Author of the BaiMian large model, offering technology and industry insights.
After DPO, the typical research trajectory moves through GRPO, DAPO, GSPO, and SAPO, each introducing new optimization objectives, sampling strategies, and reward‑shaping techniques that aim to reduce memory usage, improve gradient stability, and enhance the efficiency of large‑model reinforcement learning.
This guide explains what Coze Agent Skills are, how they are structured with YAML and Markdown, and provides step‑by‑step instructions and real‑world examples for creating, deploying, and using no‑code AI skills to automate tasks like model scouting, recipe suggestions, and camera settings.
This article systematically reviews the main reinforcement‑learning algorithms—PPO, GRPO, DAPO, GSPO, and SAPO—used for fine‑tuning large language models, explaining why supervised fine‑tuning precedes RL, how each method improves training efficiency and stability, and what trade‑offs they entail.
GLM-Image, a Chinese text‑to‑image model trained end‑to‑end on Huawei Ascend 800T A2 NPUs, combines an autoregressive decoder with a diffusion encoder, supports resolutions up to 2048×2048, and offers open‑source code, API access, and detailed prompts that demonstrate its strong layout and typography capabilities.
The article analyzes TeleChat3-105B-A4.7-Thinking, the first domestically built 100‑billion‑parameter Mixture‑of‑Experts model, detailing its multi‑dimensional evaluation, three‑stage training pipeline, hardware‑level optimizations, fine‑grained architecture, and its significance for the evolving AI competition landscape.
The GLM-4.7 model launches with record‑breaking benchmark scores across coding, reasoning, and real‑world programming tasks, outperforming both open‑source and commercial LLMs while introducing advanced interleaved, retained, and round‑level thinking modes that enhance complex task execution.
This article analyzes the training‑efficiency challenges of multi‑turn agentic reinforcement learning and compares four recent open‑source frameworks—AReal (Ant), Seer (Moonshot), Slime (Zhipu) and verl (Bytedance)—examining their asynchronous inference designs, rollout‑train separation, long‑context handling, off‑policy mitigation, and system‑level optimizations to guide framework selection.
The article analyzes the 2025 shift toward a unified "general‑type" agent architecture exemplified by Claude Code and Deep Agent, detailing industry adoption, core technical features, skill‑based extensions, long‑running capabilities, and practical steps for building domain‑specific agents.
Over recent months of extensive agent reinforcement‑learning experiments across search, data‑analysis, and multi‑source scenarios, the author shares twelve practical insights covering stability, environment‑reward‑algorithm priorities, tool‑call reliability, reward hacking pitfalls, evaluation alignment, and scaling tricks for larger models.
This article critically examines recent RL‑for‑LLM studies, revealing that reinforcement learning improves search efficiency but does not extend the intrinsic reasoning capabilities of base models, and explores the underlying model‑conditioned optimization bias, comparisons with SFT distillation, and the trade‑off with catastrophic forgetting.
