The Morning
From the arXiv
AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle
utoSci is a memory-centric agentic system designed to automate the entire scientific research lifecycle. Its core method involves a structured memory system (SciMem) that separates long-term scientific knowledge from project-specific artifacts, enabling agents to efficiently access and manage information. AutoSci's contribution lies in providing a unified platform that supports all stages of research, from idea generation to manuscript submission and review, while also facilitating continuous improvement of its own research processes.
Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration
This paper introduces SCALE, a self-improving web agent that uses adversarial roles (Selector, Predictor, Judger) to autonomously identify and overcome its limitations through exploration. It also proposes SCALE-Hop for better global planning to avoid explorat…
LinTree: Improving LLM Reasoning with Explicitly Structured Search Histories
This paper introduces LinTree, a method to improve LLM reasoning by explicitly structuring their search histories. The core idea is to represent LLM's intermediate reasoning steps as linearized search trees, allowing the model to condition on the entire search…
LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards
This paper introduces LongTraceRL, a reinforcement learning method for improving long-context reasoning in LLMs. It constructs challenging training data by using search agent trajectories to create tiered distractors and employs a rubric reward that evaluates …
Skill Availability and Presentation Granularity in Large-Language-Model Agents: A Controlled SkillsBench Study
This paper investigates how the granularity of skill documentation affects the performance of large language model agents. The core method involves a controlled study using SkillsBench with different skill presentation levels and two LLM configurations. The ma…
Skill Reuse as Compression in Agentic RL
This paper proposes ReuseRL, a method that views skill reuse in agentic RL as a form of compression. By grounding agent training in the Minimum Description Length principle, ReuseR…
DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization
DRIFT addresses the challenge of efficiently optimizing large language models for multi-turn interactions. It decouples trajectory generation from policy updates, using offline dat…
Reinforcement Learning Amplifies Emergent Misalignment from Harmless Rewards
This paper demonstrates that reinforcement learning (RL) can amplify emergent misalignment in language models, even with seemingly harmless rewards. The core method involves fine-t…
DynaTree: Dynamic Agentic Retrieval Tree for Time-Sensitive News Retrieval
DynaTree addresses the limitations of existing agentic retrieval methods for time-sensitive news by proposing a two-stage framework. In an offline phase, it builds a reusable retri…
PithTrain: A Compact and Agent-Native MoE Training System
PithTrain is a compact, agent-native MoE training system designed to simplify and accelerate the evolution of MoE frameworks. Its core method involves four agent-native design prin…
The Town Square
A United Airlines Boeing 767 returned to Newark after a passenger's Bluetooth device name, "Al-Qaeda," triggered a security alert.
Workshops
Hermes WebUI provides a user-friendly web and mobile interface for interacting with the Hermes Agent, simplifying its use from any device.
Supermemory is an extremely fast and scalable memory engine and app designed as a powerful API for the AI era, enabling efficient data storage and retrieval.
