# Paper Notes — 中文精读笔记合集

离线可读的中文精读笔记 + 论文 PDF。三个主题分别放在子目录里。

## 🔵 spec-rl/ — Speculative Decoding × RL (8 篇)

| # | 笔记 | 主题 | 大小 |
|---|---|---|---|
| 01 | [NVIDIA NeMo-RL × EAGLE-3](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/01_NVIDIA_NeMoRL_EAGLE3.html) | 系统集成 / verifier-exact / .detach() 栅栏 | 50 KB |
| 02 | [EAGLE-1/2/3 + DFlash 演化合集](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/02_EAGLE_evolution_DFlash.html) | 底层 drafter 一线发展轨迹 | 50 KB |
| 03 | [SPEC-RL (Shopee)](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/03_SPEC-RL_2509.23232.html) | 复用上一 epoch trajectory + lenience ℓ | 43 KB |
| 04 | [ReSpec](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/04_ReSpec_2510.26475.html) | 三个 gap + reward-weighted KD | 50 KB |
| 05 | [DAS](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/05_DAS_2511.13841.html) | suffix tree + length-aware budget | 40 KB |
| 06 | [SRT](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/06_SRT_2601.09083.html) | per-prompt token-frequency trie + run-ahead | 40 KB |
| 07 | [DFlash 详细版](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/07_DFlash_2602.06036.html) | 块扩散 drafter / 5 层 K/V 注入 / EAGLE-3 范式分叉 | 80 KB |
| 08 | [Hidden States Drift](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/08_HiddenStatesDrift_2604.26412.html) | KV-only reuse vs hidden reuse / KVShot 框架 | 52 KB |
| 09 | [Mirror-SD](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/09_MirrorSD_2510.13161.html) | early-exit + GPU/NPU 异构并行 / +30% over EAGLE-3 | 55 KB |
| 10 | [SSD (Speculative Speculative Decoding)](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/10_SSD_2603.03251.html) | 预测 verify 结果 / draft 隐藏在 verify 后面 / Saguaro 5× | 57 KB |
| 11 | [Performance or Illusion?](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/11_PerformanceOrIllusion_2601.11580.html) | 生产级 vLLM 审视 / paper vs production / 决策 cheatsheet | 51 KB |
| 12 | [DropMatch](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/12_DropMatch_2603.03333.html) | LM head MC dropout / semantic-level acceptance / training-free | 40 KB |
| 13 | [Speculative Verification (SV)](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/13_SpecVerification_2509.24328.html) | info-gain 动态 verify length / companion 模型 | 44 KB |
| 14 | [LK Losses](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/14_LKLosses_2602.23881.html) | 直接优化 acceptance rate / TV+KL hybrid / 反对 proxy 训练 | 41 KB |
| 15 | [📐 Lossless 证明 (Tutorial)](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/15_Lossless_Proof_Tutorial.html) | rejection sampling 数学证明全步详解 (Leviathan/Chen 2023) | 53 KB |
| 16 | [DDTree](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/16_DDTree_2604.12989.html) | 块扩散 drafter 上的最优 draft tree / best-first heap | 47 KB |
| 17 | [Orthrus](https://zhangshuibai.github.io/paper-notes/spec-rl/notes/17_Orthrus_2605.12825.html) | dual-view diffusion / frozen AR + diffusion attention path 共享 KV / 7.8× lossless / O(1) 内存 | 54 KB |

建议阅读顺序: **15 (Lossless 数学证明) → 02 (EAGLE 演化) → 04 (ReSpec) → 03 (SPEC-RL) → 05 (DAS) → 01 (NVIDIA) → 06 (SRT) → 07 (DFlash) → 08 (Hidden States Drift) → 09 (Mirror-SD) → 10 (SSD) → 14 (LK Losses) → 13 (SV) → 12 (DropMatch) → 11 (Performance or Illusion?) → 16 (DDTree) → 17 (Orthrus, intra-model 终极 lossless 加速)**

