Authors: Gerson Kroiz*, Aditya Singh*, Senthooran Rajamanoharan, Neel Nanda

Gerson and Aditya are co-first authors. This work was conducted during MATS 9.0 and was advised by Senthooran Rajamanoharan and Neel Nanda.

TL;DR

Understanding why a model took an action is a key question in AI Safety. It is a difficult question, as often many motivations could plausibly lead to the same action. We are particularly concerned about the case of model incrimination, i.e. observing a model taking a bad action and distinguishing between benign and malign motivations, e.g. scheming vs confusion.

To practice the skill of understanding motivations, we need to build high-quality environments that serve as realistic proxy tasks for investigating future models. But we’ve found that there are many ways in which environments can be contrived. It is valuable to invest effort in ruling out confounds like ambiguity, instruction conflict, or confusion.

Over the course of building and iterating on more than 20 environments, we have refined 5 design principles for creating such settings. Generally, the more principles that an environment adheres to, the more fruitful and interesting the investigation will be.

Have uncertain causes: The environment should elicit behavior that admits either multiple plausible explanations [...]

---

Outline:

(00:32) TL;DR

(02:47) Introduction

(05:10) Design principles

(05:35) 1. Have uncertain causes

(07:59) 2. Maximize investigative surprise

(09:48) 3. Clear user intent

(11:38) 4. Avoid implicit nudges

(13:26) 5. Make the environment realistic

(15:03) Example of cleaning up an environment

(20:03) Appendix

(20:06) A1. Math Tutor Environment Details

(20:49) A2. Directory Summary Environment Details

(21:48) A3. Funding Email Environment Details

(22:38) A4. Evaluation tampering environment details

(23:20) A5. Secret number environment details

(24:06) A6. Sandbagging environment details

The original text contained 3 footnotes which were omitted from this narration.

---

First published:
March 2nd, 2026

Source:
https://www.lesswrong.com/posts/8pZuQnCve6K5ZrnM8/how-to-design-environments-for-understanding-model-motives

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Figure 1: Sandbagging rates for DeepSeek R1 and OpenAI o3 under baseline and intervention conditions designed to clarify user intent. Combined interventions reduce R1's sandbagging from 60.5% to 4.2%, and o3's from 31.6% to 0%, suggesting that sandbagging in both models is fully explained by instruction ambiguity rather than self-preservation or deployment-seeking.

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

“How to Design Environments for Understanding Model Motives” by gersonkroiz, aditya singh, Senthooran Rajamanoharan, Neel Nanda

LessWrong (30+ Karma)

“How to Design Environments for Understanding Model Motives” by gersonkroiz, aditya singh, Senthooran Rajamanoharan, Neel Nanda