🥳 Introducing AEPO, an entropy-balanced agentic RL algorithm! 🙌 AEPO achieves diverse rollout sampling and prioritized learning of high-entropy tokens by balancing entropy. 📈 Impressive results: GAIA (65%), HLE (26%), and Webwalker (70%) on Pass@5! https://t.co/bOjHOAGngH

V-Thinker Interactive Thinking with Images

An interesting exploration of deeply combining visual tools with reasoning!🥳

Tencent WeChat and BUPT introduce V-Thinker A general-purpose multimodal reasoning assistant that enables "Interactive Thinking with Images" through end-to-end reinforcement learning. It actively edits, annotates, and transforms images to solve complex problems.

11 New Policy Optimization techniques ▪️ BAPO (BAlanced) ▪️ Training-Free GRPO ▪️ ASPO (Asymmetric Importance Sampling) ▪️ ICPO (In-Context) ▪️ GEPO (Graph-Enhanced) ▪️ IGPO (Information Gain-based) ▪️ AEPO (Agentic Entropy-Balanced) ▪️ AT-GRPO (Agent- and Turn-wise) ▪️ DGPO

DeepAgent: A General Reasoning Agent with Scalable Toolsets @XiaoxiLi0111 et al. introduces a deep reasoning agent that autonomously thinks, discovers tools, and executes actions within a unified reasoning process. 📝 https://t.co/4S4AMaUBDh 👨🏽‍💻 https://t.co/9OVAW80AqO

Glad to share DeepAgent! 🔍Auto Tool Search​​: Dynamically finds & uses right tools on the fly. 🧠​​Memory Folding​​: Brain-inspired memory condenses progress for efficient reasoning restarts. 🛠️​​ToolPO​​: RL training w/ fine-grained credit for tool use via a safe LLM simulator.

DeepAgent introduces Autonomous Memory Folding & ToolPO for efficient RL training. It excels on 8 benchmarks, including tasks with 16,000+ APIs! Experience DeepAgent in action: 🎥 https://t.co/SCslvSMBPJ Paper:

DeepAgent: A General Reasoning Agent with Scalable Toolsets https://t.co/PdTI0EJnLF

Demo: 🚀 DeepAgent showcases three powerful capabilities: • 16,000+ RapidAPIs for general tasks • Navigation & embodied AI in interactive environments • Deep research with web search, page browsing, code execution, ...

Interesting work!🥰

Thank your for sharing our DeepAgent🤩

1. DeepAgent: A General Reasoning Agent with Scalable Toolsets 🔑 Keywords: DeepAgent, tool discovery, action execution, autonomous memory folding, reinforcement learning 💡 Category: Knowledge Representation and Reasoning 🌟 Research Objective: - The paper introduces

DeepAgent uses an autonomous memory folding mechanism: 1. Compresses past interactions into structured memories. 2. Reduces error accumulation. 3. Preserves critical info for long-horizon interactions.

DeepAgent: A General Reasoning Agent with Scalable Toolsets Introduces ToolPO, an end-to-end RL strategy for teaching general-purpose tool use in LLMs. https://t.co/FQUZSZZ8ay

DeepAgent: The AI That Thinks, Learns, and Remembers — A Paradigm Shift for Autonomous Agents...... https://t.co/YIKYSA8D2t

DeepAgent is a new AI that figures out how to use tools all by itself to solve complex tasks. It even has a smart memory system to handle super long projects without getting confused. Details below 👇

DeepAgent A General Reasoning Agent with Scalable Toolsets

🚨BREAKING: Researchers just built an AI that teaches itself how to use new tools. It’s called DeepAgent, and it might be the first real step toward autonomous reasoning and might be the first real AGI prototype. Here’s why it’s a massive deal👇

Thank you for sharing our work!🥰

The Results Are Ridiculous Across 8 benchmarks ToolBench, TMDB, Spotify, WebShop, GAIA, HLE DeepAgent-32B beats ReAct-GPT4o and DeepSeek-R1 by huge margins. In open-set tasks (10K+ tools), it’s the first model to stay stable. This isn’t a demo it’s the first scalable general