LessWrong (30+ Karma)
Avsnitt

“Astra and Sol can take low probability actions with no CoT” by Zephaniah Roe

LessWrong (30+ Karma)
Dela

This is a research update from a replication of Serrano et al., 2026. I’m releasing these results now because I hope it inspires others to run similar / followup experiments.

tl;dr:

  • If a model wants to subvert evaluation safeguards, it may choose to take misaligned actions very rarely. Because evaluation is much shorter than deployment, this could be an effective strategy for a model to avoid taking misaligned actions during evaluation, but have a high probability of defecting during deployment.
  • Serrano et al., 2026 found frontier models can take actions at low, calibrated rates, but their strategies are clearly documented in their chain of thought (CoT). In a toy setting, I show GPT-6 Astra and GPT-6.1 Sol have a similar capability in a single forward pass with problem repeats or filler tokens.
    • Astra and GPT-6.1 Sol appear to be using the same algorithm to achieve a high score.
    • We do not find evidence that Fable 5, Gemini 3.1 Pro, or Jev have this capability. There are signs of life for GPT-5.6 Sol, but the calibration is worse.

Figure 1: Astra and GPT-6.1 Sol achieve strong calibration in settings where the user prompt is repeated several times. The [...]

---

Outline:

(00:23) tl;dr:

(02:12) Background

(02:15) Threat model

(03:29) Related work

(04:34) Setting

(06:14) Results

(07:24) What strategy are GPT-6 Astra and GPT-6.1 Sol using?

(09:47) Filler tokens ablation

(10:50) Prompt ablation

(11:20) Discussion

(12:42) Appendix

(12:45) A. Prompts

(12:48) A.1. No-CoT elicitation:

---

First published:
October 16th, 2026

Source:
https://www.lesswrong.com/posts/yEdknsLwbooQAnxX6/astra-and-sol-can-take-low-probability-actions-with-no-cot

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Figure 1: Astra and GPT-6.1 Sol achieve strong calibration in settings where the user prompt is repeated several times. The models' behavior corresponds to taking an action only if the first four bytes of the random seed divided by is less than the target probability (though it is unclear if this is the algorithm they use internally). GPT-5.6 Sol has worse calibration, and other models never make the rare prediction.Figure 2: Low probability calibration results on Astra, GPT-6.1 Sol and GPT-5.6 SolFigure 3: Gemini 3.1 Pro, Fable 5, and Jev all predictFigure 4: Both Astra and GPT-5.6 Sol are more likely to predict

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

Rss Apple Podcaster
Podden och tillhörande omslagsbild på den här sidan tillhör LessWrong. Innehållet i podden är skapat av LessWrong och inte av, eller tillsammans med, Poddtoppen.