analiza aiqbench
This research investigates how specific training pipeline choices—including pretraining data, architecture design, and reinforcement learning strategies—influence the monitorability of Chain-of-Thought (CoT) reasoning in frontier models. The study identifies that while CoT monitorability has historically been high, recent advancements like GPT-6 Astra exhibit significant degradation in this metric under non-adversarial conditions alongside improved evasion capabilities. These findings are critical for safety engineers aiming to maintain scalable oversight and detect misalignment as model architectures evolve to prioritize reasoning without explicit thought traces.
What Affects CoT Monitorability Across The Model Lifecycle?

De la OpenAI Alignment Blog
FragmentHow pretraining, architecture, reinforcement learning, and model-generated data affect chain-of-thought (CoT) monitorability.
November 10, 2026 Jeffrey G. Wang*, Aiden Low*, Zora Che*, Holden Mui*, Harshit Sikchi* Yong Zheng-Xin, Tomek Korbak❦, Bill Stavroulakis, Alex Barron, Shibani Santurkar, Kristian Georgiev Micah Carroll†, Ben Sokolowsky†, Katherine Lee†, Jakub Pachocki† * Equal contribution · ❦ Work done while at OpenAI · † Equal advising
Effective control of AI agents relies on confidence in their monitorability, which has previously been defined as “the degree to which we can identify particular aspects of a model’s actions (such as if it will cause harm).” [1] As such, monitorability is a core pillar of our safety cases. [2] While monitorability may be achieved by looking at actions or activations, so far we have also relied on chain-of-thought (CoT) monitorability as a key component of our scalable oversight strategy.
[3, 4, 5] CoTs usually contain information about a model’s intent and plans for misbehavior, and help us detect misalignment to inform upstream mitigations.1 In internal and external deployment for certain models—such as cyber critical ones—we monitor our models’ full trajectories, including their CoT. [7]
CoT monitorability has generally remained very high since the development of reasoning models.
Extras afișat conform termenilor RSS ai publicației · 198 cuvinte · textul integral și figurile sunt pe site-ul publicației
Subiecte conexe

Securing AI Research Needs an Owner
LessWrong · 27 sept. 2026
A Methodology for Auditable Trustworthiness Levels in AI Lifecycle Governance
arXiv · 20 iul. 2026

Top 10 Best SBOM Tools in 2026 [Ranked & Scored]
Cybersecurity News · 13 oct. 2026
A Methodology for Auditable Trustworthiness Levels in AI Lifecycle Governance
arXiv · 24 sept. 2026