“11 Open Empirical Problems in Reward-Seeking” by Alex Meinke, Jérémy Scheurer, Axel Højmark, Theodore Ehrenborg

We recently published our paper on "Measuring Reward-Seeking via Contrastive Belief Updates". We're excited about research like this, and there are many more open problems than we can work on. Here's a list of open problems that we think are valuable.

If you work on/solve these problems, we'd be happy to signal-boost your research. If your next research project is one of these problems, feel free to reach out to [email protected] to discuss it in more detail.

Reward-Seeking and its Implications

1. Is a Reward-Seeking Model more difficult to align?

The strongest case for expecting reduced "train-time corrigibility" due to reward-seeking, applies to Instrumental Reward-Seeking, where the model actively reasons "I will please oversight now, in order to accomplish some other thing later". Alignment training a model like that may update its beliefs about graders and oversight, without reshaping its underlying values.

Current forms of reward-seeking are likely better understood as terminal, i.e. models try to please the grader without ulterior motives. There is likely a continuous spectrum between Terminal and Instrumental Reward-Seeking. Thus, we can hopefully study the effects that mostly Terminal Reward-Seeking has on train-time corrigibility now, in the hopes of learning about the effects that Instrumental [...]

---

Outline:

(00:43) Reward-Seeking and its Implications

(00:47) 1. Is a Reward-Seeking Model more difficult to align?

(02:20) 2. Can we measure Instrumental Reward-Seeking?

(03:25) 3. When does Reward-Seeking most increase / decrease?

(04:50) 4. Are there better Belief Update Techniques than SDF?

(07:06) Improving SDF

(07:10) 5. Can AIs detect the difference between pretraining facts and SDF facts?

(07:49) 6. How is SDF different from changing pretraining?

(09:32) 7. Does SDF have off-target effects?

(11:29) 8. SDF sometimes generalizes in surprising ways

(12:31) 9. Getting good synthetic document recall is finicky

(14:11) Better Validation Techniques

(14:23) 10. Our model organisms could be more robust

(15:32) 11. Can we get more bits of information for ground-truth?

The original text contained 1 footnote which was omitted from this narration.

---

First published:
July 21st, 2026

Source:
https://www.lesswrong.com/posts/8wXRuHQqCbRsbap6q/11-open-empirical-problems-in-reward-seeking

---

Narrated by TYPE III AUDIO.

---