Synced · WeChatMay 18, 2026, 12:47 UTC+8

ICML 2026: Huawei GTS proposes EDCO for dynamic curriculum fine-tuning

ICML 2026 | 华为GTS提出AI训练数据新方法，Amazon/Google作者团队「光速跟进」：难度自适应训练正在成为新范式

Huawei GTS proposed EDCO, a dynamic curriculum method that selects fine-tuning samples by inference entropy; prefix entropy estimation cuts per-sample scoring time from 2.24 seconds to 0.37 seconds.

Why it matters: HKR-H/K/R pass: the story has a lab-race hook, a concrete entropy-based mechanism, and a 2.24s→0.37s efficiency claim. It stays below 78 because it is still a training-method paper, not a major model or product release.

Read the original ↗Export Markdown

Take

EDCO’s useful move is not the “difficulty-adaptive training” label. It plugs sample selection into the fine-tuning loop and makes the scoring cost survivable. The concrete hook is strong: full-sequence entropy costs 2.24 seconds per sample, prefix entropy drops it to 0.37 seconds, and 8-GPU parallel scoring reaches 0.04 seconds. On Datacom RLFT, accuracy moves from 40.43% with random sampling to 46.96%.

I don’t buy the “new paradigm” framing yet. Active learning, curriculum learning, and hard-example mining have been circling this idea for years. The LLM-specific question is whether inference entropy stays a reliable proxy for useful gradient signal. The paper tests Qwen3-4B, Llama3.2-3B, and three vertical domains; it does not settle larger models, longer training, or production distribution drift.

Source: Synced · WeChat · mp.weixin.qq.com
Published: May 18, 2026, 12:47 UTC+8 · May 18, 2026