The Morning
From the arXiv
Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning
rgus provides a general-purpose agentic runtime for long-horizon reasoning by employing a persistent, self-evolving architecture. It separates user intent from operational details and uses a review process for all new components, allowing agents to persist when successful and pivot when facing issues. This approach significantly improves performance on complex tasks like software engineering benchmarks compared to simpler methods.
Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite
This paper introduces HiGram, a hierarchical graph memory for LLM agents designed for long-term reasoning. Its core method organizes memories into a coarse-to-fine structure, reducing irrelevant information and improving retrieval efficiency. HiGram's key cont…
OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling
This paper introduces OctoLong, a method for creating large, dependency-rich code contexts by recursively retrieving code references. They then use these contexts to mid-train open-source language models, resulting in OctoLong-Instruct, which demonstrates enha…
When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs
This paper causally audits the claimed benefits of relaying key-value caches (instead of text) in multi-agent LLMs. The core method involves replacing the relayed cache with manipulated versions (deranged, zeroed, random) to isolate the cache's true impact. Th…
State2State: Environment-Derived Mid-Training for LLM Agents
State2State is a novel mid-training method for LLM agents that addresses the limitations of supervised fine-tuning and traditional RL. Its core method involves converting explored environment states into training objectives, challenging agents to reach specifi…
Skill-Use: Can LLMs Actually Use Skills in Agentic Harnesses?
This paper introduces Skill-Use, a benchmark to evaluate if LLM agents can effectively recognize and apply skills. The core method involves progressive disclosure, where agents mus…
A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination
A-SR introduces a novel self-evolving agentic framework for symbolic regression that moves beyond single-prompt approaches. Its core method involves coordinating specialized LLM ag…
ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
ABSeeker trains long-horizon search agents by assigning credit at a finer, step-level rather than trajectory-level. Its core method, Answer-Backtracked Credit Assignment (ABC), rec…
Agentic Reinforcement Learning with Observation-Calibrated Self-Distillation
This paper introduces Observation-Calibrated Self-Distillation (OCSD) to improve reinforcement learning for large language model agents. OCSD addresses a limitation of prior method…
ArtAnno: Annotating Implicit Semantics in Artworks through LLM Agent-Driven Bidirectional Human-AI Augmentation
ArtAnno introduces Bidirectional Human-AI Augmentation (BiHAA), a closed-loop framework for annotating implicit artwork semantics. It leverages a multi-agent LLM system where AI pr…
The Town Square
Demis Hassabis is stepping down as CEO of Google DeepMind to become its Chair, while Jeff Dean departs, as Google reorganizes its AI leadership to accelerate progress.
Workshops
Loopx is a lightweight kernel for long-running AI agent teams, offering agent-agnostic loop engineering with features like durable goals, quota-aware auto-wake, executable todos, evidence logs, and verifiable handoffs.
TencentDB Agent Memory acts as a team-level AI agent memory hub, transforming data into four reusable assets (Chat Memory, Skill, LLM-Wiki, Code-Graph) for governed sharing and cross-agent deployment.
