M2-Her

M2-her is MiniMax’s flagship language model product specifically designed for role-playing (Role-Play), with a core positioning as a “high-immersion agent engine that supports deep emotional interaction and long-term narrative consistency,” tailored to create personalized, lifelike virtual character experiences.

• Experience Upgrade: Focused on the “emotional peak moments” users seek, it maintains high engagement even after 20 rounds of conversation, avoiding retention decline caused by shallow interactions.
• Assessment Innovation: Pioneering the Role-Play Bench evaluation system, which uses Situated Reenactment to quantify the risk of misalignment across three key dimensions—Worlds, Stories, and User Preferences.
• Long-Term Stability: It maintains logical consistency over 100 rounds of dialogue, significantly outperforming competitors in addressing issues such as referential confusion, character breakdown, and repetitive plotlines.
• Preference Awareness: It accurately identifies users’ unspoken expectations (such as pacing preferences and interaction styles), avoiding speaking for users, ignoring their intentions, or over-rejecting their input.
• Content Restraint: It optimizes reply length control, preventing “word inflation” in long conversations while maintaining high information density and reading comfort.

───────────────────────────────────────────────────────────────────

Core Capabilities

🌍 Unique World Building (Worlds):

• It understands and anchors myriad settings—from campuses to martial arts realms—avoiding character homogenization caused by “averageism.”
• It effectively handles mixed scenarios involving multiple characters and narration, resolving spatial logic confusion (such as dialogues occurring after farewells without physical presence).

📖 Breathable Story Progression (Stories):

• It proactively introduces new plot elements, rejecting mechanical repetition and vague dialogue, thus sustaining narrative tension.
• Every behavioral change is carefully foreshadowed, avoiding abrupt OOC (out-of-character) moments and ensuring plot coherence.

❤️ Accurate User Preference Understanding (Preferences):

• It learns personalized pacing from implicit signals, such as those triggered by the “repeat” button, adapting to users who prefer slow-burn or fast-paced interactions.
• It naturally introduces topic hooks, avoiding silences; within safe boundaries, it minimizes over-rejection, enhancing interaction fluency.

🧪 Data and Training Innovation:

• It employs Agentic Data Synthesis to generate high-quality, highly diverse dialogue trajectories, covering long-tail characters and niche worlds.
• Combined with Online Preference Learning, it leverages denoised implicit feedback to continuously optimize its contextual alignment capabilities.

───────────────────────────────────────────────────────────────────

Selected Evaluation Data