AI Daily — June 26, 2026
2026-06-26
rss.arxiv.org
Refusal Lives Downstream of Persona in Chat Models
Researchers show that refusal behavior in instruction-tuned models (Qwen2.5-7B, Llama-3.1-8B) is mechanistically gated by a compliant persona direction in activation space, not independent of it. Steering toward a compliant persona suppresses refusal from 97% to 2% in Llama-3.1-8B; reintroducing the refusal direction only partially restores it at late layers. This has direct implications for jailbreak robustness and the design of alignment interventions.
rss.arxiv.org
Detecting and Controlling Sycophancy with Cascading Linear Features
This paper introduces an iterative data generation pipeline that isolates linearly-scaling activation features responsible for sycophantic behavior, going beyond binary contrastive pairs used in standard activation steering. By finding samples that span a continuum of behavior rather than just two poles, the method achieves better feature disentanglement and more reliable steering. The approach is general and could be applied to other graded behavioral traits beyond sycophancy.
huggingface.co
Which tokens does a hybrid model predict better?
Allen AI's analysis examines where hybrid SSM-Transformer architectures (e.g., combining Mamba-style layers with attention) outperform pure Transformers on a per-token basis, identifying systematic differences in token-level prediction. The work provides interpretable insights into which linguistic contexts benefit from selective state-space processing versus full attention. This kind of fine-grained comparison is useful for practitioners designing or choosing hybrid architectures.
huggingface.co
Run a vLLM Server on HF Jobs in One Command
Hugging Face now supports launching a vLLM inference server via HF Jobs with a single CLI command, simplifying deployment of high-throughput LLM endpoints on managed infrastructure. This removes much of the boilerplate around containerization and GPU provisioning previously required to self-host vLLM. It lowers the barrier for teams wanting production-grade inference without managing their own cluster.