agents-meet-rl
GitHub针对LLM智能体强化学习训练、评估及实验设计的故障排查助手。基于2026年6月语料库，诊断奖励不收敛、KL爆炸等具体问题，提供根因分析与引用修复方案，支持PPO/GRPO等算法选择，不执行实际训练。
Trigger Scenarios
Install
npx skills add thinkwee/AgentsMeetRL --skill agents-meet-rl -g -y
SKILL.md
Frontmatter
{
"name": "agents-meet-rl",
"description": "Troubleshooter for agentic-RL training, evaluation, and experiment design on LLM agents (single or multi-agent, multi-turn, tool-augmented). Routes a user's symptom to fixes anchored in the corpus. TRIGGER when: user is training, evaluating, or designing experiments for an RL-trained LLM agent; symptoms like reward not moving, eval flat, KL\/entropy\/length blow-ups, retokenization drift, tool-call parse failures, credit assignment, async-rollout staleness, judge inconsistency, benchmark contamination, pass@k vs pass@1; choices about ablation, baseline, framework, algorithm, reward, or data curation; user names GRPO, PPO, DAPO, veRL, OpenRLHF, slime, AReaL, RAGEN, or similar. SKIP: generic supervised LLM fine-tuning with no RL component; classical RL theory or tabular RL; non-LLM agents. Distilled from the AgentsMeetRL awesome list, snapshot 2026-06-20."
}
What this is
A corpus-anchored handbook for diagnosis and selection. It supplies knowledge — it does not read or run your training: it can't inspect your logs, wandb, or live metrics. You bring the symptom; it returns likely causes, checks, and cited fixes for you to apply.
Where things are
problems/_INDEX.md— symptom → file routing, grouped undertraining/,evaluation/,research-workflow/. Start here.problems/<cat>/<file>.md— per-symptom files. Most followSymptoms → Root causes → Diagnosis → Fixes → References; knob / decision / modality / eval-checklist / research-workflow files use task-oriented structures.references/_INDEX.md+references/<cat>.md— per-category project lists with full metadata. Each entry carries anIdea:line — one sentence on its distinctive contribution, grounded in the paper/repo. Use for "which framework / benchmark" selection, to look up project names not routed viaproblems/_INDEX.md, and to answer "what's the idea behind X" by quoting itsIdea:line.database.json— machine-readable, 312 entries (each with atakeawayfield mirroring theIdea:line) plus 3 paper-only algorithms (DAPO, Dr.GRPO, VAPO) whitelisted inscripts/lint_skill.py.
Citing fixes
Name the algorithm or idea, then anchor with whatever canonical URLs exist for that entry — typically github + arxiv + org + date, but paper-only algorithms (in the whitelist) get just the paper URL, and tools / environments without papers get just github + org + date.
Examples:
Project with paper (typical): Adapt Search-R1's outcome-only reward — code · paper · UIUC/Google · 2025.3.
Paper-only algorithm (whitelist): Try DAPO's clip-higher — paper · ByteDance Seed · 2025.3.
Tool / environment without paper: Run rollouts in atropos — code · Nous Research · 2025.4.
Cite at the idea level, not paper sections or file paths inside repos — they rot. If an entry isn't in the corpus, say so; don't fabricate.
If two corpus entries share a name (e.g. ARPO appears as both a
reasoning RL method and a GUI-agent training method), disambiguate by
including the org and paper URL — they are different works.
Staleness
Snapshot date: 2026-06-20. If the user mentions a project or paper released after that, flag explicitly that this skill's corpus may not cover it.
Version History
- 8da3504 Current 2026-07-05 14:37