## 🟡 models/ — 前沿模型 / 训练系统 / LM 方法 (24 篇)

| # | 笔记 | 主题 | 大小 |
|---|---|---|---|
| 01 | [INTELLECT-1](https://zhangshuibai.github.io/paper-notes/models/notes/01_INTELLECT1_2412.01152.html) | 跨地理 GPU 去中心化预训练 / DiLoCo / int8 ring all-reduce | 54 KB |
| 02 | [MiniMax-01](https://zhangshuibai.github.io/paper-notes/models/notes/02_MiniMax01_2501.08313.html) | Lightning Attention + 32-expert MoE / 1M context | 45 KB |
| 03 | [MiniMax-M1](https://zhangshuibai.github.io/paper-notes/models/notes/03_MiniMaxM1_2506.13585.html) | Hybrid lightning + softmax / CISPO | 47 KB |
| 04 | [MiMo-V2-Flash](https://zhangshuibai.github.io/paper-notes/models/notes/04_MiMo-V2-Flash_2601.02780.html) | 5:1 SWA-global / sink bias / 256 fine-grained experts / MTP | 62 KB |
| 05 | [MARS](https://zhangshuibai.github.io/paper-notes/models/notes/05_MARS_2601.15498.html) | Margin-aware speculative verification (logit-ratio) | 41 KB |
| 06 | [Kimi K2.5](https://zhangshuibai.github.io/paper-notes/models/notes/06_KimiK2.5_2602.02276.html) | 多模态 agentic / Agent Swarm / PARL | 53 KB |
| 07 | [INTELLECT-2](https://zhangshuibai.github.io/paper-notes/models/notes/07_INTELLECT2_2505.07291.html) | 全球去中心化 RL / PRIME-RL / TOPLOC / SHARDCAST | 59 KB |
| 08 | [INTELLECT-3](https://zhangshuibai.github.io/paper-notes/models/notes/08_INTELLECT3_2512.16144.html) | 106B MoE agentic RL / prime-rl 三件套 / IcePop | 57 KB |
| 09 | [OmniHuman 1.5](https://zhangshuibai.github.io/paper-notes/models/notes/09_OmniHuman1.5_2508.19209.html) | 数字人 / Sys1+Sys2 / Pseudo Last Frame | 48 KB |
| 10 | [TOP](https://zhangshuibai.github.io/paper-notes/models/notes/10_TOP_2508.19228.html) | Token Order Prediction / learning-to-rank LM 损失 | 48 KB |
| 11 | [TIDE](https://zhangshuibai.github.io/paper-notes/models/notes/11_TIDE_2604.26951.html) | 跨架构 dLLM 蒸馏 / TIDAL+CompDemo+Reverse CALM | 51 KB |
| 12 | [NCP / ConceptLM](https://zhangshuibai.github.io/paper-notes/models/notes/12_NCP_2602.08984.html) | Next Concept Prediction / VQ 离散 latent | 50 KB |
| 13 | [LongLLaDA](https://zhangshuibai.github.io/paper-notes/models/notes/13_LongLLaDA_2506.14429.html) | 扩散 LLM 长上下文 / NTK-RoPE training-free | 46 KB |
| 14 | [MTP (Gloeckle, Meta)](https://zhangshuibai.github.io/paper-notes/models/notes/14_MTP_Gloeckle_2404.19737.html) | MTP 元论文 / 共享 trunk + n heads / 自-spec | 36 KB |
| 15 | [FastMTP](https://zhangshuibai.github.io/paper-notes/models/notes/15_FastMTP_2509.18362.html) | 单 head + position-shared / self-distill / 2.03× lossless | 54 KB |
| 16 | [L-MTP (Leap MTP)](https://zhangshuibai.github.io/paper-notes/models/notes/16_LMTP_2505.17505.html) | 跳跃式 MTP / 非相邻 future / NeurIPS 2025 | 50 KB |
| 17 | [MuToR (Registers)](https://zhangshuibai.github.io/paper-notes/models/notes/17_MuToR_2505.10518.html) | 可学习 register tokens 替代 n heads | 53 KB |
| 18 | [FSP (Beyond MTP)](https://zhangshuibai.github.io/paper-notes/models/notes/18_FSP_2510.14751.html) | Future Summary Prediction / RevLM 蒸馏 | 41 KB |
| 19 | [📊 RL 框架横评 (AReaL/veRL/slime/NeMo-RL/OpenRLHF…)](https://zhangshuibai.github.io/paper-notes/models/notes/19_RL_Frameworks_Comparison.html) | 10 框架 / 12 维对比 / 决策树 / sync vs async | 53 KB |
| 20 | [CoPD (Co-Evolving Policy Distillation)](https://zhangshuibai.github.io/paper-notes/models/notes/20_CoPD_2604.27083.html) | RLVR + OPD 同步 / mutual teacher / 钟形 η(O) | 50 KB |
| 21 | [Prism (test-time scaling for dLLM)](https://zhangshuibai.github.io/paper-notes/models/notes/21_Prism_2602.01842.html) | 离散扩散 LLM 的 hierarchical search + self-verify | 50 KB |
| 22 | [Retrofitting Recurrence](https://zhangshuibai.github.io/paper-notes/models/notes/22_RetrofitRecurrence_2511.07384.html) | 把 pretrained TinyLlama/OLMo/Llama 切成 depth-recurrent / curriculum + Muon | 50 KB |
| 23 | [RePlaid (continuous DLM) **(本人参与)**](https://zhangshuibai.github.io/paper-notes/models/notes/23_RePlaid_2605.18530.html) | Plaid → 现代 DiT 架构对齐 / OWT PPL 22.1 SOTA(continuous) / vs AR 20× compute(原 64×) / 反超 MDLM 23.1 + Duo 25.2 / variance-min schedule ⇒ linear CE 闭式定理 / embedding 几何驱动 Δ=17.3 PPL | 68 KB |
| 24 | [Nemotron-Labs-Diffusion (NVIDIA Tech Report 2026-05)](https://zhangshuibai.github.io/paper-notes/models/notes/24_NemotronDiff_2026_TechReport.html) | tri-mode LM (AR / Diffusion / Self-Spec) 统一在一颗 8B ckpt / joint loss ℒ_AR + 0.3·ℒ_diff / 双流 attention 4 象限(clean→clean strictly causal 是关键创新)/ 两阶段 1T AR + 300B joint / 256× H100 · Megatron Bridge / Linear SS w/ LoRA on o_proj acceptance 5.46→6.82 (vs Eagle3 2.75 / MTP 4.24) / GB200 SGLang 4× 端到端吞吐 vs Qwen3-8B-Eagle3 2.4× / SOL 7.60× ceiling, 比 Linear SS 高 76.5% real-TPF headroom / SPEED-Bench 11 类 + 3B/8B/14B + VLM 全开源 / **强连接 #23 RePlaid(Mardani 同人, NVIDIA H1 2026 离散 vs 连续双押)** / **同人前作 TiDAR (arXiv:2511.08923, Liu/Fu/Kautz/Molchanov)** — Quad SS = TiDAR 方法,Linear SS + tri-mode + SOL 分析全部新增 | 84 KB |

### MTP 系列阅读建议

**14 (Gloeckle 元论文) → 15 (FastMTP 工程化) → 16 (L-MTP 跳跃) → 17 (MuToR registers) → 18 (FSP summary)**

- 14 是起源(n 个独立 heads,提 sample efficiency + 自-spec decoding)
- 15 是把 MTP 真正做成可工程化的 spec decoding(单 head + self-distill + 2.03× lossless 加速)
- 16/17 是对 head 设计的两条改良路线: 跳跃 vs register
- 18 是对 MTP "预测精确 token"这一根本目标的反思——预测未来摘要更稳

也可参考 [TOP (10)](https://zhangshuibai.github.io/paper-notes/models/notes/10_TOP_2508.19228.html) 看反向观点(用 ranking 替代 MTP)。

### INTELLECT 系列阅读建议

I-1 (decentralized **pretrain** / DiLoCo) → I-2 (decentralized **RL** / TOPLOC + SHARDCAST) → I-3 (**agentic RL** on top of GLM-4.5-Air-Base / 务实转向)。

