GRACE: Graph-Regularized Agentic Context Evolution for Reliable Long-Horizon LLM Agents
A new arXiv paper proposes GRACE, a method that maintains persistent system-level instructions for LLM agents as a typed semantic graph instead of flat text. This graph-regularized approach enables scoped verification and reliable context evolution over long horizons under distribution shift, addressing verification difficulties from accumulated instructions.
Entities
Related
Shared Selective Persistent Memory Architecture Proposed for Agentic LLMs
A new arxiv paper introduces shared selective persistent memory, an architecture that identifies and retains four categories of reusable context (task specifications, data schemas, tool configurations, and tool-use patterns) for agentic LLM systems. This approach aims to solve the context problem where each session starts from zero, while avoiding token-inefficient naive persistence of entire conversation histories. The work is relevant to developers building multi-turn code generation agents.
Study analyzes failure trajectories of CLI coding agents as temporal processes
A new arXiv paper presents the first large-scale empirical study of CLI coding-agent failure trajectories, treating failure as a temporal process rather than a final outcome. The study introduces a process-oriented framework to analyze how failures emerge, evolve, and become unrecoverable in LLM-based coding agents.
Researchers propose memory agent to combat behavioral state decay in long-horizon tasks
A new arXiv paper identifies 'behavioral state decay' as a failure mode where decision-relevant information gets buried in long trajectories. The authors propose a separate memory agent that actively updates a structured memory bank alongside an unmodified action agent, rather than relying on passive retrieval.
GATS framework eliminates LLM calls during agent planning inference
Researchers propose GATS (Graph-Augmented Tree Search), a planning framework that uses a layered world model and UCB1-based tree search to avoid LLM inference during planning, reducing computational cost and stochasticity. The approach outperforms LATS and ReAct on multi-step planning tasks.
AgentKGV: Agentic LLM-RAG Framework with Two-Stage Training for the Fact Verification of Knowledge Graphs
Researchers propose AgentKGV, an agentic LLM-RAG framework for verifying facts in knowledge graphs. It uses dynamic routing and iterative query rewriting to handle surface-form mismatches in document-level retrieval, and introduces a two-stage training process to improve accuracy and cost-efficiency for industrial deployment.