veggiebird/MATPO-single-agent-14b

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedLicense:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

veggiebird/MATPO-single-agent-14b is a 14 billion parameter causal language model developed by veggiebird, fine-tuned using the Multi-Agent Tool-Integrated Policy Optimization (MATPO) reinforcement learning framework. This model uniquely trains a single LLM to act as both a planner and worker agent, excelling at complex multi-turn, tool-integrated tasks like web search and question answering. It achieves an 18.38% relative improvement over single-agent baselines on benchmarks such as GAIA-text, WebWalkerQA, and FRAMES by mitigating context length bottlenecks and noisy tool responses.

Loading preview...

Overview of MATPO-single-agent-14b

veggiebird/MATPO-single-agent-14b is a 14 billion parameter model based on the Qwen3-14B-base architecture, fine-tuned using the Multi-Agent Tool-Integrated Policy Optimization (MATPO) framework. Developed by veggiebird, MATPO is a novel reinforcement learning approach that enables a single large language model to embody multiple specialized agent roles, specifically a planner and worker agent, through role-specific system prompts. This architecture addresses critical limitations of single-agent approaches in multi-turn, tool-integrated planning, such as context length bottlenecks and interference from noisy tool responses.

Key Capabilities & Features

  • Multi-Agent-in-One-Model: A single LLM is trained to perform both high-level planning (planner agent) and specific task execution (worker agent) with isolated contexts.
  • Enhanced Performance: Achieves an 18.38% relative improvement over single-agent baselines on complex benchmarks like GAIA-text, WebWalkerQA, and FRAMES.
  • Robust Training: Exhibits more stable learning curves and greater resilience to noisy tool responses compared to traditional single-agent methods.
  • Principled Credit Assignment: Extends GRPO with a theoretically grounded reward distribution mechanism across planner and worker rollouts.
  • Infrastructure Efficient: Eliminates the need for deploying separate models or additional rollout engines for multi-agent systems.
  • Tool Integration: Designed for tasks requiring external tools, such as web search, scraping, and analysis.

Ideal Use Cases

  • Complex Question Answering: Excels in scenarios requiring multi-step reasoning and external information retrieval.
  • Automated Web Research: Capable of planning search strategies, executing queries, and synthesizing results.
  • Multi-Turn Tool-Integrated Tasks: Applicable to any task where an LLM needs to interact with tools over multiple steps, such as code generation with execution feedback or scientific reasoning.
  • Resource-Constrained Multi-Agent Systems: Provides a performant multi-agent solution without the overhead of deploying multiple distinct models.