Tag
The paper proposes Density Feedback Control (DFC), which adaptively adjusts the sparsity coefficient during training of generator-based sparse adversarial attacks to hit a target perturbation density and improve transfer attack success rates on black-box CNN and Transformer targets.
该论文提出了一套评估神经PDE求解器中学习状态可复用性的契约框架,通过配对状态比较、匹配的目标信息与计算预算,分离了解精度、学习贡献与数值效用。在240条源训练轨迹和两类线性PDE上的实验表明,初始猜测的收益会随校正算法反转,说明复用价值需要严格的下游计算证据支撑。
该论文揭示了针对公开ASR基础模型生成的对抗扰动可有效迁移至微调后黑盒模型的风险,并提出TransferBreaker统一微调框架(结合Base Adversarial Fine-Tuning、Latent Jacobian Regularization与HybridGrad-AFT),在三种语言和四个大型ASR模型上将对抗WER从92.6降至27.8。
This paper rethinks how automatic harness evolution for agents should be evaluated, showing that gains may be due to increased compute rather than genuine improvements, and that evolved harnesses transfer poorly to unseen tasks.
This paper tests the assumption that LLMs judge better than they generate in in-context QA, finding generation accuracy exceeds self-evaluation on most benchmarks, with evaluation attending less to context. The findings challenge core assumptions in self-evaluation pipelines.
AdversaBench introduces an automated LLM red-teaming pipeline that uses five mutation operators and a three-judge panel with a meta-judge tiebreaker to confirm failures, revealing that attack difficulty varies by category and that adversarial prompts transfer from smaller to larger models.
Skill-MAS introduces a method for evolving meta-skills in multi-agent systems to improve orchestration without modifying model weights, achieving transferable performance gains across tasks and LLMs.
This paper analyzes residual scaling in looped (weight-tied) transformers, showing that weight sharing requires stronger scaling (1/N) than standard residual networks, and derives a factored parameterization that enables hyperparameter transfer across loop counts without retuning.