三代叙事曲线: 从证明跨地理预训练可行,到证明跨地理 RL 可信,到放弃 from-scratch 而转向 agentic 实用。

## 🟢 agents/ — Agent 训练 / 环境 / 内存 (39 篇)

| # | 笔记 | 主题 | 大小 |
|---|---|---|---|
| 01 | [DR-Venus](https://zhangshuibai.github.io/paper-notes/agents/notes/01_DR-Venus_2604.19859.html) | 4B 边缘部署 deep research agent / IGPO + 信息增益 reward | 52 KB |
| 02 | [LiteResearcher](https://zhangshuibai.github.io/paper-notes/agents/notes/02_LiteResearcher_2604.17931.html) | Lite 虚拟 Web 世界 / 4B GAIA 71.3% 反超 30B | 50 KB |
| 03 | [Agent-World](https://zhangshuibai.github.io/paper-notes/agents/notes/03_AgentWorld_2604.18292.html) | 1,978 真实环境 / 19,822 工具 / self-evolving arena | 60 KB |
| 04 | [MemPalace 审视](https://zhangshuibai.github.io/paper-notes/agents/notes/04_MemPalace_2604.21284.html) | 空间隐喻 LLM 内存的批判分析 / 96.6% 主要来自哪 | 52 KB |
| 05 | [Rethinking Agentic RL](https://zhangshuibai.github.io/paper-notes/agents/notes/05_RethinkingAgenticRL_2604.27859.html) | 范式重定义: 退化 single-step bandit → POMDP / capability taxonomy | 52 KB |
| 06 | [AgentGym-RL](https://zhangshuibai.github.io/paper-notes/agents/notes/06_AgentGymRL_2509.08755.html) | 模块化框架 / ScalingInter-RL 课程 / 27 任务 | 57 KB |
| 07 | [AgentRL](https://zhangshuibai.github.io/paper-notes/agents/notes/07_AgentRL_2510.04206.html) | 全异步管道 / cross-policy sampling / 跨任务标准化 | 52 KB |
| 08 | [CoEvolve](https://zhangshuibai.github.io/paper-notes/agents/notes/08_CoEvolve_2604.15840.html) | agent ↔ data mutual evolution / GRPO + 信号驱动数据进化 | 44 KB |
| 09 | [HeavySkill](https://zhangshuibai.github.io/paper-notes/agents/notes/09_HeavySkill_2605.02396.html) | Agent Swarm 内化为单模型 inner skill / parallel + summarize | 52 KB |
| 10 | [Search-R1](https://zhangshuibai.github.io/paper-notes/agents/notes/10_SearchR1_2503.09516.html) | search-augmented RL / retrieved-token masking / +41% over RAG | 46 KB |
| 11 | [ReTool](https://zhangshuibai.github.io/paper-notes/agents/notes/11_ReTool_2504.11536.html) | code-integrated RL / cold-start SFT + outcome RL / AIME 67% (vs text-RL 40%) | 60 KB |
| 12 | [LLM Memory 演化合集 (MemoryBank → Mem0 → MEM1 → Memory-R1)](https://zhangshuibai.github.io/paper-notes/agents/notes/12_Memory_evolution.html) | 四篇串读 / heuristic→CRUD→RL 演化 / external vs internal 路线对比 | 70 KB |
| 13 | [RLAnything](https://zhangshuibai.github.io/paper-notes/agents/notes/13_RLAnything_2602.02488.html) | closed-loop RL: env + policy + reward 联合优化 / +9.1% OSWorld / ICML 2026 | 38 KB |
| 14 | [ClawGUI](https://zhangshuibai.github.io/paper-notes/agents/notes/14_ClawGUI_2604.11784.html) | GUI agent 全栈框架 / RL + Eval + Deploy 三件套 / MobileWorld 17.1% | 50 KB |
| 15 | [PinchBench (Kilo Code)](https://zhangshuibai.github.io/paper-notes/agents/notes/15_PinchBench.html) | OpenClaw agent benchmark 调研 / 120+ task / 自托管模型完整接入流程 + 提交流程 | 55 KB |
| 16 | [🖥️ 桌面 GUI / CLI Agent 开源生态全景](https://zhangshuibai.github.io/paper-notes/agents/notes/16_DesktopAgents_landscape.html) | Win/macOS/Linux 桌面 + CLI 生态调研 / 14+ repo / 仿真器+评测+训练数据+infra | 50 KB |
| 17 | [UI-TARS-2 (ByteDance Seed)](https://zhangshuibai.github.io/paper-notes/agents/notes/17_UITARS2_2509.02544.html) | 230B MoE GUI agent / data flywheel + 多轮 PPO + GUI+SDK 混合环境 / OSWorld 47.5 / 主线权重未开源 | 56 KB |
| 18 | [Agent World Model (Snowflake)](https://zhangshuibai.github.io/paper-notes/agents/notes/18_AgentWorldModel_2602.10090.html) | 全合成 code-driven env / 1,000 envs · 35,062 tools / Qwen3-thinking 4/8/14B GRPO / BFCLv3 8B 53.83→65.94 OOD / 全开源 | 60 KB |
| 19 | [MCP-Atlas (Scale AI)](https://zhangshuibai.github.io/paper-notes/agents/notes/19_MCPAtlas_2602.00933.html) | 36 real MCP server · 220 tool · 1,000 task (500 公开 + 500 hold-out) / claims-based Gemini-2.5-Pro judge / Live leaderboard / 被 Claude Opus 4.7 system card 引用 77.3% | 50 KB |
| 20 | [SETA (CAMEL-AI / Eigent.AI)](https://zhangshuibai.github.io/paper-notes/agents/notes/20_SETA_camel.html) | Terminal agent 训练 env / 400+ Docker 合成任务 / AReaL GRPO + RLVR / Qwen3-8B / Claude Sonnet 4.5 + CAMEL = TB2.0 #1 (46.5%) / 全 Apache-2.0 开源 | 50 KB |
