July 30 – August 5: 12 frontier papers out of 2,246 read.
Papers rated as moving the field
How we choose“Results show that skill extraction substantially reduces attack detectability: in SkillX, safety detection drops from 98.5% for poisoned trajectories to 11.4% for extracted skills, while Anything2Skill shows a similar effect.”
The authors, in the abstract
Persistent Skill Backdoors in Self-Evolving Agents
“We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes.”
“On match outcome they average 63.9%, level with backing the bookmaker's favourite -- which is in fact what they usually do.”
“No model scores more than 2.6% Pass^8 (GPT-5.6-Sol (Max+Pro)) and the highest Pass@8 is 21.5% (Muse-Spark-1.1 (xHigh)).”
“Extensive experiments show that DSA substantially improves robustness against white-box neuron-level safety attacks while preserving the model's general language and multimodal utility.”
“Our results demonstrate that when baseline defenses face template mismatches, often accompanied by severe degradation in downstream task performance.”
Jul 30 – Aug 5: 166 cleared, 12 frontier
The week’s papers have not been sorted into fields yet.
154 papers