📄 Summary: This paper tackles the challenge of detecting not just whether text is human or AI-generated, but specifically who created it and who edited it, using a four-class classification framework. The authors propose RACE (Rhetorical Analysis for Creator-Editor Modeling), which uses Rhetorical Structure Theory to distinguish the unique signatures left by creators versus editors in text.
💡 Key Insight: Different policies apply when an AI polishes human writing versus when humans try to hide AI-generated text, so we need to detect the creator-editor split, not just the human-AI binary.
📄 Summary: Long chain-of-thought reasoning in AI models is expensive and can actually hurt performance through "overthinking," so this paper proposes CoDE-Stop, which monitors how confident the model is in intermediate answers to decide when to stop reasoning. The key insight is that correct reasoning paths show confident answers early, while incorrect paths produce long, unproductive traces with unreliable confidence signals.
💡 Key Insight: You can save compute and improve accuracy by watching when a reasoning model becomes confidently sure, rather than letting it think indefinitely.
📄 Summary: This paper solves the memory bottleneck in long reasoning by proposing TriA, which compresses the key-value cache using insights from the geometry of pre-rotation (pre-RoPE) space. The method discovers that query and key vectors cluster around stable centers and attend to keys at specific distances—patterns that can be exploited trigonometrically to select the most important keys without losing reasoning quality.
💡 Key Insight: Pre-rotation space shows that attention has predictable geometric patterns you can compress, enabling longer reasoning with less memory.
📄 Summary: Vero is an open-source visual reasoning model family that matches proprietary systems by combining reinforcement learning across six diverse task categories with a 600K-sample dataset and task-routed rewards. This breaks the black box around how strong vision-language models are built, providing a reproducible recipe for visual reasoning across charts, science, and spatial tasks.
💡 Key Insight: You can build state-of-the-art open visual reasoners by scaling RL data strategically across task categories rather than relying on proprietary methods.