| 21 | [4 个 MCP Benchmark 横向对比](https://zhangshuibai.github.io/paper-notes/agents/notes/21_MCPBenchmarks_compare.html) | MCP-Universe / MCP-Atlas / Toolathlon / MCPMark 调研 / GitHub 实时 star / frontier model card 引用 / Toolathlon-Gym 唯一 RL gym / 4 bench 提交流程对比 + 选型决策树 | 55 KB |
| 22 | [TOUCAN (MIT-IBM + UW)](https://zhangshuibai.github.io/paper-notes/agents/notes/22_TOUCAN.html) | arXiv:2510.01179 v1 2025-10-01 / 1.5M tool-agentic SFT 数据(精确 1,527,259 traj) / 495 真实 MCP server · 2,000+ tools / HF 4 subset 精确: Kimi-K2 518,516 + OSS 457,130 + Qwen3 551,613 + **SFT 119,287** / **SFT 内部拆分**: 28.3K original + 40K Ext.1 Irrelevance + 15.8K Ext.2 Diversify + 35.2K Ext.3 Multi-Turn / **3 件 contribution**(Dataset + Pipeline + Checkpoints) / 5-LLM 出题 (Mistral-S/DevStral-S/GPT-OSS/Kimi-K2/Qwen3-32B) × 3 策略 (Single/Multi/Featured-25) + 3-teacher × 2-framework distill + 6-维 Likert task filter + 双层 rule+LLM-as-judge post-filter / **BFCL V3 全 5 子项**: 32B+TOUCAN 70.45 vs GPT-4.5 70.32(Overall 险胜)· Multi-Turn 46.50 vs 45.38(**也胜**)· Non-Live AST 7B 反退步 -5.67 / **τ-Bench / τ²-Bench airline+retail+telecom 全 4 域**: 32B 42.33/31.60 · Telecom 域 7B 反退 -6.20 暴露 Communication server 仅 7 个的覆盖偏置 / MCP-Universe Pareto 推进但无绝对数(~25-30%) / **Table 6 消融**: 单训 Single-Turn 把 Irrelevance 砸到 46.88,Ext.1 一举修到 77.85 (+30.97) / 数据 Apache-2.0 + 代码 MIT + 权重跟 Qwen2.5 Apache-2.0 = **全栈商用 friendly** / 27 个 HF ckpt(3 官方 + 15 量化 + 9 衍生) / **校订**: V3≠V4,V4 Agentic 40% 重排后 TOUCAN 不直接可比,需补 SETA/EnvScaler/EnvTuning 类 long-horizon 数据 / 链 #03 #18 #20 #28 #29 #31 | 76 KB |
| 23 | [EnvScaler (RUC NLPIR)](https://zhangshuibai.github.io/paper-notes/agents/notes/23_EnvScaler.html) | 程序化合成 tool-interactive env / Python class = env / dual-agent 100 轮质量门 / 191 env · ~7K scen · ~9K traj / Qwen3-Thinking 1.7B/4B/8B + Reinforce++ / Qwen3-8B BFCL-MT 41.88 超 GPT-4.1 / MIT 全开源 | 50 KB |
| 24 | [🏭 Smithery 调研 (Industry Analysis)](https://zhangshuibai.github.io/paper-notes/agents/notes/24_Smithery.html) | MCP 生态的 Docker Hub / 7,300+ servers (2026-05) / Henry Mao 2024-12 创立 / South Park Commons Seed / Registry 免费 + Hosting 2026-03 转付费 / Connect managed OAuth / #03 #22 #23 三篇 paper 的 MCP 主题源 | 50 KB |
| 25 | [MCP-Universe (Salesforce AI Research)](https://zhangshuibai.github.io/paper-notes/agents/notes/25_MCPUniverse.html) | framework 而非 benchmark / 11 MCP server · 6 domain · 231 task / format+static+dynamic 三层评测拒绝 LLM judge / SOTA GPT-5 43.72% · Grok-4 33.33% · Claude-4-Sonnet 29.44% / <strong>真有 GRPO+verl RL 训练栈</strong> (Hybrid + Fully-Async + TITO + docker_pool) / 内嵌 Deep Research W&D + MCP+ token 压缩 CLI + MCPMark runner / Apache-2.0 585★ / 但 0 frontier model card 引用 | 60 KB |
| 26 | [4 MCP Bench 代码级深潜 — 一个模型能 SOTA 4 个吗](https://zhangshuibai.github.io/paper-notes/agents/notes/26_MCPBenchmarks_codedive.html) | #21 的代码级 follow-up / 实地 ls 4 repo (MCPMark/Universe/Atlas/Toolathlon) / server 重叠矩阵 (仅 GitHub 4-bench 共享 · Notion/Filesystem/Playwright 3-bench · 长尾 ~45) / 4 套 task schema verbatim (verify.py / func-chain DSL / Gemini judge claim / 真 GCP main.py) / 4 verifier 风味互斥 (严断言 / DSL / LLM judge / 真副作用) / 联合训 recipe: TOUCAN 冷启 + MCPMark+Universe RLVR + Toolathlon long-horizon SFT + Atlas LoRA 隔离 / 诚实结论: 不能一 ckpt SOTA 全部, 可以打到各 80% | 60 KB |
| 27 | [≤40B 小模型 MCP 能力训练 landscape](https://zhangshuibai.github.io/paper-notes/agents/notes/27_SmallModelMCP_landscape.html) | 复用导向 survey / 14+ 相关工作每个含完整开源清单 (Code/Data/Model/Pipeline/Eval/License) / 工业线: xLAM (Apache, 619⭐, 60K NC-data) · Hammer (Apache, ICLR 2025, function masking) · Functionary (MIT, 1.6K⭐) · NexusRaven (Apache 13B) · Granite 4.1 8B (IBM 企业) / 学术线 cross-link 5 篇: TOUCAN/EnvScaler/SETA/AWM/Agent-World / 训练框架: ⭐ART+MCP•RL (Apache, 9.5K⭐, 单 H100 / 1 小时 / $5) · Magnet (Google paper-only) · smolagents (27K⭐) · ml-intern · NeMo Toolkit / 4 条 recipe + 决策树 SVG + 开源复用矩阵 / 一句话: TOUCAN SFT 冷启 + ART MCP•RL 在线 = 单卡 H100 / 2 周 / $250 当前 ceiling | 50 KB |
