←
Aaron Mueller365 posts
Aaron Mueller avatar

Aaron Mueller

@amuuueller

Asst. Prof. in CS at @BU_Tweets ≡ {Mechanistic, causal} {interpretability, NLP}

◷ Joined September 2015▧ 75 Photos and videos
Aaron Mueller avatar
Aaron Mueller@amuuueller·Nov 10

This work adds to the mounting evidence that we can use our understanding of LMs to improve them. This relates to recent work from @GoodfireAI (on mechanistic shaping) and @RunjinChen (on persona vectors), who (post-)train models using concept vectors.

We cut through the mystery around interpretability in training. A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

0:11
Aaron Mueller avatar
Aaron Mueller@amuuueller·Nov 10

A clever trick @themistoklisnk devised was to make the regularization strength dependent on the confidence of the model! The more confident, the more it’s penalized for moving outside the ambiguity space. This helps us prevent *under*confidence, and results in big improvements!

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Nov 10

We derive an ambiguity subspace by getting steering vectors for many ambiguity types, and allow models to freely move in this subspace during fine-tuning. Movements outside of it are penalized.

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Nov 10

Language models should be less confident when an input is ambiguous. We derive ambiguity vectors, and use them to improve confidence estimation in language models. We can also use them to encode these improvements into the model weights! New work led by @themistoklisnk!

LLMs are highly overconfident, even when they don’t know something. We used interpretability to improve LLM confidence estimates! We propose DART, which uses ambiguity steering vectors to guide fine-tuning, improving calibration across 7 models while preserving accuracy. 🧵

0:46
Aaron Mueller avatar
Aaron Mueller@amuuueller·Nov 3

Come be my colleague at BU CS! We’re hiring in AI, broadly construed. Apply by Dec. 1 for full consideration: academicjobsonline.org/ajo/j… We’ve intentionally built a strong community of language, analysis, and ML researchers over the last few years. Looking forward to growing it!

Post attachment
Zilu Tang (Peter) avatar
⟳ Aaron Mueller retweeted
Zilu Tang (Peter)@Zilu_Tang_Peter·Nov 2

Do LMs possess genuine introspection? Are they performing anomaly detection? Or causally bypassing signals they are introspecting on? Our new preprint sheds some light on this by studying LM introspection on signals like MLP neuron activation arxiv.org/abs/2610.38911

0:25
Sheridan Feucht avatar
⟳ Aaron Mueller retweeted
Sheridan Feucht@sheridan_feucht·Oct 28

Tomorrow at 9:10am, I'll be talking about two case studies of "How LLMs Read" @BlackboxNLP. What exactly do I mean by "reading"? Come to P3 tomorrow morning and find out! :)

Post attachment

🎉 Thrilled to announce the keynote speakers for BlackboxNLP 2026 @emnlpmeeting, with three incredible perspectives on interpretability: 🔍 Ivan Titov @iatitov 🔍 Sheridan Feucht @sheridan_feucht 🔍 Michael Hahn @mhahn29 Join us on October 29th! 🇭🇺

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Oct 25

Also, if you're interested in how we can make interpretability into a more rigorous science, do stop by our tutorial on Wednesday!

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Oct 25

Is it possible that LLMs can represent biases without verbalizing them? Using steering vectors, we find evidence from QA and hiring task settings that this is possible. Do chat with Keren online!

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Oct 25

How do LLMs perform machine translation? We posit and find evidence that they read the source into an interlingua (a language-agnostic latent feature space), and read from this space to produce the output. Jake sadly couldn't make it, so I'll be presenting!

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Oct 25

How do induction heads lead to in-context learning (ICL)? When ablating induction heads during pre-training (which was not easy to do, by the way!), we find that *abstractive* ICL still emerges! Come find Kerem at the poster session.

Post attachment
Aaron Mueller avatar
Aaron Mueller@amuuueller·Oct 25

I'm at #EMNLP2026 in Budapest! 🇭🇺 Come check out the cool work from our students, and come find me at the tutorials and workshops! More details and pretty posters for each paper below: 🧵

Post attachment
Aryaman Arora avatar
⟳ Aaron Mueller retweeted
Aryaman Arora@aryaman2020·Sep 23

New paper! 🫡 We introduce Matryoshka Attribution, a new attribution method which uses gradient descent to find which parts of a neural network are responsible for a behaviour. MAttr is #1 on the Mechanistic Interpretability Benchmark by a wide margin (2.9× the runner up).

David Klindt avatar
⟳ Aaron Mueller retweeted
David Klindt@klindt_david·Sep 11

**new paper: Superposition in Brains** What if Barlow was wrong? In his neuron doctrine (1972), he suggested to search for the 'trigger features' of each neuron to understand cortex. But what if that is the wrong coordinate system to understand neural representations? biorxiv.org/content/10.64898…

Post attachment