PhD student @LTIatCMU, working with @gneubig on NLP || intern @AIatMeta @GoogleDeepMind || from ๐จ๐ฆ || on the job market
Pittsburgh, PA
Joined November 2021
- Tweets475
- Following595
- Followers3.3K
- Likes1.8K
Emmy Liu retweeted
Foundation models are trained in stages: pretraining, midtraining, and post-training. But is this separation fundamental?
We introduce unified training: a single algorithm for the full pipeline. This gives
โ๏ธ Simpler infrastructure
๐ฌ A framework for better science
(1/n)
Emmy Liu retweeted
Iโm on the job market, and I keep getting asked: โWhat do you want to work on?โ
I took some time to reflect on the art of problem selection in AI research and how that's evolved for me personally since 2017 :')
I'm no expert but if any of this resonates, lmk!
simran-khanuja.github.io/bloโฆ
Emmy Liu retweeted
๐จ New paper w/ @GoogleDeepMind ๐จ #EMNLP2026
What are AI detectors looking at when detecting generated text?
Introducing Detector Scope: a way to convert black box AI detectors into fully auditable models using sparse autoencoders.
Paper: aclanthology.org/2026.emnlp-โฆ
Thread ๐
indeed, as an aspiring "LLM raiser" there are a lot of issues to consider! Some behaviours that appear later in training may also be encouraged by earlier stages but be invisible until later...
Raising LLMs well as good kids is @_emliu โs thesis direction! Her work already shows LLMs learn stuff in a consistent order for pretraining. Exploring and controlling this across the training stack would be really cool!
Emmy Liu retweeted
Data science agents lag behind top humans on Kaggle, often due to subtle mistakes.
๐ Introducing AdaPRA, a process-reward agent that uses adaptive rubrics and writes code to audit a data science agent's output as it works.๐งต
Emmy Liu retweeted
Your agent solved the task at xhigh. Would low have solved it too? You canโt tell without rerunning it, and you still pay for every token.
We tested models x reasoning effort level x benchmarks. On tasks low effort can already solve, top effort costs up to 42ร more #LLM #AI ๐งต๐
Emmy Liu retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Emmy Liu retweeted
I built an interactive demo of the Fork-and-Flush idea basins using Naive.ai ๐งต
You simply input a natural language query, and the tool scrapes hundreds of relevant papers from arXiv, extracts a spec per paper, and builds an idea space visualization
Fork-and-Flush: help your agents escape Idea Basins!!
I love thinking of AutoResearch as Idea Descent: a metric of interest goes down, as a function of ideas/experiments, where the agents get pulled towards one idea or another. But we've seen very often an agent develop this weird tunnel vision, and honing in just one type of "idea", kinda like GD stuck in a local minimum.
We studied this and seems we can fix it with Fork-and-Flush, a super simple and kinda cute idea. Every so often, fork the agent into T copies in the same workspace but empty the context, let them run for a bit, and continue from the best one.
If you like the GD analogy, it's basically basin hopping en.wikipedia.org/wiki/Basin-โฆ.
BTW we found a silly way of also plotting "idea basins": an LLM rewrites every experiment as a spec and answers a few thousand binary queries "is Experiment A closer to B or to C?". Then we embed this NXN matrix into 2D and indeed see in many cases that each run saturates into a small region and keeps polishing it.
We call these idea basins!
At equal compute Fork and Flush beats single runs and best-of-N on 12/13 long-horizon tasks (some 4+ days), and in my own research projects it's been working really well!
BTW we only use it at test time, but it could also be part of training or RL.
Led by @jackcai1206 @christos_ziakas during Jack's internships with us, along with @vkontonis @ShivamGarg91462 @Tea_Pearce Sid Sen and Akshay Krishnamurthy.
Emmy Liu retweeted
The UT comp ling faculty (@jessyjli, @kanishkamisra, @kmahowald & myself) are looking for PhD students to join us! sites.utexas.edu/compling/ ๐ค
We are an interdisciplinary group with interests across linguistics, cogsci and NLP. Apply by 12/1: liberalarts.utexas.edu/linguโฆ
Emmy Liu retweeted
my iclr 27 papers to review are some of the strongest evidence against rsi I've seen to date. just the most dreadful claudeslop.
It was great to speak with you at COLM! ๐ฆ Looking forward to the full episode release as well (Besides this paper, I also talk a bit about my experience in the PhD and what tools I use for research!)
Congrats to @_emliu @KaiserWhoLearns @millicent_li @wordscompute @lltjuatja @JentseHuang @gneubig on winning the #COLM2026 Outstanding Paper Award!
