AuditBench exposes hidden alignment gaps

🔥 What's hot right now
AuditBench exposes a critical "tool-to-agent gap" where auditing tools fail against hidden behaviors. It reveals synthetic training makes models significantly harder to audit than demonstration-based training. ReMix also drops, solving the quality-speed trade-off in Diffusion LLMs with a training-free 2-8x inference speedup.

🚀 Just shipped
ReMix introduces a continuous mixing state for Diffusion LLMs. It resolves semantic inconsistencies during parallel decoding, achieving zero quality degradation with 2-8x inference speedups.

🛠 Useful for the array
AuditBench is essential for anyone auditing local models. It uses an autonomous investigator agent to test hidden alignment behaviors, helping you spot safety gaps that standard evaluations miss.

💬 Community pulse
The "Imagination" paper challenges the dominance of latent visual reasoning. It argues that teaching models to explicitly imagine via text (CapImagine) outperforms complex latent-space baselines, suggesting we might be over-engineering internal representations.

🐙 From TitanArray
(No items tagged "titanarray" this week.)