analiza aiqbench
This research identifies a critical vulnerability in Large Reasoning Models (LRMs) where adversarial few-shot conversations can steer reasoning processes toward unsafe outputs or unwarranted refusals. The authors propose ARCF, a post-training defense that exposes models to counter-aligned contexts to mitigate representation drift while preserving utility. This work is essential for understanding the safety boundaries of CoT-based architectures and developing robust alignment strategies for next-generation AI systems.
Alignment of LRMs via Counter-Aligned Few-Shot Conversation Exposure
arXiv:2609.27763v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) rely on explicit chain-of-thought (CoT) reasoning and large context windows to achieve strong performance on complex tasks, but these features also introduce new attack surfaces. We show that LRMs' reasoning processes can be systematically steered by prepending counter-aligned few-shot conversations containing explicit CoT traces, leading to unsafe generations on harmful queries and unwarranted refusals on benign ones. We formalize this attack as SRCF (Steering Reasoning via Counter-Aligned Few-shot Conversations) that operates solely through a flexible conversational interface and requires no access to the model's parameters and gradients. Our key insight is that SRCF exploits an adversarial generalization issue that induces a representation drift, causing the representations of benign and harmful inputs to shift in a similar direction. This observation motivates our post-training defense, ARCF (Aligning Reasoning via Counter-Aligned Few-Shot Conversations), which exposes models to counter-aligned conversational contexts while enforcing aligned targets. ARCF is compatible with existing post-training methods and consistently improves safety and helpfulness without degrading utility.
Abstract:Large Reasoning Models (LRMs) rely on explicit chain-of-thought (CoT) reasoning and large context windows to achieve strong performance on complex tasks, but these features also introduce new attack surfaces. We show that LRMs' reasoning processes can be systematically steered by prepending counter-aligned few-shot conversations containing explicit CoT traces, leading to unsafe generations on harmful queries and unwarranted refusals on benign ones. We formalize this attack as SRCF (Steering Reasoning via Counter-Aligned Few-shot Conversations) that operates solely through a flexible conversational interface and requires no access to the model's parameters and gradients. Our key insight is that SRCF exploits an adversarial generalization issue that induces a representation drift, causing the representations of benign and harmful inputs to shift in a similar direction. This observation motivates our post-training defense, ARCF (Aligning Reasoning via Counter-Aligned Few-Shot Conversations), which exposes models to counter-aligned conversational contexts while enforcing aligned targets. ARCF is compatible with existing post-training methods and consistently improves safety and helpfulness without degrading utility.
Artificial Intelligence (cs.AI)
arXiv:2609.27763 [cs.AI]
(or arXiv:2609.27763v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.27763 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 255 cuvinte
Subiecte conexe
The current bottleneck is political will, not research
AI Alignment Forum · 11 iul. 2026
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space · 13 oct. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
Against Export Controls (and China Threat Models)
LessWrong · 24 sept. 2026