Excess Separability: Nuisance-Controlled Residual-Stream Probing for Benchmark Contamination Detection
Floriane C. M. Braun
A nuisance-controlled residual-stream probing method for detecting benchmark contamination in language models.
A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Floriane C. M. Braun
A nuisance-controlled residual-stream probing method for detecting benchmark contamination in language models.
Unknown
The paper argues that implicit bias in learning dynamics controls neural scaling laws during training, linking growth to scaling behavior from perceptrons to deep networks.
Unknown
ActSWM introduces action-sensitive world models for long-horizon planning in open-world games, improving latent model-predictive control.
Unknown
ChronoMem introduces version control and semantic rollback for LLM agent memory, enabling commit semantics and global rollback via natural language.
Kai Stewart, Yasunori Toshimitsu, Robert K. Katzschmann
A real-time Jacobian estimation controller enables an anthropomorphic hand to write with a grasped pen via purely in-hand motion, achieving sub-millimeter precision without simulation training or demonstrations.
Jiaming Zhong, Reza Valiollahi Mehrizi, Mohammad Pirani, et al.
LAMPC combines agent-based MPC with Gaussian process regression to predict unknown agent contributions, improving vehicle tracking and guaranteeing constraint satisfaction.
Dexter Ong, Vijay Kumar, Pratik Chaudhari
EVPeriscope uses an upward-facing event camera on a ground robot to detect a quadrotor's propeller signature for relative localization and control in degraded conditions.
Kaushalraj Puwar, B. Thangaraju
Adaptive Bridge is a proxy layer that decouples critical ROS 2 subscribers from degraded ones via topic splitting and dynamic rate control, cutting tail latency from 15 s to 1.55 ms.
Siting Li, Zhengyang Wang, Simon Shaolei Du, et al.
A controlled autoregressive testbed shows image tokenizers act as visual languages, with task-specific losses revealing distinct scaling and downstream links.
Unknown
This paper introduces a framework for reliable tool-using agents under communication constraints, focusing on control compression and reliability frontiers.
Geoffrey J. Barton
ALSCRIPT is a tool for formatting multiple sequence alignments, providing a flexible command language to control layout and annotation for publication-quality output.
Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee
A controlled, cost-aware comparison of LLMs and embedding models across 37 tasks finds them effectively tied in quality, but LLMs cost up to 1,431x more, supporting a division of labor.