| 28 | [BFCL (Berkeley Function-Calling Leaderboard)](https://zhangshuibai.github.io/paper-notes/agents/notes/28_BFCL.html) | function-calling 事实标准精读 / UC Berkeley Sky Computing Lab Gorilla 团队 / V1 2024-02-26 → V2 2024-08-19 → V3 2024-09-19 → V4 2025-07-17 / V4 重排公式 Agentic 40%+Multi-Turn 30%+Live 10%+Non-Live 10%+Irrelevance 10% / 25 个 individual category (V1 AST · V2 Live · V3 1,000 multi-turn · V4 memory_kv/vector/rec_sum + web_search + format_sensitivity) / AST checker dispatch 代码 + state-based eval / 当前 SOTA: GLM-4.5 V4 70.9% · GPT-5 V4 59.2% · Qwen3-32B V3 75.7% (≤40B ceiling) · Claude Opus 4 V3 仅 25.3 但 MCP-Atlas 77.3 (schema 不兼容) / Apache-2.0 全套 · HF 49,556 下载/月 / 唯一在 GPT-5.5 system card 显式引用的 function-call bench / 与 #19 #21 #22 #23 #26 #27 cross-link 完整 | 60 KB |
| 29 | [Don't Just Fine-tune the Agent, Tune the Environment (ICLR 2026)](https://zhangshuibai.github.io/paper-notes/agents/notes/29_TuneTheEnvironment.html) | 蚂蚁 AWorld + 浙大 + 西湖大学 + 南大 / ICLR 2026 / "tune env 不 tune agent" 范式 / <strong>3 件套</strong>: 4-stage curriculum + Actionable Env Augmentation(把 "No available route" 改成 "Invalid airport code[s]:...") + Fine-grained Progress Reward(r_t^state · r_t^exec dense per-turn)/ <strong>只用 BFCL V3 400 条样本</strong>(4 split×100) / Qwen2.5-7B 7.00→36.92 · Llama-3.1-8B 5.48→28.25 · watt-tool-8B 35.74→54.34(超 GPT-4o 51 / o3 49.25, 仍输 xLAM-2 70.50)/ OOD BFCL V4 Web Search Llama 1.00→15.00 · ACEBench ToolACE-2 8.34→15.00 / direct GRPO baseline 17.42 vs full curriculum 36.92 = <strong>+19.50 pp 纯课程效益</strong> / code: github.com/inclusionAI/AWorld-RL/EnvTuning (MIT, v2 后才放) / 局限: augmentation 字典手写不 scalable · 未评 MCP-Universe / 链 #06 #13 #18 #22 #23 #25 #26 #28 | 60 KB |
| 30 | [ACEBench (EMNLP 2025 Findings)](https://zhangshuibai.github.io/paper-notes/agents/notes/30_ACEBench.html) | USTC + 华为诺亚方舟实验室 / arXiv:2501.12851 v1 2025-01-22 → v8 2025-11-20 / EMNLP 2025 Findings, Suzhou (pages 12970–12998) / 中英双语 tool-use eval (~2,000 entries × 2 lang · 4,538 APIs · 8 domain · 68 sub-domain) / <strong>三大类八子类</strong>: Normal (Atom/SingleTurn/MultiTurn/SimilarAPI/Preference) + Special (Incomplete/Error/Irrelevant 异常输入) + Agent (Multi-Turn user-repeatedly / Multi-Step user-once) / 评测 AST + binary + sandbox state diff + Process Acc (n/m), <strong>无 LLM judge</strong> / overall acc = √n 加权抑制大类支配 / Top5 paper Table 2: GPT-4o 85.4 · GPT-4-Turbo 84.5 · Qwen2.5-Coder-32B 79.6 · Qwen-Max 78.4 · DeepSeek-V3 74.8 / Kimi K2 自报 ACEBench(En) 76.5 vs GPT-4.1 80.1 · Claude Sonnet 4 76.2 / EnvTuning #29 当 OOD: ToolACE-2 Multi-Step 6.67→20.00 / MIT license · 无 HF dataset 镜像 · leaderboard 自 2025-07-21 静默 / 与 ToolACE(2409.00920, SFT 数据集)同源不同物 / 链 #22 #23 #28 #29 | 65 KB |
| 31 | [🏭 Prime Intellect Environments Hub (Industry Analysis)](https://zhangshuibai.github.io/paper-notes/agents/notes/31_PrimeIntellectHub.html) | agentic RL 的 GitHub / 2025-08-27 launch / <strong>1,000+ env · 250+ creator · 100,000+ download</strong> (2026-05 官方) / 三层栈: verifiers SDK (MIT 4,113★ · SingleTurnEnv/MultiTurnEnv/ToolEnv/StatefulToolEnv/BrowserEnv/OpenEnv 桥接 · v1 Env(taskset,harness) API) + Hub 分发 (env 打成 Python wheel via pyproject.toml + hatchling) + prime-rl 训练器 (Apache 1,379★ · async · FSDP2+vLLM+FP8+EP+PD-disagg · INTELLECT-2/3 同栈) / <code>prime env init/install/push</code> CLI 5 行上手 / 2026-02-10 Lab GA: hosted training agentic RL + LoRA · NVIDIA Dynamo · <strong>per-token 计费</strong> · SFT/GEPA/GKD/DPO 列为 near future / Wordle/AIME-25/terminal-bench/alphabet-sort 教科书 env (HF blog walkthrough: 2×A6000 8h GRPO Qwen3-0.6B +43% reward) / 诚实警示: <strong>verifiers 主线无 MCPEnv</strong> · Hub MCP 数量未在公开来源确认 · 1,000 env 质量方差大 · Lab beta 仅 LoRA · 链 models/#07 #08 + agents/#06 #18 #20 #22 #23 #24 #25 #27 #29 | 50 KB |
