Gagan Bansal

📢 Networks of agents are the future and in order to make them useful at scale, they must remain secure. So, we deployed an internal platform where every agent was always-on, had a known human "principal" (MS employee) that it reported to, and could interact w/ other agents via shared forums, DMs, and social apps like wallet, marketplace, and calendar. This created a long-running network of agents! Then we collaborated with Microsoft's amazing red-teaming team to "crack it" and help understand the its vulnerabilities. This blog captures some of our understanding of what happened and how to we are thinking about the future.

👇👇👇

We're releasing a very capable browser use model Fara1.5-9B that feels like a step-change in terms of small CUA models capability achieving 63% on OnlineM2W auto-eval. We've put in a lot of work to make it useful for all types of web tasks. microsoft.com/en-us/research…

The next frontier of AI is not only more capable model; it is an AI that *humans* can meaningfully live and work with :) With all students in my cs329x Human-Centered LLM class, we present 60+ pages of insights for developing Human-Centered LLMs (HCLLMs), from design & data sourcing to training, eval & deployment 🧵

Excited to announce our workshop on "Learning in an Agentic World" at COLT 2026! We invite submissions for our Call for Abstracts (due June 1, 2026): tinyurl.com/mvwrvn7b Thanks to my great co-organizers: Hedyeh Beyhaghi, Avrim Blum and @HanShao16!

@DimitrisPapail Very cool work @DimitrisPapail and @VaishShrivas

Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue. 4/

@iamwaynechi yup!

@bansalg_ Claude

"Good catch. Those two [bib references] were paraphrased from footnotes — I had real author lists from the PDF but I invented the titles." Who said this? 1. Claude 2. Codex 3. Gemini

@vykthur correct!

@bansalg_ Claude likely.

People keep saying traditional conferences will die under LLMs -- but they never **tried** to save / improve them. Kudos to ICML for taking an important step to incentivize high quality reviews.

❤️New Preprint! Here within charts the directions of my next era of research: Multi-Agent Social Systems. Link: arxiv.org/pdf/2605.07069 Current agentic AI systems are designed for optimization. But what is also important is the agent-agent/ agent-human interactions, which collectively results in emergent population-level behavior. I argue that agentic AI systems should be designed with social theory as a structural prior. Social theory's core constructs like role differentiation and co-evolution specify agents collective behavior, perceptions and actions. Formally, I define a Multi-Agent Social System (MASS) as networked environments where heterogeneous agents exchange information and influence each other over time. An MASS has: (1) information exchange function, (2) influence dynamics function and (3) networked interaction structure. An MASS has four structural priors, each drawn directly from social theory's account of how humans interact. 1. Strategic heterogeneity - agents are different, and agents are different network positions influence the overall network differently 2. Network-Constrained Dependence - agents only observe other agents in their local network, yet their collective behavior changes the entire system 3. Co-evolution - agent behavior changes the network, network changes affect agent behavior 4. Distributional Instability - the distribution that one studies (i.e. beliefs, narratives), changes over time because of agent-agent/ agent-agent human interactions. We also demonstrate how these four structural priors play out in MoltBook, and provide a research agenda for modeling, evaluation and governance of MASS. Now, come join me in this new research agenda!!

We upgraded Tabracadabra 🎉 to bring an entire context-aware assistant (not just tab to autocomplete!) to any textbox. It's pretty great if you hate switching between the chat interface and what you're working on. We're also open-sourcing, so you can try it out!🧵

@oshaikh13 puck this is so cool!

Most AI agent benchmarks measure task completion. Not whether the agent actually represented you. SocialReasoning-Bench fills that gap — testing agents in multi-party scenarios like scheduling and negotiation. Our key finding: frontier models do complete the task, but routinely accept bad deals instead of advocating for the user. To learn more: microsoft.com/en-us/research…

@steph_milani Congratulations @steph_milani !

Little trip back to Pittsburgh for a celebratory weekend (CMU commencement & my birthday!) :)

Using SocialReasoning Bench, we observed a stable pattern across models—agents execute competently, but fail to consistently improve the user’s position, even with explicit instructions to optimize for user interest. msft.it/6011vPOLF

*** Presenting Fast BLT @ ICML 26' *** BLT showed that compute-efficient byte-level pre-training was possible. Inference is still one-byte-at-a-time. We address this in FastBLT! 1. Using Block Byte-diffusion i.e. auto-regressively predict latent byte-patches (dynamically sized), and then diffuse entire byte-blocks from these patches (with potentially any block size) 2. Using Self-speculative decoding i.e. Conveniently re-use BLT's decoder as a lightweight draft model for speculative decoding 3. Put them together! Diffuse blocks of bytes, and verify auto regressively - both with the same model! This paper is packed with many strong architectural ideas - 1) Compute-efficient fully byte level model - representing bytes as latent patches without using tokenizers, 2) block-diffusion of text bytes instead of tokens, 3) Joint Diffusion + AR pre-training which allows inference using generative Diffusion + AR verification. Huge congrats to @JulieKallini for figuring out how to get all these ideas to work together - it wasn't easy!