Manuel Cherep

PhD student @MIT working on behavioral machine learning

[Manuel Cherep @manuelcherep]

Yours truly, the program committee 🙂

@manuelcherep (MIT)
@_Hao_Zhu (Stanford)
@StevenyzZhang (Georgia Tech + Stanford)
@Xinyang_Han_ (UC Berkeley)
@BenSManning (MIT)
@isi_magistrali (ETH)
Saab Mansour (Amazon)
@weronika_laj (Amazon)
@PattieMaes (MIT)
@nikhilsinghmus (Dartmouth)

[Manuel Cherep @manuelcherep]

✨Announcing the first Workshop on Agent Behavior @COLM_conf 2026 (Oct 9, San Francisco 🌅)

https://t.co/8vYnVp4uXf

We invite two types of contributions: (i) papers, and (ii) benchmark proposals. We are also seeking reviewers. More details below!

[Manuel Cherep @manuelcherep]

If you’re interested in being a reviewer, please fill out this form!

https://t.co/rGRQF9aOm9

[Manuel Cherep @manuelcherep]

@abhi__katiyar @iclr_conf Thank you!

[Manuel Cherep @manuelcherep]

ABxLab is accepted at @iclr_conf #ICLR 2026! ✨We ask: why do AI agents do what they do? 🧐

We introduce a framework for systematically studying AI agent behavior through controlled manipulations of their environments. We accomplish this by intercepting any real web environments and modifying what the agent sees in real time before they actually see it.

[Manuel Cherep @manuelcherep]

Work with Chengtian Ma, Abigail Xu, Maya Shaked, @pattiemaes, @nikhilsinghmus

🌐Web: https://t.co/04NDCCuyiZ
💻Code: https://t.co/KIkq28WEUe
📄Paper: https://t.co/sDr9LHqWqk

Would love to hear your thoughts!

[Manuel Cherep @manuelcherep]

The world is also full of visual cues 👀, and you might be wondering whether agents are sensitive to these as well. The answer is yes! Check out our new paper, where we introduce an optimization method for editing images to understand VLMs’ decisions:

https://t.co/Mpcx9P3GRG

[Manuel Cherep @manuelcherep]

Some decisions we make with our eyes 👀, but what about VLMs? Do they have structured, exploitable visual preferences that we can discover systematically before adversarial actors do?

In our new paper, we propose a new optimization method for this and show substantial effects on VLMs’ decisions.

[Nikhil Singh @nikhilsinghmus]

Excited to (finally) share this paper, accepted at @iclr_conf #ICLR 2026! ✨

In this work, we use sparse autoencoders (SAEs) to study the internal representations of generative music models (here, MusicGen) and automatically discover how they encode concepts.

[Manuel Cherep @manuelcherep]

@sanskxr02 Thank you!

[Manuel Cherep @manuelcherep]

Some decisions we make with our eyes 👀, but what about VLMs? Do they have structured, exploitable visual preferences that we can discover systematically before adversarial actors do?

In our new paper, we propose a new optimization method for this and show substantial effects on VLMs’ decisions.

[Manuel Cherep @manuelcherep]

Do you see like an agent? Try it yourself: https://t.co/ETgo1Wr4tp

Paper: https://t.co/BTx8tnf3c4

Co-Authors: Pranav M R, Pattie Maes (@PattieMaes), Nikhil Singh (@nikhilsinghmus)

[Manuel Cherep @manuelcherep]

In our recent ICLR 2026 paper, we showed how to study other kinds of sensitivities in agent behavior by using counterfactuals with our new framework (ABxLab)

https://t.co/GCkEFGBgMr

[Manuel Cherep @manuelcherep]

How does it work? ABxLAB is a "man-in-the-middle" framework.

It intercepts web content in real-time to run controlled experiments on agents by modifying the choice architecture.

Think of it as a behavioral science lab for LLMs.

Paper: https://t.co/g6rdkg108n

🧵2/9