Ximing Lu (@GXiming) / X

Ximing Lu

202 posts

Ximing Lu

@GXiming

PhD @uwcse @uwnlp.

Santa Clara, CA

gloriaximinglu.github.io

Joined February 2018

Pinned
Ximing Lu
@GXiming
Feb 3
There’s growing excitement around scaling up RLVR to get continuous gains with more compute. But in practice, improvements saturate on finite training data. 😱 Introducing Golden Goose 🦢✨, a simple trick to synthesize unlimited RLVR tasks 😎 from unverifiable internet text. 🌐
Ximing Lu
@GXiming
Apr 22, 2025
With the rise of R1, search seems out of fashion? We prove the opposite! 😎 Introducing Retro-Search 🌈: an MCTS-inspired search algorithm that RETROspectively revises R1’s reasoning traces to synthesize untaken, new reasoning paths that are better 💡, yet shorter in length ⚡️.
Ximing Lu
@GXiming
Aug 12, 2025
🚀 How far can RL scaling take LLMs? Drop ProRLv2! 🔥We keep expanding LLM’s reasoning boundaries through 3,000+ RL steps over 5 domains and set a new state-of-the-art ✨ among 1.5B reasoning models. 🔗Full blog: research.nvidia.com/labs/lpr/prorl… 🤗Open model: huggingface.co/nvidia/Nemotro…
This post is unavailable.
Ximing Lu
@GXiming
Nov 22, 2024
Are LLMs 🤖 as creative as humans 👩‍🎓? Not quite! Introducing CREATIVITY INDEX: a metric that quantifies the linguistic creativity of a text by reconstructing it from existing text snippets on the web. Spoiler: professional human writers like Hemingway are still far more creative
Ximing Lu
@GXiming
Jun 2, 2025
What happens when you ✨scale up RL✨? In our new work, Prolonged RL, we significantly scale RL training to >2k steps and >130k problems—and observe exciting, non-saturating gains as we spend more compute 🚀.
Andrew Zhao
@_AndrewZhao
Jun 2, 2025
RL scaling is here arxiv.org/pdf/2505.24864
Ximing Lu
@GXiming
Jan 22, 2024
If you're interested in this direction, check out our paper Inference-Time Policy Adapters (IPA🍺). IPA guides a frozen LLM such as GPT-3 during decoding time through a lightweight policy adapter trained to optimize an arbitrary user objective with RL.
Sebastian Raschka
@rasbt
Jan 18, 2024
There's a new promising method for finetuning LLMs without modifying their weights called proxy-tuning (by Liu et al. arxiv.org/abs/2401.08565). How does it work? It's a simple decoding-time method where you modify the logits of the target LLM. In particular, you compute the
Ximing Lu
@GXiming
Dec 9, 2023
I'm not able to travel to #EMNLP2023 due to visa issues, but my amazing co-author @faeze_brh is there and will present our work tomorrow Dec 10th at 9:00am-10:30am! ✨ Come and check it out! I'll attend #NeurIPS2023 next week. Feel free to DM me if you wanna chat! 🍹
Faeze Brahman
@faeze_brh
Dec 9, 2023
Can we efficiently tailor LLMs towards arbitrary user objectives without fine-tuning them?! The answer is ✅yes with IPA which combines RL and inference-time techniques! Come to our poster “Inference-time Policy Adapters” tomorrow Dec 10th at 9:00am-10:30am! #emnlp2023
Ximing Lu
@GXiming
Dec 23, 2024
Check out our latest work "AI as Humanity's Salieri," featured in ✨News from Science ✨. Dive into how we quantify linguistic creativity and explore: Are LLMs 🤖 as creative as humans 👩‍🎓? Paper link: arxiv.org/abs/2410.04265
Matthew Hutson
@SilverJacket
Dec 23, 2024
New research: AI writing is improving, but it still can’t match human creativity. My latest for @ScienceMagazine @NewsfromScience, on work by @GXiming of @UW et al. Thanks also @mircomusolesi and @VioletNPeng. Story link in reply.
Ximing Lu
@GXiming
Jun 7, 2021
Check out our new preprint:🍷𝓜𝓔𝓡𝓛𝓞𝓣🍷! This is joint work with my awesome co-authors: @rown🍷, @jmhessel🍷, @easycommit, @jae_sung_park96, @JizeCao, Ali Farhadi & @YejinChoinka. I'm especially grateful to work with my amazing joint first-authors @rown🍷and @jmhessel🍷!
Jack Hessel
@jmhessel
Jun 7, 2021
🍷Super excited about our new preprint!🍷 𝓜𝓔𝓡𝓛𝓞𝓣: Multimodal Script Knowledge Models! arxiv.org/abs/2106.02636 rowanzellers.com/merlot/ TL;DR: By pretraining on 6M youtube videos, we transfer with SoTA performance on 10+ tasks (e.g. Video QA) that require temporal reasoning
Ximing Lu
@GXiming
Nov 22, 2024
Replying to @GXiming
Moreover, we found that RLHF dramatically reduces the CREATIVITY INDEX of LLMs, by an average of 30.1%. This reduction is more significant at the verbatim level than the semantic level, indicating that LLMs may have converged to certain linguistic style preferred by humans during
Ximing Lu
@GXiming
Dec 22, 2024
Excited to talk about our latest work, "AI as Humanity's Salieri," at Fireside Chat today at 7 PM PST! 🔥 app.ploutos.dev/streams/innoce…
Ximing Lu
@GXiming
Nov 22, 2024
Replying to @GXiming
TLDR: We found the seemingly remarkable creativity of LLMs 🤖can be attributable in large part to the creativity of human-written texts on the web. In contrast, works by distinguished human authors 👩‍🎓cannot be easily replicated by merely assembling snippets from other works.
Ximing Lu
@GXiming
Nov 22, 2024
Replying to @GXiming
We found CREATIVITY INDEX of human authors—specifically professional writers and historical figures—is on average 66.2% higher than that of LLMs. This gap is consistent across various domains—novel snippets, modern poems, and speech transcripts—at both verbatim and semantic
Ximing Lu
@GXiming
Aug 15, 2025
The official blog of ProRLv2 is now live. 🔥 Check it out and see how we scale RL! 🚀 developer.nvidia.com/blog/scaling-l…
Ximing Lu
@GXiming
Aug 12, 2025
🚀 How far can RL scaling take LLMs? Drop ProRLv2! 🔥We keep expanding LLM’s reasoning boundaries through 3,000+ RL steps over 5 domains and set a new state-of-the-art ✨ among 1.5B reasoning models. 🔗Full blog: research.nvidia.com/labs/lpr/prorl… 🤗Open model: huggingface.co/nvidia/Nemotro…