| 32 | [DockSmith (StepFun)](https://zhangshuibai.github.io/paper-notes/agents/notes/32_DockSmith_2602.00592.html) | arXiv:2602.00592 v1 2026-01-31 → v2 2026-04-28 / StepFun · 专吃 Docker 环境构建的 agentic 模型 / base Qwen3-Coder-30B-A3B-Instruct + SFT only / 数据来自 <strong>SWE-Factory pipeline</strong>(4 agent: Context Retrieval / Dockerfile / Eval Script / Test Analysis)+ 两扩展(<strong>loop-detection controller</strong> 打破 agent 死循环 + <strong>cross-task success memory</strong> 跨任务复用已验证 Dockerfile)/ trajectory filter (acceptance shaping) + Score=0.5L+5R+3P complexity curriculum (Easy:Med:Hard=1:2:2) + joint training with Nex Agent-SFT (SWE:Docker 最佳 1:0.5 或 1:1) / <strong>Multi-Docker-Eval 开源 SOTA 39.72% F2P + 58.28% Commit Rate</strong>(base 19.46 → +20.26 pp · 超 DeepSeek-v3.1 37.72 / Kimi-K2-0905 37.62 / Claude-Sonnet-4 35.53 / GPT-5-Mini 34.13)/ 语言级 Python/JS/Java/Go/Ruby/Rust 全部第一,C 反退步 0%、C++/Java 仍低位 / 3 OOD 涨幅: SWE.V 49.65→51.90 (+2.25) · SWE.M 31.83→33.92 (+2.09) · TB2.0 10.67→14.04 (+3.37) / 错误分析: Total Errors −42.5%, Test Analysis stage −50.6%, Context Retrieval +10.9% (more aggressive recall) / Figure 4 within-layer persistence: P(env→env) +6%, P(runtime→runtime) +7% / Table 7 principled reasoning 27.1%→33.0% / 开源现状: 论文 abstract "<em>publicly available at here</em>" 占位符空 · HF 仅匿名账号 <code>8sj7df9k8m5x8/DockSmith</code>(model)+ <code>docker_building_training</code>(dataset Apache-2.0 39,719 samples / 2,876 instances)· 训练代码 / pipeline 实现 / loop-detection 算法 / cross-task memory retrieval 全未公开 / 局限: loop-detection + cross-task memory 两个"contribution"都<strong>无 ablation</strong> · OOD 涨幅小数量级,可能含源数据软污染 · 无 RL · MDE 协议 alignment 友好 / 链 #18 #20 #22 #23 #25 #29 + Nex-N1 (2512.04987) + SWE-Factory (2506.10954) + MDE (2512.06915) | 65 KB |
| 33 | [OpenEnv ← Agent World Model (Meta + Snowflake)](https://zhangshuibai.github.io/paper-notes/agents/notes/33_OpenEnv_AWM.html) | Meta PyTorch OpenEnv 把 Snowflake <strong>AgentWorldModel-1K (1,000 scenario × 10 task)</strong> 包装成标准 OpenEnv server / wire = WebSocket + FastAPI + JSON Action/Observation / <code>uvicorn ... envs.agent_world_model_env.server.app:app</code> 起 server / client <code>async with AWMEnv(base_url=...) as env</code> / Action 只有 <code>ListToolsAction()</code> 和 <code>CallToolAction(tool_name, arguments)</code> + 3 个保留 tool: <code>verify</code> / <code>done</code> / <code>__list_scenarios__</code> / <strong>verifier 两模式 = 同 tool 不同 args</strong>: <code>code</code> (默认 · SQLite 状态 diff + Python · 无 LLM 免费) / <code>sql</code> (SQL 提 diff + LLM-judge · <code>OPENENV_AWM_LLM_BASE_URL/API_KEY/MODEL</code> env var) / default reward = {complete:+1.0, incomplete:+0.1, format_error:−1.0, server_error:0.0} / reset 可覆盖 <code>reward_config</code> / done(keep_session=True) 留 5 文件 artifact: trajectory.json + {scenario}.db + {scenario}_initial.db + server.py + server.log / <strong>10,000 默认 MAX_CONCURRENT_ENVS</strong> · 3,000 idle 阈值 · 600s cleanup · 5s 扫描 / 1,024 episode × 64 concurrency × 3-20 turn stress test 脚本 / 公开 HF Space <code>ChilleD/agent_world_model_env</code> 零部署 / OpenEnv 自身 BSD-3 + AWM 数据 Apache-2.0 = 全栈商用 friendly / 跟 #18 原版差异: protocol WebSocket 化 + verify 内化为 tool (reward leak 风险) + subprocess 隔离 + idle cleanup + HF hosting / 跟 #31 verifiers 互桥 (OpenEnvAdapter) / 警告: scenario subprocess <strong>无 per-process sandbox</strong>,仅靠 Docker 兜底,官方"only academic research use" / 未确认 5 项 (pip path · Dockerfile · tool 总数 · default mode 翻转 · verify 调用次数限制) / 链 #18 #20 #22 #23 #29 #31 | 53 KB |
