VP of Applied AI Research @CapitalOne | Adjunct @Georgetown | Simple baselines, practical implementations.
Washington, D.C.
Joined April 2009
- Tweets830
- Following1K
- Followers560
- Likes1.7K
Bayan Bruss retweeted
Agents can now automate large parts of the ML training pipeline, including code generation, debugging, and model tuning. However, these autoresearch agents are often unreliable due to harness failures. We trace these failures and develop model-agnostic fixes.
Bayan Bruss retweeted
Our new guardian model lets you create LLM guardrails using natural text. This little 8B model efficiently checks in real time whether chatbots comply with bespoke moderation policies.
It's not often that academics beats industry models, but DynaGuard stacks up well!
Guardrails with custom polices are hard for models trained on safety and harm-related datasets. But what if you trained a guardian model on arbitrary rules?
Introducing DynaGuard, a guardian model for custom policies: arxiv.org/abs/2509.02563
Bayan Bruss retweeted
There is still a lot of brittleness in getting guardian models to incorporate custom policies, but we think this is a step in the right direction. Try out DynaGuard in this interactive demo (and give us feedback to improve it!): huggingface.co/spaces/tomg-g…
Bayan Bruss retweeted
Paper: arxiv.org/abs/2509.02563
Models: huggingface.co/collections/t…
Dataset: huggingface.co/datasets/tomg…
This was a collaborative effort with
@neeljain1717, @k_saifullaah, @taruschirag, @vatsalbaherwani, Joseph Vincent, Melissa Kazemi Rad, @cbbruss, @PandaAshwinee, @tomgoldsteincs
Bayan Bruss retweeted
Guardrails with custom polices are hard for models trained on safety and harm-related datasets. But what if you trained a guardian model on arbitrary rules?
Introducing DynaGuard, a guardian model for custom policies: arxiv.org/abs/2509.02563
Bayan Bruss retweeted
Your critic model is secretly a strong policy model. Stay tuned for a deep dive 🤩
Bayan Bruss retweeted
I am looking for a postdoc to lead projects related to this collaboration, on scaling laws, emergence and interpretability in pre- and post-training & inference/reasoning, in multimodal foundation models (language, time series, tabular data etc). HPC experience is a plus.
So excited to join the new @SimonsFdn Simons Collaboration on the Physics of Learning and Neural Computation, to further advance our understanding (mech interp) of learning & reasoning in large networks, including classical deep nets and other bio-inspired network models!
If you share this belief, I’ve got some cool work for you to do
Bayan Bruss retweeted
Open-weights for our Llip multimodal vision-language model led by @lavoiems are public!
LLIP proposes new pre-training objective to capture the many ways to describe an image leading to strong performance across a suite of 22-zero shot benchmarks.
xitter.cf/lavoiems/status/194582…
The code and model weights for this paper are finally open! Despite being a little late for releasing them, I hope you will find them useful!
Code: github.com/facebookresearch/…
Models:
- (ViT-G): huggingface.co/lavoies/llip-…
- (ViT-B): huggingface.co/lavoies/llip-…
Bayan Bruss retweeted
💰 Suppose you’re handed $1M and a zoo of tokenizers used by top LLMs.
You want to train a new model—but which tokenizer should you pick? 🤔
Naively, you’d train one model per tokenizer. But with just $1M, that’s a budget killer. 💸
Is there a smarter way to train once and still find the best tokenizer?
🔥 Enter YOTO: You Only Train Once 🔥
A unified framework that co-trains a language model and a distribution over tokenizers.
Instead of retraining from scratch, YOTO merges vocabularies and samples tokenizations adaptively—letting model and tokenizer evolve together.
Applied to arithmetic reasoning, YOTO discovers high-performing number tokenizers—and slashes compute costs.
🚀 One model. Many tokenizers. No compromise.
I am presenting the YOTO poster at the TokShop workshop #ICML2025 at west meeting room 111-112.
It has been so rewarding to work with @furongh. This is a groundbreaking for financial services.
As many of you know, I’ve been spending my sabbatical at Capital One with Bayan Bruss @cbbruss’s amazing team, under Prem Natarajan and Milind Naphade. Bayan has been the most supportive and visionary leader: deeply attentive to the technical work across the group.
Excited to present as a @CapitalOne associate today at #ICML, 4:30–6:00 PM at West Ballroom A, at the expo session “AI in Finance: Innovation & Emerging Opportunities.” I’ll be giving a lightning talk on “Building World Models for Financial Decision-Making.”
#ICML2025
#AI4Finance
#CapitalOne
Bayan Bruss retweeted
Check out @yucenlily's ICML poster tomorrow 11 am - 1:30 pm, East Exhibition Hall A-B # E-603, on "Out-of-Distribution Detection Methods Answer the Wrong Questions"!
In our new ICML paper, we show that popular families of OOD detection procedures, such as feature and logit based methods, are fundamentally misspecified, answering a different question than “is this point from a different distribution?”
arxiv.org/abs/2507.01831 [1/7]
When threatened that it would be turned off, ChatGPT-creator OpenAI’s o1 tried to download itself onto external servers and denied it when caught red-handed, per FORTUNE.
Readers added context they thought people might want to know
It was a controlled simulation & the model acted this way after it was prompted to pursue a goal at any cost
Tests are done to prevent it from really happening. It tried to copy itself 2% of times
x.com/bshlgrs/status…
x.com/AlexMeinke/sta…
x.com/mandeepabagga/…
The paper:
static1.squarespace.com/static/6593e70…
The work from @polkirichenko and @marksibrahim is always 🔥
The biggest AI surprise coming is not how much valuable knowledge work it can complete, but how little value most knowledge work produces.
I’ve seen this paper making the rounds again. Early on when starting a new Applied Research team I came across this paper and it was extremely informative to the approach we ended up taking. Not siloing researchers is uncomfortable but critical for impact.
Google's Hybrid Research Model
- blurs lines between research and engineering, aiming for rapid, practical user benefits.
- Teams iteratively develop production-ready code from day one, leveraging Google's infrastructure and user base for large-scale, "in vivo" experiments on real data.
- Long-term goals are factorized into shorter, measurable, often commercially valuable steps.
- Features organizational flexibility, fluid movement of people/projects, and diverse knowledge dissemination, minimizing traditional technology transfer issues.
Bayan Bruss retweeted
The result in this paper I'm most excited about:
We showed that planning in world model latent space allows successful zero-shot generalization to *new* tasks!
Project website: latent-planning.github.io
Paper: arxiv.org/abs/2502.14819
CDS PhD Vlad Sobal (@vlad_is_ai) and Courant PhD Wancong (Kevin) Zhang show that when good data is scarce, planning beats traditional reinforcement learning.
With @kchonyc, @timrudner, and @ylecun.
nyudatascience.medium.com/wh…
Bayan Bruss retweeted
In less than 24h! Zoom will be open to all! Program below:
9am -> 9:10am opening
9:10am -> 9:55am Phillip Isola
9:55am -> 10:40am Thomas Serre
11am -> 11:45am Eero Simoncelli
11:45am -> 12:30pm Yi Ma
1:30pm -> 2:15pm Yann LeCun
2:15pm -> 3pm Nicolas Ballas
3:00pm -> 3:45pm lightning talks
4:15pm -> 5pm Erik Learned-Miller
5pm. ->6pm: Panel discussion
