←
Ge Gao26 posts
Ge Gao avatar

Ge Gao

@GeGaoCS

PostDoc at Stanford | PhD in Computer Science at NCSU l ML in Healthcare & Education

◷ Joined January 2021
Alex Nam avatar
⟳ Ge Gao retweeted
Alex Nam@hjalexnam·Oct 23

Evaluating tool-use capabilities is expensive and domain-dependent. Once existing test data become saturated, expert humans need to design new tasks. Alternatively, LLMs can auto-generate challenging tasks from existing toolsets and task primitives through evolutionary search using multiple agents. With @MichaelD1729, @natashajaques, @minsuk_chang. Paper: arxiv.org/abs/2610.27307 for how to auto-generate tasks from existing benchmarks and use them for failure diagnosis

Post attachment
Anikait Singh avatar
⟳ Ge Gao retweeted
Anikait Singh@Anikait_Singh_·Oct 4

Will be in SF for #colm2026! I’m presenting GIANTS and Librarian Agents (more to come soon 👀). Interested in test-time adaptation, learning in open-ended environments, and self-improvement. Also, would be on the job market this year. Feel free to reach out (dms open)!

🚨🚨New Paper: Generative Insight Anticipation from Scientific Literature Human scientists achieve breakthroughs by "standing on the shoulders of giants," synthesizing profound insights from disparate sources. While LMs show promise in scientific discovery, they currently struggle to reliably generate ideas of true impact, diversity, and feasibility. Introducing "insight anticipation," a novel task challenging LLMs to generate a downstream paper's core scientific contribution directly from its foundational prior works. We also introduce GIANTS-4B, an RL-trained model that significantly outperforms frontier models at this task! 🧵⬇️

Post attachment
Aishwarya Mandyam avatar
⟳ Ge Gao retweeted
Aishwarya Mandyam@Aishwarya_R_M·Jul 7

✨New work on off-policy evaluation with synthetic trajectories: PERRY (Policy Evaluation with Confidence Intervals using Auxiliary Data), accepted to TMLR!✨ TL;DR: we incorporate fully synthetic trajectories into off-policy evaluation (OPE) and develop two methods to learn valid confidence intervals over the policy value. Paper: arxiv.org/abs/2507.20068 Authors: @Aishwarya_R_M , @jasonjmeng , @GeGaoCS , @JiankaiSun , @MacSchwager, @BeEngelhardt , @EmmaBrunskill

Alex Nam avatar
⟳ Ge Gao retweeted
Alex Nam@hjalexnam·Jun 3

Excited to release our latest paper, Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data (in short, MIPO), which has been accepted to ICML 2026! tl;dr Can we build a data augmentation method analogous to translation and rotation in computer vision for training LLMs without external verifiers or human supervision? We show that generating chosen responses from correct prompts and rejected responses from misspecified (random!) prompts can provide promising training signals. w/ @AimeeLiS @natashajaques paper: arxiv.org/abs/2603.19294 code: github.com/nam630/mutual_inf…

Post attachment
Perry Dong avatar
⟳ Ge Gao retweeted
Perry Dong@perryadong·May 27

Introducing EXPO-FT – Efficient, Reliable & Open-Source VLA Finetuning! EXPO-FT unlocks π0.5 for challenging manipulation tasks: Routing string lights & inserting the power connector to illuminate them Striking pool ball into pocket Inserting flower into wine bottle (1/5)

1:36
Anikait Singh avatar
⟳ Ge Gao retweeted
Anikait Singh@Anikait_Singh_·Apr 21

Heading to ICLR today and excited to be presenting two papers! 🇧🇷 If you’re interested in self-improving agents, lifelong learning, exploration, or just want to connect, I’d love to chat. DMs are open! 👇 I’ll be presenting: 📝 RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems 📅 04/23, 3:15 PM | Pavilion 3, Poster #1413 📝 FSPO: Few-Shot Optimization of Synthetic Preferences Personalizes to Real Users 📅 04/23, 10:30 AM | Pavilion 3, Poster #1509

Post attachment
Ge Gao avatar
Ge Gao@GeGaoCS·Apr 19

Excited to share our new work on anticipating scientific insights! Given prior papers, can a model predict the core idea of the next paper? With GIANTS-4B, we start to model the “conceptual leap” behind discovery. Amazing work led by @JoyHeYueya @Anikait_Singh_

Scientists often make breakthroughs by synthesizing ideas across papers. In our new paper, we ask whether a language model can anticipate this process: given two parent papers, can it generate the core insight of a future paper built on them? 🧵⬇️

Joy He-Yueya avatar
⟳ Ge Gao retweeted
Joy He-Yueya@JoyHeYueya·Apr 17

Scientists often make breakthroughs by synthesizing ideas across papers. In our new paper, we ask whether a language model can anticipate this process: given two parent papers, can it generate the core insight of a future paper built on them? 🧵⬇️

Ge Gao avatar
Ge Gao@GeGaoCS·6 Dec 2025

Please stop by if you are interested! I’d love to chat!

And our 3rd workshop poster is on off policy RL selection! ARLET RL Workshop (Rm 31abc) 3:30 p.m Behavior-Aware Off-Policy Selection in High-Stake Human-Centric Environments w/G.Gao, Aishwarya Mandyam, Joy He-Yueuya & Min Chi #neurips2025

Emma Brunskill avatar
⟳ Ge Gao retweeted
Emma Brunskill@EmmaBrunskill·6 Dec 2025

And our 3rd workshop poster is on off policy RL selection! ARLET RL Workshop (Rm 31abc) 3:30 p.m Behavior-Aware Off-Policy Selection in High-Stake Human-Centric Environments w/G.Gao, Aishwarya Mandyam, Joy He-Yueuya & Min Chi #neurips2025

Andrea Zanette avatar
⟳ Ge Gao retweeted
Andrea Zanette@Zanette_ai·24 Nov 2025

I’m recruiting several PhD students at Carnegie Mellon University! If you’re interested in LLM reasoning, agents, or diffusion language models, consider applying to the CMU ECE PhD program. Applications are due Dec 15. ece.cmu.edu/admissions/gradu…

Ge Gao avatar
Ge Gao@GeGaoCS·30 Oct 2024

(3/n) Tested in intelligent tutoring and sepsis treatment, FPS can improve outcomes by up to 208% in education in a retrospective study and outperformed conventional methods in healthcare. FPS is a leap in adapting RL policies to real-world, personalized applications!

Ge Gao avatar
Ge Gao@GeGaoCS·30 Oct 2024

(2/n) Our recent work published in #neurips2024, "Off-Policy Selection for Initiating Human-Centric Experimental Design", introduces the First-Glance Off-Policy Selection (FPS) framework, which segments participants based on traits to assign intervention policies.