71-author survey maps code foundation models to coding agents, with practical guide
mdancho84 · x · 2026-11-11
Recommends a book-markable survey: "From Code Foundation Models to Agents and Applications" (arXiv:2511.18538) by Jian Yang and 70 co-authors, covering the full landscape from code foundation models to coding agents and applications.
The paper argues LLMs have fundamentally transformed automated software development via tools like GitHub Copilot, Cursor, Trae, and Claude Code. Key takeaways from the author's thread:
- The "alignment tax" on code is real: safety alignment can reduce coding capability; the paper discusses mitigation.
- Multi-agent debate can cut hallucinations — if configured well; done poorly, agents reinforce each other's mistakes.
- "Lost in the middle" persists: huge context windows don't guarantee usefulness; mid-context info still gets missed.
- Structure > perfect facts in reasoning training: models learn the decomposition template more than literal correctness of intermediate steps.
- Code models are insecure by default: training on public repos inherits a decade of insecure patterns; safety fine-tuning doesn't fix code security.
- MoE vs Dense = capacity vs stability: MoE is more fragile during SFT; hyperparams and routing stability matter more.
- Coding scaling laws aren't uniform: different languages need very different data amounts; Python/JS are ironically harder than "enterprise-y" languages.
- Python is weirdly hard: dynamic typing causes negative transfer vs statically typed languages, while 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
- Gemini 4 Carbon Developer Guide: When to Pick Carbon, Thinking Levels & Verify Loops — rseroter · 2026-11-11
- LaunchDarkly launches Release Autopilot; already ships 20% of PRs agent-direct — annetgriffin · 2026-11-11
- Mercor pays experts $5M/day as it bets big on RL environments for vertical models — ccerrato147 · 2026-11-11
- Mercor acquired Deeptune, building hundreds of simulated apps for agents to practice in — ccerrato147 · 2026-11-11
- Peter Yang: building a kart racer with Fable all morning — actually a good game, not one-shot slop — petergyang · 2026-11-11
- Anthropic's real bet is MCP connectors embedding HubSpot, Adobe and Atlassian inside Claude — petrusenko_max · 2026-11-11