours/trial

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 22

World models go beyond robotics. Autoresearch and agentic ML engineering face a familiar bottleneck: expensive experiments. Introducing MLE-World: learn from past experiments to predict new outcomes before running them. Paper, model, data, code: mle-world.github.io/

Post attachment

🚀 Toward RSI: AI agents automate ML research, but each step trains and evaluates a model, limiting test-time scaling. Our new paper, MLE-World, asks: can agents explore more ideas by predicting experiment outcomes before running them? 🧵⬇️

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 19

Why should agentic robots rediscover what humans already know? Excited to share our new work, MoCaP (mobile Code-As-Policy), which turns passive human videos of everyday life into a head start for agentic robots: where things are and how to approach them. Human experience → less robot trial and error.

Coding agents are emerging as general robot policies, but they are still inefficient (e.g., requiring long real-world exploration), especially in mobile manipulation. Humans become efficient by watching others. Can coding agents learn from passive human videos, just like we do? Introducing MoCaP: Mobile Code-as-Policies for Zero-Shot Mobile Manipulation from Passive Human Videos ↓🧵

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 15

This work was inspired by the paper "Between MDPs and Semi-MDPs" by @RichardSSutton, Doina Precup, and Satinder Singh from nearly three decades ago. Applying conceptual frameworks such as semi-MDPs in practical scenarios of modern agentic RL for LLMs is a reminder of how important the foundations of RL remain.

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 14

I was most excited about the NanoGPT results: reward-rate RL achieves lower pretraining loss faster than vanilla RL. With growing interest in autoresearch and MLE agents from @karpathy @quocleix @JeffDean, time-aware RL could be an important ingredient. Choosing the next experiment means deciding how to spend a limited research budget. Agents should learn to weigh an experiment’s potential payoff against the time it takes to run. x.com/muhang_tian/status/210…

6/7 On the modded NanoGPT speedrun, RPG beats vanilla RL by 85.7% at the final budget.

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Oct 14

Excited to share Reward-Rate Policy Gradient: RL that accounts for execution time. For autoresearch and ML engineering agents, actions can take seconds or hours. A fixed-step MDP misses this distinction. We optimize MLE agents' reward per unit time in a semi-MDP. arxiv.org/pdf/2609.36393 muhang-tian.com/reward-rate See details below:

Post attachment

1/7 Excited to share our new work on agentic RL! 🛠️ Some tasks in agentic RL involve actions that incur costs. One example is machine learning engineering (MLE) agents, where a proposed script might run for minutes or hours. We propose Reward-rate Policy Gradient (RPG) using a Semi-Markov Decision Process formulation that optimizes reward per unit time. 🌐muhang-tian.com/reward-rate/

Yann LeCun avatar
⟳ Sherry Yang retweeted
Yann LeCun@ylecun·Oct 8

Au contraire. C'est une nouvelle ère qui s'ouvre pour les mathématiques. Une ère où la démonstration formelle est largement automatisée et où l'accent sera reporté sur le développement de nouveaux concepts, nouvelles abstractions, nouvelles définitions, et nouvelles conjectures. L'invention du bateau a réduit l'importance de la nage, mais a permis la découverte de nouvelles terres.

Sherry Yang avatar
Sherry Yang@sherryyangML·Sep 18

I've been interested in the intersection of robot learning × AI for science for a while. WetRobo makes me particularly excited about coding agents as an interface between foundation models and scientific instruments. Moving from agents that reason about experiments to agents that can physically carry them out and adapt along the way is an extremely exciting future direction of robot learning × AI for science.

Sherry Yang avatar
Sherry Yang@sherryyangML·Sep 18

One aspect I found particularly interesting is how adaptation happens through physical interaction. The agent tries an action, observes the result, and modifies its program – for example by adding calibration, grasp checks, perception, or motion logic – before trying again. The adaptation happens in code rather than through model retraining.

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Sep 18

We also looked at what happens when the physical setup changes across labs. A VLA trained on demonstrations from Lab X achieved: - Lab X → 10/10 - Lab Y → 0/10 After on-site adaptation, the coding-agent system achieved: - Lab X → 10/10 - Lab Y → 10/10 Same task, different lab setup.

Post attachment
Sherry Yang avatar
Sherry Yang@sherryyangML·Sep 18

During adaptation, the agent can also bring in existing tools as needed: - SAM for segmentation - OpenCV for perception - NumPy for geometry - MuJoCo for collision checking - RGB-D for 3D reasoning So the resulting solution is often a composition of several existing tools.

Post attachmentPost attachment