Zining Zhu
@zhuzining
Assistant Professor @FollowStevens (2024-) PhD @UofT, @VectorInst Areas: #NLProc #Explainable #AI
New paper: What happens when an LLM reasons? We created methods to interpret reasoning steps & their connections: resampling CoT, attention analysis, & suppressing attention We discover thought anchors: key steps shaping everything else. Check our tool & unpack CoT yourself 🧵
Reviewers should perhaps be prohibited from changing the scores they give on the day of seeing the scores of their own papers submitted to @ReviewAcl.
🚨Deadline alert: If you work on LLM explainability for reasoning and planning, submit your work by June 23! - Non-archival, two formats (long/short) - Welcome recently accepted papers and dual submissions - 🏆Two awards will be announced! Details: …reasoning-planning-workshop.github.io
Dear MAGA friends, I have been worrying about STEM in the US a lot, because right now the Senate is writing new laws that cut 75% of the STEM budget in the US. Sorry for the long post, but the issue is really important, and I want to share what I know about it. The entire…
Re: “Every major breakthrough in AI has been American”: America does itself no favors when it overestimates its specialness. Yes, the center of the AI industry is the US (California!), but many of the breakthroughs of (neural, gradient-based) AI happened elsewhere: • LSTMs,…
DeepSeek is a wake up call for America, but it doesn’t change the strategy: - USA must out-innovate &race faster, as we have done in the entire history of AI - Tighten export controls on chips so that we can maintain future leads Every major breakthrough in AI has been American
We replicated the DeepSeek-R1-Zero and DeepSeek-R1 training on 7B model with only 8K examples, the results are surprisingly strong. 🚀 Starting from Qwen2.5-Math-7B (base model), we perform RL on it directly. No SFT, no reward model, just 8K MATH examples for verification, the…
arxiv.org/abs/2501.11183 LLM safe-tuning can learn from lessons in cybersecurity.
My 2024 predictions (agents, multimodality, synthetic data, sparsity) all came true, but those were generic and somewhat obvious in hindsight. So I'll make some bolder, more concrete predictions about AI in 2025: 🤖 On agents 1. We will surely see a rapid improvement in AI…
A uniquely interesting book with a lot of new information, and I feel the urge to take notes (either to echo or to debate) while reading. Highly recommend.
Glad to see this paper appearing in Stevens News. Nowadays the improved LLMs know to reject many very wrong problems, but still, it's quite easy to find adversarially helpful examples. We recommend LLM users (especially when using LLMs to explain unknown problems) fact-check the…
Everybody at school taught us morals and values. Most people in academia who I know e.g., at #NeurIPS are honest, morally upright, and are against racism.
ACL 2025 is looking for reviewers and ACs in Interpretability and Analysis of NLP Models! - DM if interested in emergency reviewer or AC roles (late March/early April) - Self-nominate for reviewer/AC roles here (early-mid March): docs.google.com/forms/d/e/1FAI…
Work led by myself with @elliot_creager, @zhuzining, and @SPOClab. arXiv: arxiv.org/abs/2412.04576 If you happen to be in Vancouver for NeurIPS, come chat with Elliot about our work at the Workshop on Creativity & Generative AI [creativity-ai.github.io] on Sat Dec 14! N/N
Survey results identify how scientific research and teaching contribute to the graduate student mental health crisis go.nature.com/3Z5GqHl rdcu.be/d0K8f