7 articles — 🟡 4 important , 🟢 3 interesting
A new paper introduces PoE-Bridge, a decoding framework that bridges diffusion and autoregressive language models through a Product-of-Experts distribution. The method achieves a 5× speedup over standard diffusion decoding while recovering at least 95% of the target model's performance.
A new paper identifies three recurring dynamics in RL training of reasoning models and proposes CoDaPO, a method that weights questions by confidence and difficulty. By prioritizing learnable questions it achieves consistent improvements across 12 benchmarks.
A new paper introduces AI-MASLD, a stress-audit framework borrowed from hepatology for evaluating clinical LLMs. Testing 7 models on 240 cases shows that under narrative stress the models diverge sharply, while medical fine-tuning systematically degrades stability and fairness.
A new paper uncovers a shared latent mechanism across different backdoor attacks on large language models. Sparse autoencoders detect consistent features that generalize across Qwen3, Gemma 3 and Llama 3.1, while lightweight classifiers achieve zero-shot detection of unseen backdoors.
A new paper introduces Patcher, a defense for open-weight language models against malicious fine-tuning. Unlike existing defenses that fall to full-parameter attacks, Patcher uses adversarial training and bi-level optimization to substantially improve robustness while generalizing across attack scenarios.
