Ph.D. student in CS@Max Planck & MPI-SWS
Saarland, Germany
Joined March 2020
- Tweets42
- Following461
- Followers77
- Likes206
I’m in Budapest for EMNLP 🇭🇺 happy to chat about topics about understanding model behavior, agent tool calling, personalization and multi-agents (and any other interesting topics!)
I’m also near graduation and planning for the next positions in those areas 👻
#EMNLP2026
Qinyuan Wu retweeted
I will be hiring phd students to work on the economics of generative AI, counterfactuals in ML, and human-ai collaboration through @ELLISforEurope and the MP-AIX program. Please, spread the word. Happy to chat in person at NeurIPS in Paris or via [email protected] (1/n)
We do need controllers for LLM agents! But when to call a tool depends on the underlying LLM’s capabilities. How does Jev account for that, especially with unseen models?
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Qinyuan Wu retweeted
📢I am hiring interns at the Max Planck Institute!
Topics of interest:
(1) LLM factuality, reliable information synthesis and reasoning (esp. in agentic settings), personalization
(2) AI creativity
(3) Data-centric interpretability
Apply here by Sep 15: shorturl.at/RmSb1
Qinyuan Wu retweeted
We’ve decided to open-source a multi-agent harness we use internally at YC.
We call it “QM” and it’s meant to be easy to customize, like Hermes or OpenClaw, but useful for a whole company. We use it across accounting, legal, events, and engineering (including building QM itself!).
The whole project is under an MIT license. It is cloud-first and has Slack and web UI natively.
Qinyuan Wu retweeted
🎉 Excited to share that our paper has received a Senior Area Chair (SAC) Highlight at ACL 2026! 🎉
The SAC Highlight recognizes one paper in each Senior Area Chair's batch that stands out for its originality, significance, and potential impact on the field.
Qinyuan Wu retweeted
I'll be in Rio attending ICLR b/w 22nd and 30th!
Looking forward to talking about Pretraining, Fairness with AI Agents and much more. Feel free to DM if anything related interests you or if just want to talk about the field in general and discuss hot takes about AI alignment
🙋♀️ Excited to be at ICLR 2026 in Rio (22–27)! 🌍
We have 4 papers — 2 in the main conference, 2 at the Agents in the Wild workshop — on trustworthy ML: post-training, interpretability, and AI reliability. Opening the black box to understand how models learn and behave in deployment.
Would love to connect with anyone working on impactful AI problems — research or industry! 🎉
#ICLR #LLM #TrustworthyML
2) In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
Pavilion 4 P4-#3818 | Apr 23, 3:15 PM – 5:45 PM
Paper Link: openreview.net/pdf?id=JpEZIM…
3) Characterizing Web Search in The Age of Generative AI
Apr 26 -
Workshop Agents in the Wild: Safety, Security, and Beyond
Paper Link: openreview.net/forum?id=q4aM…)
4) The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT
Apr 26 -
Workshop Agents in the Wild: Safety, Security, and Beyond
Paper Link: openreview.net/pdf?id=iUMNkB…
Qinyuan Wu retweeted
This week, I will be in #Paris for the @IASEAIorg conference, presenting another work on how all three aspects- privacy, utility, and efficiency are achievable during fine-tuning LLMs. I’ll be talking about it this Wednesday in Track 9 on assessing risks in human-AI interaction.
This week, I will be in #Paris for the @IASEAIorg conference, presenting our work on memorization and generalization of LLMs on Wednesday in Track 15: Detecting and responding to AI failures.
Check the preprint paper: arxiv.org/abs/2507.21914
#IASEAI #LLM #Memorization #Generalization
Qinyuan Wu retweeted
This week, I will be in #Paris for the @IASEAIorg conference, presenting our work on how latent source preferences steer LLM generations on Tuesday in Track 6 on Truth, persuasion, and AI design for human users. All kudos to @Aflah02101 for leading this awesome piece of work!
Qinyuan Wu retweeted
This week, I will be in #Paris for the @IASEAIorg conference, presenting our work on ChatGPT memories on Wednesday in Track 9 on assessing risks in Human-AI interactions.
The preprint of the work is available here : arxiv.org/abs/2602.01450
#Memory #ChatGPT #Personalization #AI
Qinyuan Wu retweeted
New paper 🚨: "In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations"
Here, we focus on a novel consideration that arises when designing trustworthy LLM agents: how does the latent knowledge of LLMs shape which information they privilege?
Qinyuan Wu retweeted
Really happy to share that three of our papers have been accepted to ICLR 2026!
1️⃣In Agents We Trust, but Who Do Agents Trust? Latent Preferences Steer LLM Generations
Preview: openreview.net/pdf?id=yTUNl6…