Aligning to What? Rethinking Agent Generalization in MiniMax M2

Summary

MiniMax M2 addresses the performance gap between agent benchmark scores and real-world deployment across diverse scaffolding frameworks and environments. While initial alignment efforts relied on scaling tool variety, models continued to fail when subjected to slight variations in prompts, scaffolding, and tool responses. To resolve this instability, the team introduced interleaved thinking, which enables models to execute internal reasoning at arbitrary steps throughout long-horizon tasks to adapt dynamically to external tool outputs. Additionally, developers constructed a data pipeline designed for full-trajectory generalization by introducing perturbations across system prompts, user objectives, environments, and tool feedback. In internal evaluations, MiniMax M2 maintained robust tool-calling and instruction-following performance when tested against unfamiliar, cold-start agent scaffolding frameworks.

Context

Large language model agents often demonstrate strong performance on standardized tool-use benchmarks but fail in practical deployments due to environmental shifts. Variations in agent scaffolding frameworks, toolsets, system prompts, user inputs, and unexpected tool outputs create perturbations that degrade tool calling and instruction following during long-horizon tasks.

Approach / What changed

The MiniMax M2 team addressed agent alignment by adopting interleaved thinking and building a comprehensive data pipeline for full-trajectory generalization. Interleaved thinking permits internal reasoning at any point in a workflow, allowing continuous adaptation to tool outputs while requiring users to retain session thinking history. The data pipeline generates training trajectories exposed to perturbations across system prompts, user goals, environments, and tool responses rather than solely scaling tool counts.

Takeaways

• Interleaved thinking allows an agent to generate internal reasoning steps throughout a task rather than solely at the beginning, aiding error recovery and long-context coherence.
• MiniMax M2 requires retaining the complete session history, including interleaved thinking steps, because the context functions as the model's operational memory.
• Tool scaling alone failed to ensure agent generalization across different scaffolding frameworks; robust generalization required training data covering perturbations across prompts, environments, and tool responses.

Related reading

• Salesforce · Aug 17, 2026

  How to Evaluate Production AI Agents: Measure System Outcomes, Not Conversations

  Evaluating AI agents solely on conversational transcripts fails when agents must invoke tools to modify external production systems. Language models can convincingly narrate that an action occurred while completely omitting the underlying tool call, leaving database or billing records unchanged. Outcome-based evaluation addresses this divergence by treating external system mutations rather than generated dialogue as the actual source of truth. Frameworks such as CRMAgentBench implement persistent, stateful environments with strict all-or-nothing validation, verifying tool arguments, execution sequence, final record modifications, and the total absence of unauthorized actions. Furthermore, measuring repeated success across independent executions using pass^k metrics exposes operational variance and inconsistency that standard single-shot scores and pass@k metrics conceal.

  Scott Nyberg

• Amazon · Aug 21, 2026