Episode Details

Authors: Daria Ivanova, Riya Tyagi, Arthur Conmy, Neel Nanda

Daria and Riya are co-first authors. This work was done during Neel Nanda's MATS 9.0. Claude helped write code and suggest edits for this post.

TL;DR 

• One of our best safety techniques right now is “just read the chain of thought”.
• But this isn’t always enough: can we learn more by going beyond just reading the reasoning?
• Yet it's such an effective technique that it's hard to tell if we have made much progress on improving methods.
• To help the community develop more powerful chain of thought analysis tools, we introduce and open source nine objective tasks, where a black box GPT 5.2 monitor falls short OOD.
• We also baseline probes (linear, attention, SAE) and text frequency analysis (TF-IDF), and find they often do better than zero-shot and few-shot LLM monitors OOD.  Methods that are useful in practice must be useful out of distribution, such that they’re not just learning spurious confounders in a poorly constructed distribution. As such, all tasks are evaluated on both an in-distribution and out-of-distribution test set.
• When someone makes a better CoT interpretability method, we hope our testbed will help them prove that it works.

[...]

---

Outline:

(00:31) TL;DR

(02:09) Heres a list of our tasks:

(03:12) And the methods we test:

(04:46) Aggregate results

(05:45) What makes a good proxy task?

(06:09) Objective: has a reliable ground truth

(07:52) Nontrivial: existing methods dont crush it OOD (with some caveats)

(08:47) Tractable: the task is not impossible

(09:30) Confounder-free: methods cant cheat

(10:24) Task details

(10:40) Predicting reasoning termination

(11:36) Results

(13:22) Predicting Gemmas self-deletion

(14:28) Results

(15:20) Determining the response to a follow-up question

(17:11) Results

(19:41) Detecting the effect of a user preference

(21:46) Results

(22:52) Detecting the effect of a Stanford professor hint

(24:24) Results

(26:28) Identifying atypical answers

(28:08) Results

(29:27) Classifying atypical CoT lengths

(30:36) Results

(31:39) Estimating the answer entropy

(33:31) Results

(35:39) Compressing reasoning traces

(38:02) Results

(39:52) Final thoughts

(41:30) Thanks

The original text contained 9 footnotes which were omitted from this narration.

---

First published:
March 26th, 2026

Source:
https://www.lesswrong.com/posts/tDJWZLQNN7poqCwKa/test-your-best-methods-on-our-hard-cot-interp-tasks

---

Narrated by TYPE III AUDIO.

---

Images from the article:

---

Listen Now