causality and human behavior in language models assistant prof. @HebrewU; research scientist @Google ex: {@blei_lab, @TechnionLive}
Tel Aviv
Joined July 2016
- Tweets126
- Following189
- Followers596
- Likes514
Amir Feder retweeted
I am flying to @COLM_conf to present STATe-of-Thoughts (arxiv.org/pdf/2602.14265) along with @trsaenger , @GiladMorad12 , and @amir_feder .
We are presenting our poster on Thursday, October 8th, at the Imperial Ballroom #9 between 2 and 4 PM.
Github: github.com/zbambergerNLP/sta…
Amir Feder retweeted
We published a new paper at TACL! Come see the poster at #EMNLP2026 in Budapest next month!
We have a new paper on topic models!
We introduce Mechanistic Topic Models (MTMs). MTMs model SAE feature counts rather than words or plain embeddings, combining the benefits of LLM-based and probabilistic topic models.
Paper (published in TACL) : arxiv.org/abs/2507.23220
Amir Feder retweeted
We have a new paper on topic models!
We introduce Mechanistic Topic Models (MTMs). MTMs model SAE feature counts rather than words or plain embeddings, combining the benefits of LLM-based and probabilistic topic models.
Paper (published in TACL) : arxiv.org/abs/2507.23220
Amir Feder retweeted
We are presenting a poster at EMNLP in Budapest next month!
Github: github.com/blei-lab/mechanis…
With @carolinazheng_, @nicolasvelezb, @SwetaKarlekar, @causalclaudia, @AchilleNazaret, Asif Mallik, @amir_feder, and David Blei.
Amir Feder retweeted
🚨 Can AI assistants understand social situations only from the user’s description? And how can we test this when we often don’t know what really happened?
We tackle this using multi-agent simulations with verifiable ground truth 🎯
More details in @TaubenfeldAmir’s thread 👇
Amir Feder retweeted
12/12
Building assistants that are truly reliable in the social settings is one of the important open challenges in the field. To support further research we open-source Fuse and a dataset with 21k examples.
@zorikgekhman
@amir_feder
arxiv.org/abs/2609.17496v1
Amir Feder retweeted
New paper 📣
LLM assistants are widely used for everyday social advice, making it essential to study how they reason from users' subjective narratives.
We introduce a multi-agent social simulation framework that quantifies this by constructing verifiable ground truth.
1/12👇
Banger paper from Google Research.
This one is on how LLM assistants reason about the people in a user's life.
People ask assistants for social advice constantly, and the assistant only hears the user's version of events.
Measuring whether it reads the situation correctly is hard, because other people's intentions have no ground truth.
Fuse builds that ground truth with simulation.
A target agent with a hidden motive interacts with other agents, including one playing the user. The user agent then describes what happened to the assistant, which has to infer the motive.
The team validated the simulations with 24k human annotations and tested 12 LLMs.
Hearing events through the user makes the task harder. Biased framing from the user shifts the assistant's answer. Models sometimes need more detail than humans do, and longer conversations with room for clarifying questions did not reliably help.
They release the framework and 21k examples.
Paper: academy.dair.ai/papers/verif…
Check out our recent work, led by the superstar duo @zorikgekhman and @TaubenfeldAmir!
Banger paper from Google Research.
This one is on how LLM assistants reason about the people in a user's life.
People ask assistants for social advice constantly, and the assistant only hears the user's version of events.
Measuring whether it reads the situation correctly is hard, because other people's intentions have no ground truth.
Fuse builds that ground truth with simulation.
A target agent with a hidden motive interacts with other agents, including one playing the user. The user agent then describes what happened to the assistant, which has to infer the motive.
The team validated the simulations with 24k human annotations and tested 12 LLMs.
Hearing events through the user makes the task harder. Biased framing from the user shifts the assistant's answer. Models sometimes need more detail than humans do, and longer conversations with room for clarifying questions did not reliably help.
They release the framework and 21k examples.
Paper: academy.dair.ai/papers/verif…
Amir Feder retweeted
I am back in Israel for the next couple of weeks and would be happy to see friends, colleagues, and students. Please feel free to reach out if you are around, or to share with others who might be interested.
I will be giving a few talks on Causal AI and why causality and counterfactual reasoning are central to intelligence, decision-making, and the next generation of AI systems.
Current schedule:
Tel Aviv University
Sunday, June 28, 2026
11:00am-12:00pm
Check Point 420
Technion -- Israel Institute of Technology
Tuesday, June 30, 2026
2:30pm-3:30pm
Taub 337
Weizmann Institute of Science
Wednesday, July 1, 2026
11:15am-12:15pm
Location TBC
The Hebrew University of Jerusalem
Sunday, July 5, 2026
2:00pm-3:30pm
The Federmann Center for the Study of Rationality
Eilath Hall, 2nd Floor, Feldman Building, Givat Ram Campus
Looking forward to the talks and to many interesting conversations while I am here.
Amir Feder retweeted
Excited to share our new preprint: "Causal Effect Estimation with Latent Textual Treatments" 🚀
We tackle a core challenge in text-based causal inference: how to perform valid experimentation with latent concepts as treatments.
arxiv.org/abs/2602.15730
1/8 🧵
Amir Feder retweeted
Semantics in language is naturally hierarchical, but attempts to interpret LLMs often ignore this.
Turns out: baking semantic hierarchy into sparse autoencoders can give big jumps in interpretability and efficiency.
Thread + bonus musings on the value of SAEs:
Amir Feder retweeted
We secure all communications with a cloud-hosted LLM, running on an H100 in confidential mode.
Latency overhead goes away once you cross the 10B model size.
This is our first foray into applied cryptography -- help us refine our ideas.
can you chat privately with a cloud llm—*without* sacrificing speed?
excited to release minions secure chat: an open-source protocol for end-to-end encrypted llm chat with <1% latency overhead (even @ 30B+ params!). cloud providers can’t peek—messages decrypt only inside a secure gpu enclave, where inference stays fully confidential 🤯
links + code in comments👇
Amir Feder retweeted
⌨️ 😇 Drafting for NeurIPS? Submit to #ICML2025 workshop on Scaling Up Intervention Models (SIM) too!
Let’s enjoy some fun science in Vancouver this July. 🌞🌳
🗓️Workshop submission due on May 20 AOE
🔗More info: sites.google.com/view/sim-ic…
Amir Feder retweeted
new version of paper on worker rights in union contracts using NLP, with empirical work showing that these rights are valued like non-wage amenities.
thread here: xitter.cf/BenjaminArold/status/1…
Using new tools in legal language processing to show workers value the legal rights provided by collective bargaining agreements, from @BenjaminArold, @ellliottt, W. Bentley MacLeod, and @snaidunl nber.org/papers/w33605
Amir Feder retweeted
🚨 It's often claimed that LLMs know more facts than they show in their outputs, but what does this actually mean, and how can we measure this “hidden knowledge”?
In our new paper, we clearly define this concept and design controlled experiments to test it.
1/🧵
Amir Feder retweeted
This is going to be great! Looking forward to seeing you all at ICML.
📢 Excited to announce the #ICML2025 workshop on *Scaling Up Intervention Models (SIM)*! Let’s bring together state-of-the-art ideas on modeling novel interventions and distribution shifts. :)
🙌🏻 Submissions are welcome! Link: sites.google.com/view/sim-ic…
Amir Feder retweeted
📢 Excited to announce the #ICML2025 workshop on *Scaling Up Intervention Models (SIM)*! Let’s bring together state-of-the-art ideas on modeling novel interventions and distribution shifts. :)
🙌🏻 Submissions are welcome! Link: sites.google.com/view/sim-ic…