James Oldfield

Interested in interpretability and AI safety. Postdoc at Oxford

[James Oldfield @jamesaoldfield]

Excited to share that our work on Dynamic Safety Monitoring for Language Models is accepted at ICLR 2026!! Looking forward to chatting with people there :)

Thanks a lot to @philiptorr @ioannispatras @Adel_Bibi @FazlBarez !!

[James Oldfield @jamesaoldfield]

How can we efficiently monitor LLMs for safety? Strong monitors waste compute on easy inputs, but lightweight probes risk missing harms ⚠️

𝙏𝙧𝙪𝙣𝙘𝙖𝙩𝙚𝙙 𝙥𝙤𝙡𝙮𝙣𝙤𝙢𝙞𝙖𝙡 𝙘𝙡𝙖𝙨𝙨𝙞𝙛𝙞𝙚𝙧𝙨 (TPCs) address this by generalizing linear probes for dynamic monitoring! 💫

[James Oldfield @jamesaoldfield]

@Turn_Trout The OG strategy

[William MacAskill @willmacaskill]

I had a very fun and wid-ranging conversation with @campbellclaret and @RoryStewartUK for The Rest is Politics: Leading, including around what European countries should do to avoid disempowerment in the face of AI progress.

One of the most lively conversations I've had in many years! Link below.

[Shawn Im @shawnim00]

I’ll be in Brazil for ICLR! 🇧🇷

I’ll be talking about how we can use theory to interpret models during the Thursday morning poster session and afternoon oral session! (Oral in 201 A/B, P4-#4006)

Happy to talk about interp, theory, or other things! Send a DM!

[Ben Hayum @BenHayum]

1/ AI agents are increasingly powerful. Security has not yet caught up.

New from CNAS: our response to CAISI’s RFI on AI Agent Security, with @janet_e_egan and @CalebWithersDC. 🧵

[James Oldfield @jamesaoldfield]

@johnhewtt Indeed!! Thanks for the writeup :)

[Shawn Im @shawnim00]

Excited to share our recent work selected as an ICLR Oral!


We work towards answering how models learn to associate tokens and build semantic concepts. We find that early-stage features in attention-based models can be written as compositions of three basis features.

[James Oldfield @jamesaoldfield]

Paper: https://t.co/HJvyxlecRE

[James Oldfield @jamesaoldfield]

Excited to share that our work on Dynamic Safety Monitoring for Language Models is accepted at ICLR 2026!! Looking forward to chatting with people there :)

Thanks a lot to @philiptorr @ioannispatras @Adel_Bibi @FazlBarez !!

[James Oldfield @jamesaoldfield]

How can we efficiently monitor LLMs for safety? Strong monitors waste compute on easy inputs, but lightweight probes risk missing harms ⚠️

𝙏𝙧𝙪𝙣𝙘𝙖𝙩𝙚𝙙 𝙥𝙤𝙡𝙮𝙣𝙤𝙢𝙞𝙖𝙡 𝙘𝙡𝙖𝙨𝙨𝙞𝙛𝙞𝙚𝙧𝙨 (TPCs) address this by generalizing linear probes for dynamic monitoring! 💫

[James Oldfield @jamesaoldfield]

@crabshellman @philiptorr @ioannispatras @Adel_Bibi @FazlBarez Thank you Jiaxing!

[Fazl Barez @FazlBarez]

2 papers accepted at #ICLR 2026!

Congrats to @elkmf & @SimonSchrodi and @jamesaoldfield for the hard work!

[Arthur Conmy @ArthurConmy]

Our new @GoogleDeepMind paper studies novel activation probe architectures for classifying real-world misuse risks.

Our research has informed live deployments of probes in Gemini. 🧵

[Fazl Barez @FazlBarez]

🚨New AI Safety Course @aims_oxford!

I’m thrilled to launch a new called AI Safety & Alignment (AISAA) course on the foundations & frontier research of making advanced AI systems safe and aligned at @UniofOxford
what to expect 👇
https://t.co/r9YHS3XJhR

[James Oldfield @jamesaoldfield]

@yisongyue @Tsinghua_Uni @YueLabCaltech Amazing!! Well deserved @YueSong48287250 !!

[Tony Wu @TonyWu1105]

🚨New paper!
Excited to share my paper with @fazlbarez on Query Circuit Discovery! 𝙒𝙚 𝙪𝙣𝙘𝙤𝙫𝙚𝙧 𝙖 𝙨𝙥𝙖𝙧𝙨𝙚 𝙘𝙞𝙧𝙘𝙪𝙞𝙩 𝙞𝙣𝙨𝙞𝙙𝙚 𝙖𝙣 𝙇𝙇𝙈 𝙩𝙝𝙖𝙩 𝙖𝙘𝙩𝙪𝙖𝙡𝙡𝙮 𝙙𝙧𝙞𝙫𝙚𝙨 𝙞𝙩𝙨 𝙧𝙚𝙨𝙥𝙤𝙣𝙨𝙚 𝙩𝙤 𝙖 𝙪𝙨𝙚𝙧 𝙦𝙪𝙚𝙧𝙮.

[James Oldfield @jamesaoldfield]

@dmkrash Hi Dima!! Thanks a lot for sharing both of these--looks like great work. We'll take a detailed read and update the pre-print to discuss :)

[James Oldfield @jamesaoldfield]

How can we efficiently monitor LLMs for safety? Strong monitors waste compute on easy inputs, but lightweight probes risk missing harms ⚠️

𝙏𝙧𝙪𝙣𝙘𝙖𝙩𝙚𝙙 𝙥𝙤𝙡𝙮𝙣𝙤𝙢𝙞𝙖𝙡 𝙘𝙡𝙖𝙨𝙨𝙞𝙛𝙞𝙚𝙧𝙨 (TPCs) address this by generalizing linear probes for dynamic monitoring! 💫

[James Oldfield @jamesaoldfield]

Please find many more results on 4 LLMs (across base models, instruction-tuned models, and reasoning models), and ablations in the paper!

📰 Project: https://t.co/urC7CPTZAO
💻 Code: https://t.co/1nBL6XXP3w
📄 Paper: https://t.co/HJvyxldF26

[James Oldfield @jamesaoldfield]

A big thank you to the fantastic coauthors! @philiptorr, @ioannispatras, @Adel_Bibi, @FazlBarez!