Yonatan Belinkov ✈️ COLM2025

🤔What happens when LLM agents choose between achieving their goals and avoiding harm to humans in realistic management scenarios? Are LLMs pragmatic or prefer to avoid human harm?

🚀 New paper out: ManagerBench: Evaluating the Safety-Pragmatism Trade-off in Autonomous LLMs🚀🧵

Reach out if you'd like to discuss anything related to model interpretability and controllability, robustness, multi-agent communication, biological LMs, etc.
Also happy to talk about PhD and Post-doc opportunities!

Protein2Text: Providing Rich Descriptions for Protein Sequences Understanding the functionality of proteins has been a focal point of biological research due to their critical roles in various biological processes. Unraveling protein functions is essential for adv...

In the #LM4Science workshop, Friday 111:30am-1pm, we have a poster on generating rich text descriptions from protein sequences, work by Edo Dotan, who couldn't travel.
Preprint: www.biorxiv.org/content/10.1...

In #Interplay25 workshop, Friday ~11:30, I'll present on measuring *parametric* CoT faithfulness on behalf of
@mtutek.bsky.social , who couldn't travel:
bsky.app/profile/mtut...

Later that day we'll have a poster on predicting success of model editing by Yanay Soker, who also couldn't travel

In the LLM explainability in reasoning and planning workshop, Friday 9am, I'll talk about scaling up model interpretability (xllm-reasoning-planning-workshop.github.io)

@zorikgekhman.bsky.social
will present on Wednesday a poster on hidden factual knowledge in LMs
bsky.app/profile/zori...

@itay-itzhak.bsky.social
presenting today, morning, a spotlight talk and poster on the origin of cognitive biases in LLMs
bsky.app/profile/itay...

Traveling to #COLM2025 this week, and here's some work from our group and collaborators:
Cognitive biases, hidden knowledge, CoT faithfulness, model editing, and LM4Science
See the thread for details and reach out if you'd like to discuss more!

What's the right unit of analysis for understanding LLM internals? We explore in our mech interp survey (a major update from our 2024 ms).

We’ve added more recent work and more immediately actionable directions for future work. Now published in Computational Linguistics!

Opportunities to join my group in fall 2026:
* PhD applications direct or via ELLIS @ellis.eu (ellis.eu/news/ellis-p...)
* Post-doc applications direct or via Azrieli (azrielifoundation.org/fellows/inte...) or Zuckerman (zuckermanstem.org/ourprograms/...)

Excited to join @KempnerInst this year!
Get in touch if you're in the Boston area and want to chat about anything related to AI interpretability, robustness, interventions, safety, multi-modality, protein/DNA LMs, new architectures, multi-agent communication, or anything else you're excited about!

@robinjia.bsky.social
speaking at
@kempnerinstitute.bsky.social
on Auditing, Dissecting, and Evaluating LLMs

Thrilled that FUR was accepted to @emnlpmeeting.bsky.social Main🎉

In case you can’t wait so long to hear about it in person, it will also be presented as an oral at @interplay-workshop.bsky.social @colmweb.org 🥳

FUR is a parametric test assessing whether CoTs faithfully verbalize latent reasoning.

BlackboxNLP is the workshop on interpreting and analyzing NLP models (including LLMs, VLMs, etc). We accept full (archival) papers and extended abstracts.

The workshop is highly attended and is a great exposure for your finished work or feedback on work in progress.

#emnlp2025 at Sujhou, China!

Join our Discord for discussions and a bunch of simple submission ideas you can try!
discord.gg/n5uwjQcxPR

Participants will have the option to write a system description paper that gets published.

Have you started working on your submission for the MIB shared task yet? Tell us what you’re exploring!

New featurization methods?
Circuit pruning?
Better feature attribution?

We'd love to hear about it 👇

Working on feature attribution, circuit discovery, feature alignment, or sparse coding?
Consider submitting your work to the MIB Shared Task, part of this year’s #BlackboxNLP

We welcome submissions of both existing methods and new or experimental POCs!

VLMs perform better on questions about text than when answering the same questions about images - but why? and how can we fix it?

In a new project led by Yaniv (@YNikankin on the other app), we investigate this gap from an mechanistic perspective, and use our findings to close a third of it! 🧵

Tried steering with SAEs and found that not all features behave as expected?

Check out our new preprint - "SAEs Are Good for Steering - If You Select the Right Features" 🧵

🚨New paper at #ACL2025 Findings!
REVS: Unlearning Sensitive Information in LMs via Rank Editing in the Vocabulary Space.
LMs memorize and leak sensitive data—emails, SSNs, URLs from their training.
We propose a surgical method to unlearn it.
🧵👇w/ @boknilev.bsky.social @mtutek.bsky.social
1/8

Interested in mechanistic interpretability and care about evaluation? Please consider submitting to our shared task at #blackboxNLP this year!

Slides available here: docs.google.com/presentation...

I think a scalable open source implementation would have many uses! Let’s say I can’t run all pretraining data because of cost. And I run a subset and get influential examples. What would that mean concerning what I’m missing?

Looks great! What would it take to run this on another model and dataset?

This has been a huge team effort with many talented contributors. Very thankful for everyone’s contributions!

See the list here:
bsky.app/profile/amuu...

Excited about the release of MIB, a mechanistic Interpretability benchmark!

Come talk to us at #iclr2025 and consider submitting to the leaderboard.

We’re also planning a shared task around it at #blackboxNLP this year, located with #emnlp2025

We release many public resources, including:

🌐 Website: mib-bench.github.io
📄 Data: huggingface.co/collections/...
💻 Code: github.com/aaronmueller...
📊 Leaderboard: Coming very soon!

Hope to see you there!

1/13 LLM circuits tell us where the computation happens inside the model—but the computation varies by token position, a key detail often ignored!
We propose a method to automatically find position-aware circuits, improving faithfulness while keeping circuits compact. 🧵👇

Wondering about the mech interp passage where you say cognitive science can offer normative theories for ToM. Why is it important for a model to have a specific cog sci theory to be able to say that it has ToM? Could it implement a different mechanism for ToM than people?