Chengyu Du

LLM Engineering @Minimax_AI Building MiniMax Agent, MiniMax Audio and @Hailuo_AI Ask me anything!

[Amanda Huang @amandaH_333]

OpenRoom is our first step toward an agent-native playground.

Huge thanks to the team for making this happen in such a short time!

It already feels like a miracle build — but more importantly, our iteration speed is only getting faster. OpenRoom is still evolving rapidly.

We’re moving fast and shaping it with real user feedback.

DM me for early access & share your thoughts.

[Chengyu Du @adososerious]

@amandaH_333 try it！

[Chengyu Du @adososerious]

@OpenRoom_AI_ how to use it！

[MiniMax (official) @MiniMax_AI]

Proud to have @FireworksAI_HQ as our official day-0 launch partner for MiniMax M2.5!

If you're building production agent, this is the fastest way to get started.

[karminski-牙医 @karminski3]

STOP!!! 你们真的不要再发了😇

[Pengyu Zhao @zpysky1125]

Guess if M2.5 also good at real-world entertainment 👀

[Amanda Huang @amandaH_333]

Really really happy to have worked on this project!

The key takeaway is not model capability in isolation, but how to close the loop between model iteration and real consumer usage in a domain that is both non-verifiable and inherently preference-driven.

In the work we reframe three core questions:
1.What is Role-Play?
We define Role-play as an agent’s capacity to navigate specific coordinates: {World} × {Stories}, conditioned on {User Preferences}.

https://t.co/FTBVHj490v do we evaluate it when there is no ground truth answer?
If correctness is subjective, then optimize for not being wrong.

3. How do we iterate model performance in production?
Online preference learning on denoised user signals, A/B testing for validation and iteration.

If you’re thinking about AI entertainment, or online learning in production usage— would love to discuss more!

[Skyler Miao @SkylerMiao7]

weights here, enjoy!
guess I could finally enjoy my Chinese New Year now😉

[Skyler Miao @SkylerMiao7]

Our RL researcher is on live with Alex, wanna know something about RL on M2.5? come in rn

[MiniMax (official) @MiniMax_AI]

Introducing M2.5, an open-source frontier model designed for real-world productivity.

- SOTA performance at coding (SWE-Bench Verified 80.2%), search (BrowseComp 76.3%), agentic tool-calling (BFCL 76.8%) & office work.

- Optimized for efficient execution, 37% faster at complex tasks.

- At $1 per hour with 100 tps, infinite scaling of long-horizon agents now economically possible

MiniMax Agent: https://t.co/aIzrFYcfUz
API: https://t.co/fHRdSV7BwZ
CodingPlan: https://t.co/FDhZBBjQrX

[Skyler Miao @SkylerMiao7]

Honestly I wanna release this beast ASAP — I’m dying to go back to my hometown for Spring Festival 😂
But the more training compute we put in, the more it keeps rising.
Painfully happy problem. We hear you guys. M2.5 soon.

[Chengyu Du @adososerious]

🤝 Thanks to: • CoSER: https://t.co/bbog1o0KfZ • MiniMax Role-Play Bench: https://t.co/uMNWkpY4Mg • Technical Report: https://t.co/a7NFjafd7e

👏 Discussions & Stars welcome!

[Chengyu Du @adososerious]

🚀 Excited to share our work in MiniMax on role-playing LLMs:

HER: Human-like Reasoning and Reinforcement Learning for LLM Role-playing

📄 Paper: https://t.co/rcoJ5jsTb1
💻 GitHub: https://t.co/aRzxDpSkSN

Thread below⬇️
#MiniMax #LLM #roleplay

[Chengyu Du @adososerious]

[4/4] Resources & Acknowledgments:

🔗 Access Our Work: 🤗 HF Paper: https://t.co/Yi1G5zyItE 🤖 HER Model: https://t.co/ARtLghBVQx 🤖 HER-RM: https://t.co/VmSDJrMGNh 📚 Dataset: https://t.co/xbUdg4aNwO

[Chengyu Du @adososerious]

[3/4] Results:

📊 Comprehensive Evaluation on Two Authoritative Benchmarks: • +30.26% on CoSER Bench (20 turns, 4 major dimensions)
• +14.97% on MiniMax Role-Play Bench (100 turns, 3 major dimensions)

[Chengyu Du @adososerious]

2️⃣ GenRM+RL : GenRM identifies implicit preferences from dialogue context and dynamically generates by-case principles to guide scoring, rather than relying on fixed criteria. Used in RL phase to guide role-playing model training.

[Chengyu Du @adososerious]

[2/4] Our Solution:

✨ We propose HER (Human Emulation Reasoning):

🛠️ Core Contributions: 1️⃣ Data : Constructing role-playing data with reasoning processes through reverse synthesis methods

[Chengyu Du @adososerious]

[1/4] The Challenge:
🧠 On the path toward cognitive-level persona simulation, existing research faces two major bottlenecks:
1. Lack of high-quality role-playing data with reasoning processes
2. Lack of reward signals that align with human preferences and are stable and reliable

[Chengyu Du @adososerious]

We address two critical bottlenecks in persona simulation through high-quality reasoning data + dynamic reward modeling + RL, achieving significant improvements on authoritative benchmarks.

[Skyler Miao @SkylerMiao7]

Get started in seconds with MiniMax OAuth. No API keys to copy!