The Morning
From the arXiv
COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
OBRA-Skills optimizes LLM agent skills by treating it as a budgeted sequential optimization problem. It uses a contextual bandit to guide the selection of skills for evaluation, prioritizing promising or informative candidates. This approach efficiently refines the skill population using execution feedback, significantly reducing optimization costs and data requirements while achieving superior performance across diverse benchmarks.
Artificial Id: Drive and Persistent Alignment in Agentic AI
This paper introduces an "artificial id" as an internal, adaptive drive for agentic AI systems. This mechanism allows agents to autonomously decide when to continue, stop, or change their behavior, moving beyond externally defined control. The core contributio…
Autonomy, Social Norms, and Alignment: Towards a Developmental Framework for Autonomous Artificial Agents
This paper proposes a developmental framework for autonomous AI agents, moving beyond static datasets to enable learning through environmental interaction. Its core method involves incorporating intrinsic motivations like curiosity and competence to guide expl…
Domain-Specific Hallucination Detection in Large Language Models
This paper proposes a multi-signal pipeline for detecting hallucinations in LLM-generated text, combining a fine-tuned classifier, Monte Carlo Dropout for uncertainty, and temperature scaling. The core contribution is a robust detection method that achieves hi…
Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents
Ecdysis addresses the inefficiency of training LLM agent runtime harnesses by introducing principled failure diagnosis. Instead of simply iterating on observed failures, Ecdysis distinguishes between model-specific and harness-specific issues. This allows for …
LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
LOCUS is a task-aware, low-rank adaptation method for post-training large language models. It efficiently minimizes generation length by adapting only a small subspace of parameter…
MAPLE: Memory-Augmented Planning with Language and Evolution
MAPLE is a system that helps domain experts solve optimization problems using natural language. Its core method combines language understanding, mathematical programming, and evolu…
MindTopo: Can Foundation Models Reason in Topological Space?
This paper introduces MindTopo, a benchmark designed to evaluate foundation models' ability to reason about topological spatial relations, which are invariant under continuous defo…
RetroThinker: Enabling Retrospective Thinking in Speech LLMs
RetroThinker enables streaming SpeechLLMs to improve their reasoning by dynamically revising their thought processes during inference. It achieves this through a post-training fram…
SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control
SIRF addresses industrial content risk control by internalizing complex platform policies directly into model weights through continued pretraining. This "spec-internalization" all…
The Town Square
Anthropic's September 2026 report details emerging AI misuse threats and outlines strategies for detection and mitigation, focusing on evolving adversarial techniques.
Workshops
This repository offers a spy satellite simulator in your browser, displaying live, open-source spatial intelligence on a photorealistic 3D globe, making real-world data accessible.
Superpowers is an agentic skills framework and software development methodology designed to enhance productivity and effectiveness through a structured approach to building and utilizing AI agents.