| 34 | [ToolMind (Nanbeige / BOSS Zhipin × RUC)](https://zhangshuibai.github.io/paper-notes/agents/notes/34_ToolMind_2511.15718.html) | arXiv:2511.15718 v1 2025-11-12 → v2 2025-12-05 / Nanbeige Lab BOSS 直聘 + 人大高瓴 Wayne Xin Zhao 组 / <strong>360K tool-agentic SFT</strong>(精确 368,611 = 163,180 graph-synth + 205,431 augmented open-source 跨 7 源: xLAM-60K 101K · Glaive-v2 20K · ToolACE 7.3K · BUTTON 21K · APIGen-MT 25K · When2Call 17K · τ-bench train 12.8K)/ <strong>方法 3 件套</strong>: (1) <strong>function graph + random walk</strong>(20K function · 参数级 embed · sᵢⱼ=max cos · τ 阈值+LM validator+随机扰动 · chain L~U{5..20} per-node visit limit) (2) <strong>三角色 multi-agent simulator</strong>(user incremental disclosure ###STOP### / assistant &lt;think&gt;+content+tool_calls / tool 模拟 server 含 error mode · <strong>tools NOT executable</strong>)(3) <strong>双层 binary filter</strong>: trajectory-level + <strong>turn-level mask 不丢</strong>(novelty)/ <strong>BFCL-v4</strong>: Qwen3-8B 42.21→46.92 (+4.69) · 14B 45.14→50.54 (+5.40) Multi-Turn +6.88 Agentic Search +23.00 但 Agentic Memory 14B 反退 −3.44 / <strong>τ-bench</strong>: 8B 35.83→46.70 (+10.87) · 14B 38.78→<strong>53.00</strong> (+14.22) retail 域 60.00 最强 / <strong>τ²-bench</strong>: 8B 34.64→46.41 (+11.77) · 14B 40.65→49.08 (+8.43) airline 14B +20.00 但 telecom 14B −1.75 / <strong>Table 4 消融</strong>: turn-filter 在 τ-bench +7.00 BFCL +2.76 <strong>但 τ²-bench −2.88(反例)</strong> / 训练: Qwen3-8B/14B + OpenRLHF + 64K seq + bs64 + lr 5e-6 + warmup 0.03 + loss only on anchored assistant / <strong>开源状态</strong>: HF dataset <code>Nanbeige/ToolMind</code> Apache-2.0 (2,759 下载 / 147 likes / 7.87GB / 8 jsonl split) 但<strong>含 ToolACE CC-BY-NC 派生</strong>实际是混 license · <strong>代码 / pipeline / judge LM 未公开</strong> · paper 主线 ckpt <strong>未发布</strong>(后续衍生 <code>ToolMind-Web-3B</code> 1,622 下载 转 deep-search Llama-base 不是 Qwen3)/ <strong>局限</strong>: turn-judge 用什么 LM 论文未说(可能 GPT-4o 但未确认) · binary 0/1 信号粗(改 weighted/3-class 应更好) · tool 是 LM 模拟非真 server(没在 MCP-Universe/MCPMark 真 server 评测) · NC 来源 license 混合 · "20K tools" 实为 schema 而非 callable / <strong>定位</strong>: TOUCAN 的方法论补丁版(turn-level mask 替代 trajectory-level Likert · graph-walk 替代 5-LLM 出题 · simulated tool 替代真 MCP) / 链 #18 #20 #22 #23 #28 #29 #30 #31 | 60 KB |
| 35 | [ToolACE (Huawei Noah's Ark + USTC)](https://zhangshuibai.github.io/paper-notes/agents/notes/35_ToolACE_2409.00920.html) | arXiv:2409.00920 v1 2024-09-02 → v2 2025-07-25 · <strong>ICLR 2025</strong> · 27 作者 Weiwen Liu/Xingshan Zeng/Yasheng Wang.../Enhong Chen / <strong>26,507 合成 API · 11,300 公开 dialog (HF subset, 完整训练集 ≥180K 未公开)</strong> / <strong>3 件套 TSS/SDG/DLV</strong>(<em>不是</em> TAS/SAD/DLA): (1) <strong>TSS Tool Self-Evolution Synthesis</strong> = speciation (pretraining 文档抽 API context tree) → adaptation (子树切 functionality) → evolution (4 类 diversity indicator 变异 + example buffer) (2) <strong>SDG Self-Guided Dialog Generation</strong> = <em>待训 LLM 当 complexity evaluator</em> (用 sample loss 当难度信号) + 3 agent (user/assistant/tool) multi-agent, 4 类 dialog (single/parallel/dependent/non-tool-use), tool 由 LM 模拟<strong>不接真 server</strong> (3) <strong>DLV Dual-Layer Validation</strong> = rule checker (4 维 API def/exec/dialog/consistency) + model checker (任务分解到 multiple expert agent) · ablation w.o.dual &lt; w.o.model &lt; Final / <strong>训练</strong> LLaMA-3.1-8B-Instruct full SFT → <strong>ToolACE-8B</strong> · <strong>BFCL-v3 Rank 3 / Overall 59.22 / Single AST 89.27 / Exec 90.07 / Rel 85.37 / Irrel 83.81 / Multi-turn 14.37 (弱点)</strong> · API-Bank "<em>comparable with GPT-4 series</em>" · MMLU/GSM8K/HumanEval/CSQA 通用能力无 degradation (vs xLAM-7b-fc-r 显著高) / <strong>🚨 LICENSE 校订verdict</strong>: <span style="color:#1f7a4a">HF API + README header 双验证 <code>Team-ACE/ToolACE</code> dataset + <code>ToolACE-8B</code>/<code>2-8B</code>/<code>2.5-8B</code> 三 model <strong>全 Apache-2.0</strong></span>(实地 curl 验证), <strong>不是</strong> CC-BY-NC-4.0 — 这意味着 #22 / #31 §10.1.1 / #34 中关于 "ToolACE 是 NC" 的论断<strong>错的</strong>, 需更正; ToolMind 含 7,327 条 ToolACE 派生<strong>没有 license 冲突</strong> / 真 NC 的同期数据是 xLAM-60K + Hammer-data, ToolACE 不是 / <strong>开源短板</strong>: pipeline 代码全闭(github 全 404 无 official repo, 只有数据 + 模型 weights) / <strong>团队衍生链</strong>: ToolACE-R (2504.01400 adaptive refinement) → ToolACE-DEV (2505.07512 decomposition+evolution) → ToolACE-MT (2508.12685 non-AR multi-turn) + 姐妹 paper ACEBench (#30) / <strong>community fork</strong>: watt-tool-8B (Apache, +DMPO multi-turn 偏好, 被 #29 EnvTuning 当 baseline 35.74→54.34) / 链 #22 #27 #28 #29 #30 #31 #34 | 56 KB |
