Frank M. Freimann Collegiate Professor at Notre Dame CSE | Data Mining | NLP | AI
Notre Dame, IN
Joined August 2012
- Tweets224
- Following556
- Followers1.6K
- Likes361
A good step forward. Likely more cost-effective than hiring a dedicated team to respond to issues on the repository, assuming issue submissions are enabled.
We want to support the mathematics community in turning progress in AI into deeper human understanding and new discoveries. To support that goal, we are contributing $5 million to a new Simons Foundation initiative to help mathematicians examine, understand, and build on results produced by AI, including the results we recently published.
The funding, managed and awarded by the @SimonsFdn will support workshops, conferences, and special programs. We will also provide access to our models and computing resources to Simons Foundation-supported researchers.
An idea for making AI math results more accessible & accountable (if had unlimited tokens): Generate a 45min lecture for each result, as if presented by its author; Enable comments to identify where humans get lost or need clarification; Release a follow-up video every two weeks.
Students who were RAs are now more like PIs, as they can manage one or multiple teams of artificial RAs if they know how to collaborate with them effectively.
Professors who were PIs are now more like senior PIs of large-scale research programs or directors of institutes. ...
Would OpenAI dare to give a 0.01% equity stake to anyone who finds an error in its AI-gen outputs (math results)? Mathematicians stake their reputations & intellectual integrity on their work. Tech giants reap too much, risk too little & end up wasting humanity’s brightest minds.
Meng Jiang retweeted
Can an LLM learn to write user profiles that improve recommendations?
Excited to share BLUE, work from my Google internship, accepted to #NeurIPS2026!
We train a profiler using feedback from an embedding model.
arxiv.org/abs/2605.06981
It's our great pleasure to host Dr. Lu's visit at Notre Dame! Looking forward to his talk in 30 minutes!
CSE Seminar Speaker: Zhiyong Lu, NIH - "Building Trustworthy AI Agents for Biomedical Discovery and Clinical Decision Support": cse.nd.edu/events/building-t…
☘️ Join us today!
🕞 3:30pm
📍303 Cushing
Meng Jiang retweeted
We built Agent as Policy (AGP), a robot control harness on top of GPT-6 Astra.
With AGP, Astra performs zero-shot manipulation on real robots across precision manipulation, long horizon tasks, dynamic actions, and deformable objects.
How it works. 🧵
agent-as-policy-2026.github.…
Please submit papers to NeurIPS 2026 Workshop on Continual Learning for Enterprise AI Agents (CLEA)! We welcome submissions on continual learning, AI agents, reinforcement learning, agent memory, workflow learning, planning, evaluation, and enterprise AI. clea-neurips.github.io/
Benchmarks are lenses on reality. Their motivations, construction, and evolution encode signals about the real world. Can we use signals across benchmarks to find samples that best predict deployment performance before seeing any new app data?
Thought #3: Is the widespread use of AI code-generation tools amplifying inequalities among programming languages, where some languages are more popular and therefore more widely created, used, and represented, while others remain less popular and are created and used far less?
As this imbalance intensifies, are the flaws of programming languages more likely to be exposed because of the dominance of a single language, or less likely to be exposed because of the availability of variety?
Random thought #2: When the solution space is huge, humans may only know <1% of the strategies for finding optimal solutions. If AI can explore more broadly and receive accurate feedback, it may discover much of the remaining 99%. The key is feedback: ...
My students keep telling me to tweet more, and I keep making excuses. One excuse is thinking every social media post has to be polished or joyful. I know that’s not true, so I’m practicing: one random thought a week.
Best conference experience this year so far! Conference organizers are awesome; speakers are super super good!
@CAISconf was sold out but small, focused on a particular well-scoped area (compound agentic systems). Some workshop papers were submitted as late as just a couple weeks before (fresh content), and good thing, too! Agentic systems already look quite different today than even 3 months ago.
Perhaps the systems focus naturally attracts researchers who are inclined to build prototypes and libraries (not just publish), but I was still pleasantly surprised to see how many members of the Laude community had independently identified this conference as one they wanted to submit to and attend. On top of that, throwing a lounge after hours with food, drinks, demos, salon-style conversation, (and a mini podcast studio, because why not?) made it easy to get an even higher concentration of interesting people with interesting half-baked ideas in a particular niche that happens to be on fire right now.
Loved the experience. Well-done, CAIS organizers (@heathercmiller, @lateinteraction, @matei_zaharia, @deeptir18, et al). Same time, same place next year?
Meng Jiang retweeted
🔥 New survey: Inference-Time Control for Trustworthy LLMs.
Once a model ships, training-time alignment can't keep up. Inference-time control is how we patch the gap.
🚀 We collected 200+ papers (2020–2026) and built the first pipeline-grounded taxonomy that unifies seven method families: context engineering, guardrails, decoding, representation engineering, unlearning, pruning, and multi-agent orchestration.
📄 Paper: doi.org/10.20944/preprints20…
Existing surveys cut the field by harm type or by one method family. We treat inference-time interventions as a unified control plane over the generation pipeline.
💡 The Framework: Three tiers, by where the intervention happens.
🛡️ External: around the model (prompts, guardrails, decoding)
🧠 Internal: inside the model (steering, unlearning, pruning)
🤝 System-level: across models (debate, verification)
📊 The Coverage:
🔍 200+ papers indexed (2020 → 2026)
🧠 7 method categories, one pipeline-grounded map
📈 4 trustworthy dimensions ✕ 5 evaluation axes = a meta-axis grid
🗂️ Repo (live, 200+ papers indexed):
github.com/leopoldwhite/Awes…
🌐 Website: leopoldwhite.github.io/Aweso…
👇 Full thread below.
