50 researchers from ByteDance, Alibaba, Tencent release 303-page code AI field guide
mdancho84 · x · 2026-11-11
About 50 researchers from ByteDance, Alibaba, Tencent and universities published a 303-page survey and practical guide, "From Code Foundation Models to Agents and Applications," covering the full lifecycle of code LLMs: data curation, pre-training, post-training (SFT/RLHF/RLVR), prompting paradigms, and coding agents, with comparative analysis of GPT-4, Claude, LLaMA, StarCoder, DeepSeek-Coder and QwenCoder, plus experiments on scaling laws, framework selection and hyperparameter sensitivity.
Two counterintuitive takeaways highlighted: small open models can close the gap with giants on reasoning-style coding tasks if RL is done right via RLVR; and mixing languages in pretraining helps only up to a point—Python's dynamic typing causes negative transfer vs. statically typed languages, while pairs like Java↔Cand JS↔TS show strong synergy.
Related event: Researchers Release 303-Page Survey on Code Intelligence and Coding Agents(2 posts)→
More from coding & agent
- Claude Adds Interactive Connectors for Adobe, Atlassian and Box — claudeai · 2026-11-11
- thimble: open-source Claude Code plugin makes agent outputs explainable in your terminal — AlexTamkin · 2026-11-11
- Gemini 4 Carbon lands on Vercel AI Gateway at 50% off through Nov 30 — steren · 2026-11-11
- Agent tunes a live camera for ML training: focus lock 190, white balance 4600K — TheZachMueller · 2026-11-11
- Fable 5.2 autonomously trains a local PDF-to-markdown model in 2 hours — IgorCarron · 2026-11-11
- Fable 5.2 one-shot pelican-on-a-bicycle video, full agent trace published — victormustar · 2026-11-11