#SOTATV just interviewed @_emliu @COLM_conf on ๐๐ก๐ ๐๐ฆ๐ฉ๐ฅ๐ข๐๐ข๐ญ ๐๐ฎ๐ซ๐ซ๐ข๐๐ฎ๐ฅ๐ฎ๐ฆ ๐๐ฒ๐ฉ๐จ๐ญ๐ก๐๐ฌ๐ข๐ฌ. language models pick up skills in the curriculum: copying, then word forms, logic, translation, arithmetic, composite tasks last.
Emmy also shared why she gives 2/10 for ICLR workshop paper written by AI, and her secrete sauce on getting 100K GPU hours compute.
More on ๐๐๐๐ ๐๐ interview, coming soon.
Emmy Liu retweeted
If youโre applying to CMUโs School of Computer Science @SCSatCMU and are interested in the Language Technologies Instituteโs PhD or MLT programs @LTIatCMU, come to our info-session on November 11!
Weโll have faculty there to answer any questions you may have about the department and application process
The LTI is holding a virtual info-session for prospective MLT and PhD applicants on Wednesday, November 11 from 12pm-1pm, ET! Our faculty will be answering any questions about the application process. Register at the following link to attend: forms.gle/pxsdNCY4JXZmjUxJ6
Emmy Liu retweeted
Recently, @_emliu et al.'s paper on "The Implicit Curriculum Hypothesis" won an outstanding paper award at COLM: arxiv.org/abs/2604.08510
This is of course due to the ingenuity and hard work of the authors, but it's also thanks to an @NSF ACCESS grant of compute resources, for which we are very grateful.
Just wanted to explicitly point this out as a positive story of public support for computational resources for open research!
Emmy Liu retweeted
๐งตStill cleaning pretraining corpora with a heuristic HTML scraper plus dozens of hand-written filters? We replaced the whole stack with ReScraper, one unified 0.6B model that beats the strongest curation baseline at every scale ๐
๐ Blog: cxcscmu.github.io/ReScraper
Emmy Liu retweeted
Introducing ๐๐ฉ๐๐๐ฎ๐ฅ๐๐ญ๐ข๐ฏ๐ ๐๐๐๐จ๐๐ข๐ง๐ : ๐๐จ๐ฐ ๐๐ญ ๐๐ฏ๐จ๐ฅ๐ฏ๐๐, ๐๐ก๐๐ง ๐๐ญ ๐๐ญ๐๐ฒ๐ฌ ๐๐จ๐ฌ๐ฌ๐ฅ๐๐ฌ๐ฌ, ๐๐ง๐ ๐๐ก๐๐ญ'๐ฌ ๐๐๐ฑ๐ญ. An interactive tutorial @Madisonkanna and I built for the NeurIPS Education Track.
Every LLM you use generates one token at a time. Autoregressive decoding is the bottleneck for inference. Speculative decoding accelerates this, and today it runs under nearly every hosted LLM. It is a cornerstone topic to learn in the LLM stack.
Blog: neurips2026-speculative-decoโฆ
Emmy Liu retweeted
Tired of AI slop and spaghetti code? We study how to measure and reduce slop in long-horizon coding agents in this work, which will be presented at the #COLM Workshop on Agent Behavior (October 9, 10 - 11:30 AM at Continental Ballroom) ๐งต๐
Emmy Liu retweeted
When LLM demand exceeds capacity, which requests should go first?
Model providers already offer a few priority tiers. But even within a tier, requests compete for compute, and users value latency differently. Your coding agent might be blocked due to a request that belongs to an unattended background job.
Why not let the market decide?
In our new paper, Inference Auctions, users bid for request priority, and our auction allocates that priority to those who value it most.
Unfortunately, naively selling to the highest bidders destroys prefix-cache locality, increasing SGLang's average latency by up to 12x in our experiments.
Our fix is to run an auction over depth-first search traversals of the request prefix tree, ordering subtrees by average bid. This maximizes social welfare while retaining cache optimality, with VCG payments incentivizing truthful bidding, and the resulting request order preserves key performance metrics (cache hits and mean TTFT) of the underlying scheduler.
w/ @keegan_w_harris @sidpra5ad @nhaghtal and Michael Jordan
Paper: arxiv.org/abs/2609.40070
Project page: inferenceauctions.com
Emmy Liu retweeted
๐ขMarin 535B-A23B training has crossed the halfway mark! Some observations and learnings so far:
Emmy Liu retweeted
I sadly won't be participating in #COLM2026 (instead will be teaching the CMU AI Agents course), but my wonderful students will be there.
Say hi to them, especially the ones on the job market this year.