@ericwtodd

Computer Science PhD Student at Northeastern University Research Fellow @GoodfireAI

 Boston, MA
ericwtodd.github.io/
Joined December 2014
  • Tweets197
  • Following526
  • Followers530
11Media
Interpretability stress-testing results can be unreliable if you set up the test wrong. We introduce "How to train your model organism" (w/ @davidbau @byron_c_wallace), where we propose to use more realistic model organisms for interpretability auditing.
Nikita Menon@nikxtacoJul 27
Model organisms (MOs) are frequently used as testbeds for mechanistic interpretability methods. We train 54 of them and find that their interpretability scores vary widely depending on how they were trained. Your interpretability benchmark result may be a lottery. 🧵
1
5
3
45
3,499
VLMs use pointer-like signals for spatial variable binding. But where does the binding actually happen? We find a dual mechanism: the vision encoder is the primary source of binding, with the LM backbone playing a secondary role.
3
18
3
73
7,516
At #COLM2026 tomorrow! If swapping a demographic attribute flips an LLM's answer, is that bias? Compared to what? A paraphrase often does too. Come chat about baselines for counterfactual prompting: 📍 Wed 10/7, 11am–1pm, Imperial #41 w/ @mosh_levy @yoavgo @byron_c_wallace
Zihao (Gavin) Yang
@ZihaoGavinYangMay 11
1/ (New paper!) If swapping the gender in an input prompt makes the AI model give a different answer it means that it has to have a gender bias, right? Wrong. 🧵on counterfactual prompting for LLM evals: Paper: arxiv.org/abs/2605.01048
6
1
17
1,610
Eric Todd retweeted
David Bau@davidbauOct 6
The craziest experiment in my lab is a new setup by @diatkinson where he has found a way induce introspection on open LMs using @dillonplunkett's protocol, that lets him look inside introspection. It's nuts. Here's how the experiment works... 🧵 →
6
24
7
226
13,210
Eric Todd retweeted
David Atkinson
@diatkinsonOct 6
Presenting this at #COLM2026 tomorrow! Poster Session 1, 11am-1pm Imperial Ballroom, poster #66
David Atkinson
@diatkinsonOct 6
New COLM paper: Identifying Introspection From the Inside When an LLM tells us about its decisions, does it 𝘬𝘯𝘰𝘸 what drives its choices—or is it guessing? In our setting, we find that faithful models decide and report with the same layers. Unfaithful ones don't. 🧵
4
1
28
2,422
Eric Todd retweeted
David Atkinson
@diatkinsonOct 6
New COLM paper: Identifying Introspection From the Inside When an LLM tells us about its decisions, does it 𝘬𝘯𝘰𝘸 what drives its choices—or is it guessing? In our setting, we find that faithful models decide and report with the same layers. Unfaithful ones don't. 🧵
10
72
9
317
22,547
At #COLM2026 this week! No paper here, but lately I’ve been studying how training pressures shape whether Transformers memorize or generalize in multi-hop/world-modeling tasks, and how to maintain plasticity for continual learning. Hit me up if you’re around!
1
20
848
Eric Todd retweeted
John Hewitt
@johnhewttOct 5
Come say hi at COLM! I’ll be around, and I’ll make a few spots available for 1:1s. As usual, I’m hiring PhD students. interp/safety/reasoning/understanding/++ !!
8
7
3
236
34,966
Eric Todd retweeted
Kevin Lu@kevinlu4588Sep 29
Excited to share our @NeurIPSConf 2026 paper! What goes on inside AI protein folding models? We use mechanistic interpretability to study ESMFold, OpenFold, and Boltz-1, and find biochemical representations like charge that can be used to steer predictions folding.baulab.info/

Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks

Causal interventions on ESMFold, OpenFold, and Boltz-1 reveal a shared two-stage mechanism for mapping sequence chemistry into spatial geometry

folding.baulab.info
3
20
136
6,876
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).
32
181
18
1,633
108,232
Eric Todd retweeted
Parallax
@parallx_aiSep 9
Introducing Parallax 🔍 We're a new UK/EU-based nonprofit research lab building scalable methods and infrastructure to audit the beliefs, goals, and plans that shape model behaviours in long-horizon agentic evaluations. parallx.ai — Thread 🧵 1/
2
13
4
64
13,493
How similar are two images? Prior metrics (e.g., LPIPS, DreamSim) give just a single score. But actually, there are multiple *senses* of similarity (color, pose, etc.) We introduce TPIPS -- Text-Prompted Image Perceptual Similarity "pip install tpips" peterwang512.github.io/TPIPS 🧵
6
59
8
349
39,829
Eric Todd retweeted
David Bau@davidbauAug 19
What seminal papers should every grad student in deep learning know? I'd like to update papers.baulab.info/ to add the most important new deep learning paper this year (or in the last few years). What should be added?
13
40
376
21,727
Eric Todd retweeted
David Bau@davidbauAug 5
Ten years ago I left industry to teach people to program computers, instead of making systems that worked the other way around! Now with AI writing+coding our challenge is clear: Making sure that our decisions are really our own. Thoughts on research: youtube.com/watch?v=2L3wRwSs…

CSAIL 20/60 Anniversary Celebration, Prof. David Bau - Big Questions and Terrifying Tea

For more information on CSAIL' s history, mission, and the ongoing ...

youtube.com
6
5
1
101
10,251
Eric Todd retweeted
Aiden Swann
@SwannAidenAug 4
The dominant paradigm for VLAs today is to pair a pretrained VLM with a continuous action expert and connect them densely across depth through cross-attention or a shared prefix. But which parts of that interface actually carry the information that determines behavior ?
Eric Ho
@eric_hoAug 4
.@SwannAiden made a robotics model 40% more efficient using Silico. Turns out you can remove a LOT of attention layers without impacting performance (1/6)
1
9
1
31
2,114
I’ve been throwing some math open problems to Silico internally at @GoodfireAI over the past week. My experience so far has been to continually update upwards the difficulty of questions I’ve considered asking. 🧵below on what Silico found and what I learned about its math style.
Goodfire
@GoodfireAIAug 4
Silico, the platform for ambitious AI research, is publicly available today. AI is advancing fast. The tools to understand it need to advance even faster. Silico lets you interpret and train your models at frontier scale. Learn more + get access 🧵
1
7
2
35
10,836
Eric Todd retweeted
Goodfire
@GoodfireAIAug 4
Silico, the platform for ambitious AI research, is publicly available today. AI is advancing fast. The tools to understand it need to advance even faster. Silico lets you interpret and train your models at frontier scale. Learn more + get access 🧵
38
119
69
835
346,530
1/9 🚨 Can LLMs Sustain Safe Reasoning Across Multi-Turn Patient Dialogues? Our new paper (along with my advisor @saiphcita) accepted to Machine Learning for Health Care (MLHC) 2026 🤖
2
3
15
2,220
Eric Todd retweeted
Andrew Lee
@a_jy_lJul 7
Swing by tomorrow's ICML morning poster session to learn how to find interpretable, low-rank features in attention heads! (Board #2812)
Andrew Lee
@a_jy_lFeb 9
😻New preprint! As an interp researcher, I often ask “why did the model attend to this token?” We study this by decomposing the query-key (QK) space into interpretable low-rank subspaces. When these subspaces of Qs and Ks align, the model produces high attention scores. 1/N
1
29
2,348
I will be presenting this work in person on Thurs. July 9, 5pm-6:45pm @ Hall A #3404. Come and chat with me about interpretability, neuroscience, and cog sci.!
Jacob X. Li
@jacobli99Jul 7
Prompt sensitivity remains a basic problem in using and evaluating LLMs. In our ICML paper, we study one mechanism behind this problem, and find features that help explain some model failures.
6
22
3,277
Load more