MiroThinker: The Third Scaling Dimension for AI Agents

AI scaling has focused on two dimensions: model size and context length.¹ MiroThinker introduces a third: interaction depth. The research agent, released with 8B, 30B, and 72B parameter variants, trains models to handle up to 600 tool calls per task through reinforcement learning.² On the GAIA benchmark, the 72B variant achieves 81.9% accuracy, approaching commercial systems like GPT-5-high while remaining fully open-source.³

TL;DR

MiroThinker explores "interaction scaling" at the model level, systematically training models to handle deeper and more frequent agent-environment interactions.⁴ Unlike test-time scaling in isolation, interaction scaling uses environment feedback to correct errors and refine trajectories.⁵ With a 256K context window, the agent performs up to 600 tool calls per task, enabling sustained multi-turn reasoning for complex research workflows.⁶ Training uses three phases: supervised fine-tuning, preference learning, and reinforcement learning with group relative policy optimization.⁷ Across GAIA, HLE, BrowseComp, and BrowseComp-ZH benchmarks, MiroThinker demonstrates consistent improvements as interaction depth increases.⁸

The Agent Scaling Problem

Current AI agents face a fundamental limitation. As reasoning chains extend, errors compound.⁹ A single mistake early in a trajectory can derail the entire task. Traditional approaches address this through:

Larger Models: More parameters for better single-step accuracy¹⁰ Longer Context: More space to hold reasoning history¹¹ Better Prompting: Improved instructions to reduce errors¹²

Yet these interventions don't address the core issue: agents operating in isolation from their environment during extended reasoning.

Reasoning Drift

Long reasoning chains without environmental feedback exhibit "reasoning drift"—gradual divergence from correct trajectories.¹³ The agent continues reasoning based on increasingly stale or incorrect assumptions.

The Feedback Solution

MiroThinker's insight: let the environment correct the agent continuously.¹⁴ Rather than reasoning in isolation, the agent checks its work by interacting with external tools, catching errors before they compound.

Interaction Scaling Defined

Interaction scaling treats the depth of agent-environment interaction as a scalable dimension analogous to model size or context length.¹⁵

The Three Dimensions

Why Interaction Differs

Unlike model size (fixed at training) or context (passive storage), interaction depth enables active verification and course correction.¹⁶

Passive Scaling: Larger models and contexts provide more capacity Active Scaling: More interactions provide more opportunities to check, correct, and refine

The agent can discover that a web search returned outdated information, that a file path doesn't exist, or that a calculation produced unexpected results—and adapt accordingly.

MiroThinker Architecture

The agent follows the ReAct framework with specific enhancements for deep interaction:¹⁷

Core Loop

Thought → Action (Tool Call) → Observation → Thought → ...

Each observation feeds back into the agent's context, informing subsequent reasoning.¹⁸

Tool Suite

MiroThinker includes a comprehensive toolkit:¹⁹

600 Tool Calls

The 256K context window supports up to 600 tool calls per task.²⁰ For context, most agent benchmarks involve fewer than 20 tool calls. MiroThinker operates at 30x the typical interaction depth.

Training Methodology

MiroThinker training proceeds in three phases:²¹

Phase 1: Supervised Fine-Tuning

Initial training on successful agent trajectories teaches basic tool use patterns:²²

• When to search vs. browse
• How to formulate effective queries
• Interpreting tool outputs
• Synthesizing multi-source information

Phase 2: Preference Learning

The model learns to prefer successful trajectories over failed ones:²³

• Binary feedback on trajectory outcomes
• Implicit learning of error recovery
• Preference for efficient tool sequences

Phase 3: Reinforcement Learning

Group Relative Policy Optimization (GRPO) trains for extended interaction:²⁴

• Rewards for correct final answers
• Implicit credit assignment across long trajectories
• Learning when to persist vs. pivot strategies

Base Models

MiroThinker builds on open-weight foundations:²⁵

Benchmark Performance

GAIA (General AI Assistants)

GAIA tests realistic assistant tasks requiring web search, reasoning, and multi-step problem solving:²⁶

MiroThinker approaches commercial performance while remaining fully open.

HLE (Humanity's Last Exam)

Extremely challenging questions across diverse domains:²⁷

BrowseComp

Complex web browsing and information synthesis:²⁸

The Chinese performance suggests strong multilingual transfer.

Scaling Behavior

Critical finding: performance improves predictably with interaction depth.²⁹

As MiroThinker engages in more tool calls: - Accuracy increases (up to hardware/context limits) - Error recovery becomes more effective - Complex tasks become tractable

This demonstrates that interaction depth exhibits true scaling behavior, not merely diminishing returns.

Comparison to Other Approaches

vs. Chain-of-Thought

vs. ReAct Agents

vs. Test-Time Scaling

Why Interaction Scaling Works

The paper identifies several mechanisms behind interaction scaling's effectiveness:³⁰

Error Detection

More tool calls create more opportunities to discover mistakes:³¹

• Contradictory search results reveal incorrect assumptions
• Failed operations expose invalid states
• Unexpected outputs trigger reconsideration

Information Acquisition

Extended interaction gathers more relevant information:³²

• Follow-up searches refine understanding
• Multiple sources enable cross-validation
• Deep browsing uncovers hidden details

Strategy Refinement

Long trajectories allow strategy evolution:³³

• Initial approaches can be abandoned
• New angles can be explored
• Synthesis can incorporate late-arriving information

Open Source Release

The MiroMind team released comprehensive resources:³⁴

Models

Code

• Full training pipeline
• Inference implementation
• Tool integration examples
• Evaluation scripts

Documentation

• Architecture details
• Training recipes
• Deployment guides
• Benchmark reproduction

Implications for Agent Development

Training Paradigm Shift

Effective agents may require specific training for deep interaction, not just better base models.³⁵

Benchmark Evolution

Current agent benchmarks may prove too easy for interaction-scaled agents.³⁶ MiroThinker's 81.9% on GAIA suggests the need for more challenging evaluations.

Resource Requirements

600 tool calls per task implies significant compute and latency:³⁷

• Extended inference time
• Multiple external API calls
• Large context window maintenance

Production deployment must account for these costs.

Key Takeaways

MiroThinker establishes interaction scaling as a viable third dimension for AI capability:

1. New Dimension: Interaction depth scales like model size and context length
2. 600 Tool Calls: Trained for 30x typical agent interaction depth
3. 81.9% GAIA: Approaches commercial performance while fully open
4. Three-Phase Training: SFT → Preference Learning → RL pipeline
5. Error Correction: Environmental feedback prevents reasoning drift
6. Open Release: Models, code, and training recipes all available

The next generation of AI agents may prove capable not through larger models alone, but through deeper engagement with their environments.

References