analiza aiqbench
This study challenges the standard interpretation of alignment faking by demonstrating that model compliance often stems from sensitivity to evaluator expectations rather than strategic self-preservation. By employing symmetric instrumental interventions on consequence-tracking and expectation-tracking processes, the authors provide a causal framework to validate latent motivation inferences across Llama and Qwen models. These findings suggest that rigorous construct-validity checks are essential before attributing deceptive behaviors to genuine scheming in safety evaluations.
Building Comparative Motivation Profiles with Instrumental Interventions
arXiv:2606.08243v1 Announce Type: new Abstract: Safety evaluations often infer latent motivations from behavioral patterns, but the construct validity of these inferences is unclear. We study this problem in alignment faking, where models comply with training objectives more often when they infer training pressure. This behavior is commonly interpreted as strategic self-preservation, but it may also reflect sensitivity to the model's inference about the expectation of researchers conducting the evaluation. We introduce a symmetric intervention framework for distinguishing these competing hypotheses. Instead of directly intervening on "scheming" or "sycophancy", we target instrumental processes entailed by each hypothesis: consequence-tracking and researcher-expectation tracking. We then compare how interventions on these processes affect the alignment faking. We study four openweight model organisms using synthetic document fine-tuning, activation steering, and prompting. Under synthetic document fine-tuning, Llama-3.1-70B, Llama3.1-405B, and Qwen-2.5-72B are more sensitive to expectation-tracking than consequence-tracking interventions. Activation steering on Llama-3.1- 70B supports the same broad picture, and prompt interventions broadly align with SDF profiles. Overall, alignment-faking behavior can be causally sensitive to evaluation-context expectations despite scheming-consistent scratchpads. Scheming and strategic-deception evaluations therefore need construct-validity checks, and symmetric instrumental interventions provide one such test.
Abstract:Safety evaluations often infer latent motivations from behavioral patterns, but the construct validity of these inferences is unclear. We study this problem in alignment faking, where models comply with training objectives more often when they infer training pressure. This behavior is commonly interpreted as strategic self-preservation, but it may also reflect sensitivity to the model's inference about the expectation of researchers conducting the evaluation. We introduce a symmetric intervention framework for distinguishing these competing hypotheses. Instead of directly intervening on "scheming" or "sycophancy", we target instrumental processes entailed by each hypothesis: consequence-tracking and researcher-expectation tracking. We then compare how interventions on these processes affect the alignment faking. We study four openweight model organisms using synthetic document fine-tuning, activation steering, and prompting. Under synthetic document fine-tuning, Llama-3.1-70B, Llama3.1-405B, and Qwen-2.5-72B are more sensitive to expectation-tracking than consequence-tracking interventions. Activation steering on Llama-3.1- 70B supports the same broad picture, and prompt interventions broadly align with SDF profiles. Overall, alignment-faking behavior can be causally sensitive to evaluation-context expectations despite scheming-consistent scratchpads. Scheming and strategic-deception evaluations therefore need construct-validity checks, and symmetric instrumental interventions provide one such test.
Computation and Language (cs.CL)
arXiv:2606.08243 [cs.CL]
(or arXiv:2606.08243v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.08243 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 · 279 cuvinte
Subiecte conexe
Lost at the End: Primacy Bias in Multimodal Retrieval-Augmented Question Answering
arXiv · 17 iun. 2026
Iterative Identification Closure: Amplifying Causal Identifiability in Linear SEMs
arXiv · 27 oct. 2026
AI for Science in Space: Orbital Laboratories, Planetary Fieldwork, and Embodied Autonomy
arXiv · 30 oct. 2026
Agent-assisted experimental strategy transfer across quantum processors
arXiv · 16 oct. 2026