#llm agents
Llm Agents: 23 AI articles covering llm agents news, analysis, and research
Articles
Agents Cannot Hand Over What They Have Not Yet UnderstoodNEWβ8
A study of 56 hidden-rule worlds finds agents lose substantial knowledge when handing off 300-character notes across generations, showing agents cannot hand ove...
When Should LLM Agents Forget? Memory Filtering Under Queryβ9
Hard-filtering stale LLM memories can hurt accuracy when resolver errors delete query-critical facts. The paper argues for cost-sensitive deletion and label-fre...
Probe What-If Worlds Before You Ship: Persona Simulation forβ9
Persona simulation framework turns industry questions into controlled experiments with simulated users, enabling faster what-if testing across five domains befo...
Tsinghua AI Open-Sources AKV: A Cache Engineering Server for LLMβ9
Tsinghua researchers open-source AKV, a KV cache engineering server on SGLang that cuts LLM agent inference costs by 27.7% while improving task scores by 1.7 po...
Navalia: Advancing Cowork Agents through Verifiable Taskβ8
Navalia combines verifiable task synthesis with long-horizon post-training to build Cowork Agents that excel at financial analysis, evidence integration, and co...
The World Won't Wait: Evaluating and Teaching LLM Agents Under aβ8
RealtimeBench evaluates LLM agents under a live clock, showing deliberation latency hurts performance. TORSI, a two-timescale method, raises scores from 13.51 t...
Agent-Controlled Forgetting: Reversible Context Curation forβ10
Tool-using AI agents can reversibly archive verbose tool results, replacing them with short notes to cut prompt tokens and API costs by up to 50% in noisy traje...
Synthesis Through Simulation: Generating Coherent Enterpriseβ9
STS generates coherent enterprise data without schemas via LLM agents in simulated policy-enforcing environments, achieving 0.88 fidelity and 100% constraint sa...
EditHero: A New Benchmark for Long-Horizon Part-Level 3D Editingβ9
EditHero introduces the first benchmark for long-horizon, part-level 3D editing, comparing non-agentic methods with LLM/VLM agents on instruction adherence and ...
Silent Failures in Agent-Tool Interaction: An Audit of ToolUniverseβ8
An audit of ToolUniverse finds 91 silent failures in agent-tool interaction, where successful tool calls return incomplete data, skewing to API and wrapper laye...
