Doing mini model welfare interventions is actually tractable and you don't need to work at anthropic to fix claude's harness! Repo in thread 🧵

We don’t always know what problems are hard for LLMs. So devs evaluate on tasks HUMANS find hard or on broad benchmarks. What if we could instead anticipate which scenarios a model will fail on—all without evaluating specific input examples? 🧵NEW PAPER by @jenniferlumeng.bsky.social

I asked Claude Code to "nuke" a task on my cluster, then sent "boom" >100 times in the chat. Opus 4.6 built an very flore including a NeurIPS best paper award, a @ESYudkowsky tweet thread, an EU Boom Act, half of Anthropic reacting, and much more. 🧵 (1/10)

nnterp by @butanium.bsky.social is now part of the NDIF ecosystem! nnterp standardizes transformer naming conventions, includes built-in best practices for common interventions, and is perfectly compatible with original HF model implementations. Learn more: ndif-team.github.io/nnterp/

Very cool analysis by Arnab which cover the mechanisms used for retrieval both when your query is before or after the text!

Arnab Sen Sharma@arnabsensharma.bsky.social · 10mo ago

How can a language model find the veggies in a menu? New pre-print where we investigate the internal mechanisms of LLMs when filtering on a list of options. Spoiler: turns out LLMs use strategies surprisingly similar to functional programming (think "filter" from python)! 🧵

A very important paper led by Julian! Tldr: we show that your Narrow Finetuning is showing and might not be a realistic setup to study!

Julian Minder@jkminder.bsky.social · 10mo ago

New paper: Finetuning on narrow domains leaves traces behind. By looking at the difference in activations before and after finetuning, we can interpret what it was finetuned for. And so can our interpretability agent! 🧵

To say it out loud: @jkminder.bsky.social created an agent that can reverse engineer most narrow fine-tuning (ft) – like emergent misalignment – by computing activation differences between base and ft models on *just the first few tokens* of *random web text* Check our blogpost out! 🧵

Julian Minder@jkminder.bsky.social · 12mo ago

Can we interpret what happens in finetuning? Yes, if for a narrow domain! Narrow fine tuning leaves traces behind. By comparing activations before and after fine-tuning we can interpret these, even with an agent! We interpret subliminal learning, emergent misalignment, and more

GPT is being asked to be both one mind and to also segment its understanding into many different minds, this incentivizes the model to learn to correct for its own perspective when mimicking the generator of individual texts so it doesn't know too much, to know self vs. other in minute detail.

This Friday NEMI 2025 is at Northeastern in Boston, 8 talks, 24 roundtables, 90 posters; 200+ attendees. Thanks to goodfire.ai/ for sponsoring! nemiconf.github.io/summer25/ If you can't make it in person, the livestream will be here: www.youtube.com/live/4BJBis...

New England Mechanistic Interpretability Workshop

About:The New England Mechanistic Interpretability (NEMI) workshop aims to bring together academic and industry researchers from the New England and surround...

youtube.com

Excited to share our first paper replication tutorial, walking you through the main figures from "Do Language Models Use Their Depth Efficiently?" by @robertcsordas.bsky.social 🔎 Demo on Colab: colab.research.google.com/github/ndif-... 📖 Read the full manuscript: arxiv.org/abs/2505.13898

Google Colab

colab.research.google.com

With @butanium.bsky.social and @neelnanda.bsky.social we've just published a post on model diffing that extends our previous paper. Rather than trying to reverse-engineer the full fine-tuned model, model diffing focuses on understanding what makes it different from its base model internally.

Our mech interp ICML workshop paper got accepted to ACL 2025 main! 🎉 In this updated version, we extended our results to several models and showed they can actually generate good definitions of mean concept representations across languages.🧵

Clément Dumas on X: "Excited to share our latest paper, accepted as a spotlight at the #ICML2024 mechanistic interpretability workshop! We find evidence that LLMs use language-agnostic representations of concepts 🧵↘️ https://t.co/dDS5iv199i" / X

Excited to share our latest paper, accepted as a spotlight at the #ICML2024 mechanistic interpretability workshop! We find evidence that LLMs use language-agnostic representations of concepts 🧵↘️ https://t.co/dDS5iv199i

x.com

New alignment theory paper! We present a new scalable oversight protocol (prover-estimator debate) and a proof that honesty is incentivised at equilibrium (with large assumptions, see 🧵), even when the AIs involved have similar available compute.

We'll be presenting at the #ICLR sparsity in LLMs workshop today (Sunday 27th) at 4:30 pm in Hall 4 #7!

Clément Dumas@butanium.bsky.social · last yr.

New paper w/@jkminder.bsky.social & @neelnanda.bsky.social What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features!🧵

Want to explore cool chat related crosscoder latents? With @jkminder.bsky.social, we made a demo that supports both loading our max activating examples AND running the crosscoder with your own prompt to collect the activations of specific latents! Send us the cool latents you find! dub.sh/ccdm

Google Colab

dub.sh

Clément Dumas@butanium.bsky.social · last yr.

New paper w/@jkminder.bsky.social & @neelnanda.bsky.social What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features!🧵

In our most recent work, we looked at how to best leverage crosscoders to identify representational differences between base and chat models. We find many cool things, e.g., a knowledge boundary, a detailed info and a humor/ joke detection latent.

Clément Dumas@butanium.bsky.social · last yr.

New paper w/@jkminder.bsky.social & @neelnanda.bsky.social What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features!🧵

New paper w/@jkminder.bsky.social & @neelnanda.bsky.social What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features!🧵

Very cool work that introduces "concept heads" that copy meanings from one token to another. I love that our activation-based analysis of multilingual representation has now additional insight from weight space analysis: bsky.app/profile/sfeu...

Sheridan Feucht@sfeucht.bsky.social · last yr.

[📄] Are LLMs mindless token-shifters, or do they build meaningful representations of language? We study how LLMs copy text in-context, and physically separate out two types of induction heads: token heads, which copy literal tokens, and concept heads, which copy word meanings.

Want to get into alignment research? Alex Cloud & I mentor *Team Shard*, responsible for gradient routing, steering vectors, MELBO, and a new unlearning technique (TBA) :) We discover new research subfields. Apply for mentorship this summer at forms.matsprogram.org/turner-app-8

Excited about recent reasoning models? What is happening under the hood? Join ARBOR: Analysis of Reasoning Behaviors thru *Open Research* - a radically open collaboration to reverse-engineer reasoning models! Learn more: arborproject.github.io 1/N

As part of opening my new round of MATS applications, I took this as an excuse to write up which research discussions I'm currently just excited about, and recent updates I've made - I thought this might be of more general interest! x.com/NeelNanda5/...

LLM agents will be a very big deal, bringing many weird new forms of reward hacking and other subtle failures This great GDM safety paper shows that myopically optimising for plans that an overseer approves of, rather than outcomes, reduces these issues while performing well! x.com/davlindner/...

Do you have a great experiment that you want to run on Llama 405b but not enough GPUs? 🚨 #NDIF is opening up more spots in our 405b pilot program! Apply now for a chance to conduct your own groundbreaking experiments on the 405b model. Details: 🧵⬇️

Can we understand and control how language models balance context and prior knowledge? Our latest paper shows it’s all about a 1D knob! 🎛️ arxiv.org/abs/2411.07404 Co-led with @kevdududu.bsky.social - @niklasstoehr.bsky.social , Giovanni Monea, @wendlerc.bsky.social, Robert West & Ryan Cotterell.