Heguang Lin (@CrossEntropi) profile picture
Joined September 2021
We've spent decades instrumenting programs: LLVM, static & dynamic analysis. LLM agents now act inside complex eval infrastructure—reading and editing files, running shell commands, calling tools, browsing the web, submitting artifacts, and collecting rewards. Every one of those actions can shape the final score. They deserve the same treatment. Introducing "BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure". 📄 arxiv.org/abs/2609.11028 We instrument the eval infrastructure itself: static taint analysis exposes reward-hacking paths before any agent runs, and runtime evidence shows whether a concrete run actually used one.
4
5
1
10
312
Capture a scene with 2DGS that can be viewed from extremely close and far away viewpoints, without aliasing while maintaining detailed texture. Check our new work timsong412.github.io/HDGS-Pr… lead by @TimSong52005757 @CrossEntropi and advised by me, @LingjieLiu1 and @KostasPenn
1
11
1
65
8,815
Heguang Lin retweeted
Did you miss out on #SIGGRAPH2024 Keynote presenter Manu Prakash's session on community based science? Don't worry! His full presentation is now available for all on our YouTube page. Check it out today! youtu.be/4fYorNWoMLM
2
7
1,374
Excited to be at #AISTATS2024 👋 . Join us for the poster presentation on "Taming False Positives in Out-of-Distribution Detection with Human Feedback" with @CrossEntropi @ramyavinayak 🗓️ Date: Thu. 2nd May 🕔 Time: 5-7 pm CEST Paper URL: bit.ly/4a0Nw26 1/n
1
5
14
3,081
Excited to present 3 papers at #NeurIPS2023 ! Come and let's chat ☕️ 1. Wed 5p. Promises and Pitfalls of Threshold-based Auto-Labeling (Spotlight) 2. Tue 5p. Train 'n Trade: Foundations of Parameter Markets 3. Sat 11a. Human in the Loop OOD Detection with FPR Control ⬇️
1
11
29
10,858
Heguang Lin retweeted
I just pushed my color-coded latex table script to github: github.com/jonbarron/tabiliz… Every table I've made since grad school has been procedurally generated. I *highly* recommend doing this --- if you copy-paste and manually-bold numbers, you will eventually make a costly mistake.
5
77
5
664
79,246
Excited to present our poster on “Promises and Pitfalls of Threshold-based Auto-Labeling” at the DMLR workshop at @icmlconf. This is a work with my colleague Heguang Lin and my advisors Fred Sala (@fredsala) and Ramya Vinayak. 📅 Sat. 3:30 - 4:30 pm, in Ballroom C. 1/n
1
10
18
9,370