There's been a lot of excitement about pluralistic value alignment 🌈 — AI that reflects the full range of human perspectives But no formal way to benchmark whether we're actually making progress. 🤔 Introducing 𝐎𝐕𝐄𝐑𝐓𝐎𝐍𝐁𝐄𝐍𝐂𝐇. 🎉Accepted to #ICLR2026 1/n 🧵

This is a really lovely position piece, laying out a unified framework for using self-consistency as a training objective! Intuition pump: I sometimes need to decide whether to trust people who are smarter than me. One way I do this is by judging their self-consistency.

None of this would be possible without my amazing co-authors @belindazli, @LauraRuis, and @jacobandreas, alongside @CarlGuo866 , @HuLillian39250, @MehulDamani2, @EkdeepL, and @ishapuri101. This was a true team effort!

Paper: https://t.co/KpHHwnUeWo If you've been thinking about consistency (or think we're wrong), we'd love to connect.

We argue that unification under our framework buys you something concrete: shared data augmentation pipelines, shared optimization strategies, and shared evaluation metrics.

Take introspection, where models explain themselves 🤖🪞 This can be trained by: 1) Asking the model to describe its behavior on a large number of inputs 2) Get dataset D of output behaviors 3) Optimize φ(explanation, D) that rewards explanations consistent with behavior.

This unifies many previously siloed problems, including sycophancy, factual coherence, and more. But this also introduces new/emerging classes of *meta-model* capabilities, such as: self-description / introspection, self red-teaming, AI for science (see paper for more).

However, current training (SFT, RLHF, etc.) scores each response in isolation. We propose a framework for optimizing consistency directly across groups of related inputs and show how common failures and new capabilities fall naturally under this lens.

Consider sycophancy: An LLM’s response may change depending on user beliefs, even when those signals are irrelevant. Evaluated one interaction at a time, each response appears coherent (A). The issue only becomes clear when multiple interactions are considered together (B).

New paper: It's time to optimize for 🔁self-consistency 🔁 We’ve pushed LLMs to the limits of available data, yet failures like sycophancy and factual inconsistency persist. We argue these stem from the same assumption: that behavior can be specified one I/O pair at a time. 🧵

Very thoughtful post by @belindazli on training LLMs to be introspective! IMO especially insightful point: We could train for *bidirectional* consistency between model self-explanations and revealed behaviors.

Check out Belinda's insightful post! Really excited about model introspection for interpretability and training.

Can LMs learn to faithfully describe their internal features and mechanisms? In our new paper led by Research Fellow @belindazli, we find that they can—and that models explain themselves better than other models do.

New paper! Language has rich, multiscale temporal structure, but sparse autoencoders assume features are *static* directions in activations. To address this, we propose Temporal Feature Analysis: a predictive coding protocol that models dynamics in LLM activations! (1/14)

Do AI agents ask good questions? We built “Collaborative Battleship” to find out—and discovered that weaker LMs + Bayesian inference can beat GPT-5 at 1% of the cost. Paper, code & demos: https://t.co/lV76HRKR3d Here's what we learned about building rational information-seeking

Can LLMs reason like a student? 👩🏻‍🎓📚✏️ For educational tools like AI tutors, modeling how students make mistakes is crucial. But current LLMs are much worse at simulating student errors ❌ than performing correct ✅ reasoning. We try to fix that with our method MISTAKE 🤭👇

Outstanding paper 2🏆: Shared Global and Local Geometry of Language Model Embeddings https://t.co/DT7L520ELV

🚨Why can’t Transformers learn multiplication?🧮 Even with billions of params, models struggle with multi-digit multiplication. In our new work, we reverse-engineer two models: a standard fine-tuned (SFT), and an implicit chain-of-thought (ICoT) model to see why. Read on! 1/n 🧵

Super excited to be joining @GoodfireAI! I'll be scaling up the line of work our group started at Harvard: making predictive accounts of model representations by assuming a model behaves optimally (i.e., good old rational analysis from cogsci!)

🧵New paper! Why do AI Experts Disagree on Existential Risk? When it comes to existential risk, AI researchers disagree: @ylecun: "effectively 0% chance of human extinction" @romanyam: "99% chance of catastrophe" I surveyed 111 AI experts to understand why.