LiveMACEBench: process-aware eval of LLM agents in live financial markets
Fudan-University · hf · 2026-10-19
Fudan University released LiveMACEBench, a benchmark evaluating persistent LLM agents in live crypto and US equity markets. Built on a Base ReAct agent, it scores decision traces across tool use, persistent memory, rule following and multi-agent collaboration using LLM judges plus objective metrics, producing capability profiles for GPT, DeepSeek, Gemini, Grok and Qwen — exposing gaps between outcomes and underlying capabilities.
More from coding & agent
- The Sane AI Code Review Loop: Two Agents, Automated Checks, Human Fallback — TheZachMueller · 2026-10-19
- Survey of 100 veteran Apple developers: 95% use AI nearly every day — op7418 · 2026-10-19
- Matt Pocock to speak at React Advanced on why software fundamentals matter in the AI era — mattpocockuk · 2026-10-19
- Obsidian Kanban Plugin 2.5.0 Cuts a 4,888-Card Board's Open Time From 9s to 3s — dSebastien · 2026-10-19
- RWA Pipe MCP Server Connects AI Agents to Tokenized Asset Data Across Blockchains — modelcontextprotocol · 2026-10-19
- Census Trade MCP Brings US Census Bureau International Trade Data to AI Agents — modelcontextprotocol · 2026-10-19