AI-generated summaries.
Research and policy news on artificial intelligence and robotics, updated three times a day from the public feeds of research laboratories, universities and EU institutions, with a daily selection of new research papers. Each item has a short summary and a link to the source.
MIT professor Paris Smaragdis describes his research group's work on machine listening, aiming to build systems that can hear and interpret sound at or beyond human capability. The group explores machine learning for sound source separation, smart audio-augmenting devices, and tools for understanding and creating music, with applications ranging from traffic accident detection to industrial equipment monitoring.
OpenAI announces an updated version of its Model Spec, the document that defines how its AI models should behave. The revised specification describes five core values intended to guide increasingly capable systems toward sound judgment, preservation of human agency, and faithfulness to human intent.
OpenAI describes its approach to detecting misalignment and unauthorized behavior in frontier models. The company says it uses layered monitoring systems designed to identify problematic agent actions so that human responders can intervene before incidents escalate.
MIT Department of Civil and Environmental Engineering head Ali Jadbabaie discusses how AI, climate change, aging grid infrastructure, and geopolitical complexity are simultaneously reshaping demands on physical infrastructure. He argues that meeting these challenges requires rethinking how infrastructure is designed, financed, and operated rather than repeating resource-intensive 20th-century models.
The authors propose Concordat, a Python library that enforces finite-state session-type protocols on LLM agents at runtime, restricting tool access to only those legal in the current protocol state and rejecting out-of-protocol calls before side effects occur. They report that on tau-bench airline tasks, the Concordat agent made zero unauthorized changes even under adversarial customer pressure, while a standard agent violated confirmation requirements in 13 to 29 of tasks depending on the model.
The author assesses empirical evidence for functional valence in LLMs, examining recent claims of functional pleasure, pain, and emotions. They argue that key aspects of valence as understood in philosophy and cognitive science remain untested, and that current experimental paradigms inadequately distinguish functional valence from neighboring constructs like preferences, value judgments, and role-play.
The authors introduce ThinkWorld, a VLA learning framework that represents action reasoning as a Chain-of-Dynamics, a sequence of discrete motion tokens capturing intermediate dynamics between current and future states, serving as a shared action representation across embodiments. They report that reinforcement learning with world imagination enables human videos without action labels to provide direct supervision, demonstrating effective human-to-robot transfer and generalization to unseen tasks in both simulation and real robots.
This paper addresses in-hand object pose reaching with a LEAP hand, training policies in simulation using diverse resets, generated grasp datasets, and an adaptive friction curriculum, then transferring them zero-shot to real hardware. A learned residual safety filter reduces drops by 75% while maintaining 91.5% goal success, with the state-based policy reaching 95.2% of goals across 4,800 unseen targets in 13.3 hours of evaluation.
The authors introduce CrossProfile, a benchmark of 60 templates yielding 480 tasks that disentangles fact ownership from usefulness in hidden-profile group decisions made by LLMs. Testing across multiple models reveals that both rarity-based and relevance-based memory heuristics fail predictably, and that removing or randomizing agent attribution tags raises accuracy by 13 to 24 points in classic configurations.
The authors propose Adaptive Residual Correction, a post-training module that applies reinforcement learning to learn residual corrections atop a frozen imitation policy, using sparse human interventions to localize when and where corrections are needed. Across three real-robot precision manipulation tasks, this approach improves average success from 33.3% to 75.0% while confining exploration to low-dimensional subspaces retaining over 96% of correction magnitude.