Zhipu's long-term vision work and RL details with Curriculum Sampling
teortaxesTex · x · 2026-08-23
A reminder that Zhipu has been working on vision models for a long time (since 2023).
The quoted text adds technical specifics:
- Solid data engineering on multimodal data.
- Insightful RL details, including answer extraction and reward system design, the utilization of Curriculum Sampling, and improvements to effectiveness and stability.
Related event: Zhipu's CogVLM2 Hits SOTA on OCR Tasks(2 posts)→
More from Models
- Does training on OBLIQ tasks bake in specific similarity notions? — antoine_chaffin · 2026-08-23
- ox model reviewed: meticulous PhD janitor as a long-horizon subagent — teortaxesTex · 2026-08-23
- Qwen3.8-27B GGUF Release with Speculative Decoding Support — z-lab · 2026-08-23
- MiniMax H3's high prompt adherence creates stiff, frozen videos lacking subtle motion — DifficultAd5938 · 2026-08-23
- Princeton's i1: A fully open text-to-image model backed by 300 controlled experiments — 机器之心 · 2026-08-23
- AI models exhibit 'Fablish' writing quirks: garbled negations and OSV word order — alexisgallagher · 2026-08-23