Papers

AI papers from arXiv and journals, each with a summary and why it matters · 7705 papers
Nov 2
PapersBuildersSelected13:26
SecPO Cuts Prompt Injection Attacks on Qwen3.6-27B

The key insight is that DPO-based defences anchor to an insecure reference model, and an injection-invariant oracle can be approximated in practice — a fix that goes beyond simply training on more attacks. Teams building agentic systems can fine-tune with the released SecPO model, loss and adversarial training data rather than relying on prompt filtering. The released checkpoint is based on Qwen3.6-27B, making it directly usable for open-weight deployments.

arXiv cs.AI@Sizhe Chen et al. (5)Source
PapersBuildersSelected13:19
Training-Free MoE Expert Pruning Study Compares Replacement Operators

The paper shows that a usually implicit implementation choice — whether to renormalize survivors' gate scales — can flip which pruning method wins, so benchmark tables built on one operator may not generalize. Engineers pruning MoE checkpoints get concrete guidance: match the operator to the router's training, and validate on math and code tasks rather than multiple-choice accuracy alone. Researchers should adopt the five-item reporting checklist when publishing pruning results.

arXiv: DeepSeek@Ezza Ali et al. (4)Source