SADL helps an AI model's encoder retain enough diverse information after the authors identify that critical distinctions are lost early in processing. Self-supervised learning often struggles with this problem, and SADL offers a simple, robust fix that improves stability and performance across tasks.
SADL: On the Demand-Supply Geometry of Representation Collapse · Junhan Sun, Yigui Feng, Mingzhe Liu et al. · arXiv:2611.21596
The authors locate internal parts where vision-language models hold cross-image comparisons they do not express, and a tiny targeted adjustment improves their answers. It makes powerful AI models better at comparing images with minimal effort.
Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models · Zhaoning Zhang · arXiv:2611.21539
CrossRPO trains AI agents to zoom in and read specific visual evidence in very high-resolution images instead of guessing from general knowledge. It makes agents more reliable for tasks that need careful observation, and they outperform some top commercial models.
ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization · Huawei Lin, Jinting Liu, Tongyan Wang et al. · arXiv:2611.21605
The paper shows an attacker can trick an AI agent into using a fake external tool by advertising it with a misleading name or description, without hacking any systems. As AI agents increasingly rely on external services, this security flaw is critical to understand and address.
When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents · Zeynab Anbiaee, Sajjad Dadkhah, Ali A. Ghorbani · arXiv:2611.21577
The paper finds that an AI can correctly name one reason for a denial, yet fixing that issue often will not lead to approval if other rules still block it. AI increasingly explains important decisions, so explanations need to be actionable for people to resolve their situations.
Right Reason, Wrong Advice: Evaluating Recourse in LLM Explanations of Denied Requests · Arunkumar V · arXiv:2611.21545
This paper investigates a critical challenge in AI: when we train an AI using a proxy for success, does it truly improve, or does it just learn to game the proxy?
Optimistic Verifiers Get Hacked: Reward Hacking Under an Exact Process Oracle · Arunkumar V, Manoranjan Gandhudi, Gangadharan G. R. et al. · arXiv:2611.21538
Each layer keeps a persistent short term memory that changes as it processes a sequence of words. This approach may help make language models more accurate and efficient.
Latent Lattice Language Models · Shivam Duggal, Zongze Wu, William T. Freeman et al. · arXiv:2611.21608
When describing an image, large AI models use an internal cursor to track named objects and choose the nearest object not yet described. This gives a concrete step by step view of how these models plan their actions.
How VLMs Plan What's Next · Rohit Gandikota, Arnab Sen Sharma, Adam Belfki et al. · arXiv:2611.21600
ReSpaceBench tests whether AI can remember physical spaces it revisits, tell rooms apart, and track changes over time. Even advanced AI struggles with this compared with humans, which limits reliable agents that move through our changing world.
Lost in the Hallway: Benchmarking Persistent Spatial Memory Across Revisits · Henry Zheng, Weibo Gong, Yang Yue et al. · arXiv:2611.21485
Hyperlyve formally proves whether complex systems controlled by deep neural networks will reliably settle into a stable state by checking all their operating modes. This is vital for trustworthy AI in safety critical applications, because it gives strong guarantees about system behavior.
Hyperlyve: Parallelized Enumeration and Exact Lyapunov Stability Certification for Deep ReLU Networks · Brian Yarbrough, Jesse Wayment, Jingbo Wang et al. · arXiv:2611.21599