• dLLMs hallucinate in fundamentally different ways than autoregressive models. The first controlled comparison identifies three unique failure modes (premature termination, incomplete denoising, context intrusion), meaning existing detection tools need redesign.

• Contrastive mutual information cuts process reward labeling cost by two orders of magnitude. Step-level signals extracted directly from model probabilities, no repeated rollouts needed. Accepted at ACL.

• RAG knowledge base defense shifts from static rules to runtime adversarial games. Canary tokens borrowed from stack canary concepts enable continuous detection, plug-and-play with no architecture changes.

• TorchUMM unifies mainstream multimodal models into one codebase. Covers understanding, generation, and editing, enabling the first apples-to-apples comparison across architectures.

Also Notable

• Hierarchical Analogical Reasoning Replaces Rule Matching for Content Moderation — Analogies handle gray-area cases more flexibly than hard rules.
• Chain-of-Analogy Counters Decision Shortcuts in Moderation — Companion paper to CHAIRO above, using DPO to strengthen analogical reasoning quality.
• Strip Textures, Keep Wireframes, Test VLM Geometric Understanding — Checks whether models truly understand spatial structure or just read texture cues.
• Multi-Agent Structured Reasoning for Legal Consultation — Includes a large-scale Chinese legal QA dataset.
• 2.5M Spatially Aligned Samples for Remote Sensing Multimodal Pretraining — Semantic supervision for geospatial foundation model pretraining.
• LLM Code Summaries Are Getting Longer, Evaluation Can't Keep Up — Reference-free fine-grained factual consistency evaluation.
• Teaching Navigation Agents to Recognize Nonexistent Targets — Handle false-premise instructions instead of searching blindly until timeout.
• Unsupervised Domain Adaptation for Low-Light Pose Estimation — No annotated dark-scene data required.

Read the full edition →