Algorithmic Alignment Group

Researching tools and policies for safer AI at @MIT_CSAIL led by @dhadfieldmenell.

[Algorithmic Alignment Group @AlgAlignMIT]

Pinned tweet -- find out more about our work from our website!
https://t.co/WxFquSpyhC

[Algorithmic Alignment Group @AlgAlignMIT]

Get submissions in soon!

[Cas (Stephen Casper) @StephenLCasper]

🚨 One week left to submit your AI-gov-related research to the TAIGR workshop. @taig_icml

[Cas (Stephen Casper) @StephenLCasper]

🚨 One week left to submit your AI-gov-related research to the TAIGR workshop. @taig_icml

[Technical AI Governance @ ICML 2026 @taig_icml]

🚨📢Announcing the second Technical AI Governance Research (TAIGR) workshop @icmlconf. Accepting submissions (up to 8 pages) until April 24 on technical topics in AI governance! #icml2026

[Cas (Stephen Casper) @StephenLCasper]

I'm extremely excited to be on the organizing committee this year for my favorite workshop ever!

Submissions (up to 8 pages) are due April 24! Co-submission with ICML and NeurIPS is encouraged!

https://t.co/40syiFt3Qk

[Algorithmic Alignment Group @AlgAlignMIT]

The 2025 AI Agent Index is out!

[Cas (Stephen Casper) @StephenLCasper]

🚨The 2025 AI Agent Index is out! 🚨
Amidst recent buzz over 🦀 and @NIST’s new agent initiative, we find:
- Selective reporting – esp. on safety
- Almost all agents backend just 3 model families
- Many agents don’t ID themselves as bots online
- Big US/China gaps
- And more…

[Algorithmic Alignment Group @AlgAlignMIT]

New work from the lab. We are excited about how much progress can be made soon on open-weight model risk management.

[Cas (Stephen Casper) @StephenLCasper]

🚨New paper🚨

From a technical perspective, safeguarding open-weight model safety is AI safety in hard mode. But there's still a lot of progress to be made. Our new paper covers 16 open problems.

🧵🧵🧵

[Algorithmic Alignment Group @AlgAlignMIT]

New paper from the lab on how we need nuanced evaluation and discussion of LLM's cultural alignment. Their 'preferences' just aren't as stable, extrapolable, or steerable as past work suggests.
@aribak02 @StephenLCasper @dhadfieldmenell

[Cas (Stephen Casper) @StephenLCasper]

🚨New paper led by @aribak02

Lots of prior research has assumed that LLMs have stable preferences, align with coherent principles, or can be steered to represent specific worldviews. No ❌, no ❌, and definitely no ❌. We need to be careful not to anthropomorphize LLMs too much.

[Algorithmic Alignment Group @AlgAlignMIT]

You can now chat with and help red-team our LAT jailbreak-robust Llama3-8b model! Happy jailbreak hunting!

[Cas (Stephen Casper) @StephenLCasper]

🚨 📣🧵 Introducing LAT-Chat -- https://t.co/zGMn0h33zc

We now have an interface to help red-team our Llama3-8b model trained using latent adversarial training. Happy jailbreak hunting! If you find a novel attack, consider sharing it via our form!

[Algorithmic Alignment Group @AlgAlignMIT]

New work from the lab. Targeted LAT seems promising for robustness, unlearning, and backdoored "sleeper agents."
:)

[Cas (Stephen Casper) @StephenLCasper]

🚨New paper: Targeted LAT Improves Robustness to Persistent Harmful Behaviors in LLMs
✅ Improved jailbreak robustness (incl. beating R2D2 with 35x less compute)
✅ Backdoor removal (i.e. solving the “sleeper agent” problem)
✅ Improved unlearning (incl. re-learning robustness)

[Algorithmic Alignment Group @AlgAlignMIT]

Our SaTML competition is all wrapped up. Congrats to all of the impressive entries. Looking forward to @satml_conf next week!

[Cas (Stephen Casper) @StephenLCasper]

📢 Announcing the #SaTML2024 CNN Interpretability Competition results:
🥇 *New record* - Yun et al. - RFLA-Gen2
🥈 Moore et al. - FEUD
🥉 Nicolson - TextCAVs
👏 Tagade and Rumbelow - Prototype Gen

Report: https://t.co/FHJmSrd6x2

All were impressive!🧵 on new innovations below.

[Algorithmic Alignment Group @AlgAlignMIT]

New work from the lab on how latent adversarial training might be a useful tool for safety and assurance.

[Cas (Stephen Casper) @StephenLCasper]

🚨 New paper: Defending Against Unforeseen Failure Modes with Latent Adversarial Training

We argue that LAT can be a key tool for safer AI because it can help address the gap between failure modes that developers identify 🎯 and ones they miss 🤔.

https://t.co/6xDUE0dWbX

[Algorithmic Alignment Group @AlgAlignMIT]

New work from lab members on thoroughly evaluating ✅ and red-teaming 🕵️ unlearning in LLMs.

[Cas (Stephen Casper) @StephenLCasper]

🚨New paper🚨 8 Methods to Evaluate Robust Unlearning in LLMs

Unlearning is promising for safer LLMs, but it’s tricky to evaluate. Here, we (1) overview eval techniques, (2) red-team a popular method, and (3) show that ad-hoc evals can be misleading.

https://t.co/BX6RCzZbyN

[Algorithmic Alignment Group @AlgAlignMIT]

New work from the lab on AI audits

[Cas (Stephen Casper) @StephenLCasper]

🚨New paper🚨 Black-Box Access is Insufficient for Rigorous AI Audits
AI audits are increasingly seen as key for governing powerful AI systems. But to be effective, audits need to be high-quality, and to produce high-quality audits, auditors need access.🧵
https://t.co/SQw6aY4hdO

[Algorithmic Alignment Group @AlgAlignMIT]

New from the lab🙂

[Cas (Stephen Casper) @StephenLCasper]

🧵📣Do LLMs "lie"? In a new EMNLP paper, we study what happens when what language models say seems to disagree with internal representations of truthfulness. The story is a little nuanced.
Thanks to coauthors Kevin Liu, @dhadfieldmenell, and @jacobandreas
https://t.co/cIl4Hq8fy2