Low-Precision-LLM-Efficient-Computation-Challenge

IEEE ICME 2026 Challenge

Low-Precision Large Language Model Efficient Computation Challenge

This challenge is an official challenge track of the
IEEE International Conference on Multimedia & Expo (ICME) 2026.

Latest release: v1.0 (2026-01-23)
If you are seeing an older version, please hard refresh the page (Ctrl/Cmd + Shift + R). —

🔹 Quick Start

• 💻 Complete Project Repository (GitCode):
  Low-precision LLM Efficient Computation Challenge – GitCode

• 📄 Challenge Description (PDF):
  Download the Challenge Description (PDF)

• 📝 Registration Form:
  Register for the Challenge

• 🧰 Code, Tools, and Datasets:
  Browse Code and Resources on GitCode

---

1. Challenge Background

With the rapid growth of large-scale language and multimodal models, computational cost and energy efficiency have become major bottlenecks. Low-precision numerical formats such as HiFloat8, HiFloat4, and MXFP4 offer promising opportunities to significantly reduce memory footprint and computation cost while maintaining model accuracy.

This challenge aims to promote algorithmic and system-level innovations in low-precision training and inference for large models, with a strong emphasis on reproducibility and practical deployment.

The challenge consists of two independent sub-challenges:

• Sub-Challenge 1: W4A4 Quantization for Inference
• Sub-Challenge 2: W8A8 Quantization for Training

Each participant or team may choose to participate in only one sub-challenge. Submissions across multiple sub-challenges are not allowed.

📄 Full technical details are available in the challenge description PDF:
👉 Download the Challenge Description (PDF)

---

2. Participation and Registration

🔗 Registration Form:
https://forms.gle/eFQW3FXuPB8bVjWX9

For registration inquiries:
📧 icme_challenge_registration@your-domain.com

---

3. Sub-Challenges

Sub-Challenge 1: W4A4 Quantization for Inference

• Supported formats: HiFloat4 / MXFP4
• Reference model: Qwen3-32B
• Requirement: average accuracy degradation < 0.5% compared with FP16

Sub-Challenge 2: W8A8 Quantization for Training

• Supported format: HiFloat8
• Test model: Wan2.1
• Requirement: accuracy degradation < 0.5%

---

4. Evaluation Metrics

Objective (60%)

• Loss (30%)
• Inception model / FID (30%)

Subjective (40%)

• University volunteer scoring (40%)
  • Authenticity (20%)
  • Clarity (20%)

---

5. Important Dates

Phase	Description	Date
Registration	Challenge registration period	Jan 30, 2026 – Mar 10, 2026
Submission	Result submission window	Feb 10, 2026 – Mar 15, 2026
Review	Review & reproducibility verification	Mar 16, 2026 – Mar 29, 2026
Announcement	Final results announcement	Mar 30, 2026

---

6. Submission

📧 Submission Email:
submission_icme_challenge@your-domain.com

Each submission must be sent as a single archive (.zip / .tar.gz) and include:

• Quantized model weights (.pt / .ckpt / equivalent)
• Source code (training and/or inference)
• Technical report (PDF)
• Reproducibility scripts + instructions

---

7. Code, Tools, and Datasets

All engineering materials (baseline code, simulation tools, datasets, evaluation scripts) will be maintained in the GitCode repository:

👉 https://gitcode.com/your-username/your-repo-name

---

8. Official Contact

📧 icme_challenge@your-domain.com

---

Note: This challenge is jointly released by Huawei Technologies Co., Ltd. and Tsinghua University. Baseline code, simulation tools, datasets, and evaluation scripts will be released in stages. Please watch the GitCode repository for updates.