PulseAugur
EN
LIVE 20:29:48
中文(ZH) 别光给Agent加Tool了,它根本选不明白!复旦×通义提出全新CUA训练范式

Fudan, Tongyi Lab unveil ToolCUA for agents choosing between GUI and tools

Researchers from Fudan University and Tongyi Lab have developed ToolCUA, a new training paradigm for agents that can effectively utilize both graphical user interface (GUI) operations and tool calls. Experiments revealed that simply equipping agents with tools does not automatically improve performance, as models often struggle to choose between GUI and tool actions, leading to decreased accuracy. ToolCUA addresses this by first synthesizing interleaved GUI-Tool trajectories and then employing online agentic reinforcement learning with a novel tool-efficient path reward to guide the agent in selecting optimal action paths. AI

IMPACT This new training paradigm could enable more capable agents that efficiently leverage both graphical interfaces and external tools, improving task completion and reducing errors.

RANK_REASON The cluster describes a new training paradigm and methodology for AI agents, presented in a research paper, with open-sourced code and model weights. [lever_c_demoted from research: ic=1 ai=1.0]

Read on 量子位 (QbitAI) →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

COVERAGE [1]

  1. 量子位 (QbitAI) TIER_1 中文(ZH) · Jay ·

    Stop just giving Agents Tools, they can't even choose them properly! Fudan x Tongyi proposes a new CUA training paradigm

    下一代CUA训练范式