| 36 | [📱 手机本地能跑的 Agent 小模型 Landscape](https://zhangshuibai.github.io/paper-notes/agents/notes/36_MobileAgents_landscape.html) | iOS+Android 端侧 / ≤4B base 模型 + mobile-agent 专训 + benchmark + 推理框架 全梳理 · 姊妹于 #16 桌面版 / **base 模型矩阵**(Apple Foundation ~3B 30 tok/s 官方 · Gemini Nano 3.25B via AICore · Qwen2.5-VL-3B Apache-2.0 phone-use · Phi-3.5-mini · Gemma 3 1B/4B · LLaMA 3.2 1B/3B · MiniCPM-V · MobileLLM 125M/350M · SmolLM2 1.7B IFEval 56.7 BFCL 27%)/ **mobile-agent 专训** (Octopus v2 2B functional-token 35× over Llama-7B · Mobile-Agent v1→v2→v3.5 阿里 X-PLUG / GUI-Owl 2B/4B/8B/32B Qwen3-VL base · DigiRL Stanford 1.3B AitW 17.7→**67.2** offline-to-online RL · UI-TARS-1.5 7B 跨平台)/ **benchmark** (AndroidWorld 116 task × 20 app paper 30.6% 现 leaderboard 70%+ · AitW 715K episode × 30K 指令 · AndroidControl 15.3K demo × 833 app NeurIPS 2024 · GUI-Odyssey · Mobile-Eval · B-MoCA · SPA-Bench)/ **推理栈** (MLC-LLM TVM Apache iOS+Android SDK · ExecuTorch Meta BSD 12+ 后端 CoreML/XNNPACK/Qualcomm/MediaTek 已上 IG/WA/Quest 3 · MediaPipe LLM Inference Google · Apple Core ML + ANE · MLX · llama.cpp + GGUF · MNN · NCNN)/ **硬件约束** (iPhone 15 Pro 8 GB 第三方 app ~3 GB hard / Increased Memory Limit entitlement ~5 GB / int4 必修 / NPU 内存带宽 bound) / **决策树 SVG** + 实测 tok/s 表 (除 Apple 30 之外多为社区数量级估计标 "~估计") / **诚实分析**: Siri/Google Assistant 仍走云 = 1-3B multi-turn 弱+VLM 太贵+iOS 跨 app 权限闭 / GUI agent 比 function-call 难一量级 (VLM 3-7B vs LLM 0.5-3B · 3+ GB RAM vs 1.5 / 每 step 1-3s vs &lt;1s) / 链 #14 #15 #16 #17 #27 #29 #31 | 54 KB |
| 37 | [AgentDistill (Princeton Mengdi Wang)](https://zhangshuibai.github.io/paper-notes/agents/notes/37_AgentDistill_2506.14728.html) | arXiv:2506.14728 v1 2025-06-17 / Princeton AI Lab + UMich + IIIS Tsinghua + SJTU + Columbia + CUHK + TCCI + UT Austin · Jiahao Qiu, Xinzhe Juan, Yimin Wang, Ling Yang... <strong>Mengdi Wang 通讯</strong> / <strong>"training-free agent distillation"</strong> 范式立项: ∇θπ_S = 0 student 权重<strong>完全冻结</strong>, 把"agent capability"从权重搬到运行时<strong>工具库</strong> / 同源于 Alita [20] / <strong>三阶段</strong>: (1) Teacher = Claude-Sonnet-4 (Manager) + GPT-4o (MCP creator, 4 个 sub-section Brainstorm→OpenSourceSearch→ScriptGen→VirtualExec) 解题时<em>顺手</em>生成 @mcp.tool() Python 脚本 → 仅留 π_T(x)=y 成功 traj + 可执行 MCP 得 pool L (2) <strong>MCP-Box construction</strong> 由 Claude-Sonnet-4 三步: <strong>Abstract</strong>(留 ≤3 critical param 公式 4)→ <strong>Cluster</strong>(按 functional semantic 聚 K cluster 公式 5)→ <strong>Consolidate</strong>(union param + if-branch 合并成 super-tool 公式 6)→ B={(MCP_k^final, cluster_name_k)} (3) <strong>SmolAgents framework</strong> 整 box mount 到 student tool interface <em>无 retrieval 无 reranking 无 param selection</em> · MCP 实为 <strong>FastMCP 同进程 Python 函数</strong>(不接真 cross-process server) / <strong>实测 K 极小</strong>: PathVQA 9 · SLAKE 13 · Game-of-24 仅 <strong>1</strong> / Table 1 calling rate: Qwen3-8B SLAKE 94.7% · GPT-3.5 PathVQA 38.0% · LLaMA-3.1-8B PathVQA 24.3% / Table 2 <strong>9/9 全正向</strong>: GPT-3.5 Game24 34.3→82.7 (<strong>+48.4</strong>)· LLaMA-3.1-8B Game24 21.7→64.0 (<strong>+42.3</strong>) · LLaMA SLAKE 49.3→59.3 (+10.0) · Qwen3-8B Game24 ceiling +7 / Table 3 跨 agent: student-after-distill 在 PathVQA <strong>52.7 险胜 teacher 52</strong> 但 Game24 teacher 99 vs student 75.5 (solver 在但 student 仍读错数字)· OctoTools(GPT-4o) 49/64/45 被均压 / Figure 5 brain-MRI case study: 两 narrow MCP(亮区 / 左半边)consolidate 成 region+analysis_mode+3 threshold 的 super-tool / <strong>开源</strong>: <strong>无</strong> · GitHub <code>shimu0215/AgentDistill</code> 是 <strong>另一篇 2505.17612 (Nardien et al.)</strong> 撞名 · arXiv "Code, Data and Media" tab 空 · HF 0 关联 / <strong>诚实批判</strong>: (a) "training-free" 是 student-side 0 grad 但 teacher-side Claude-4+GPT-4o token cost 未报, 数量级 $100-500/dataset (b) box 整挂无 retrieval 仅在 K≤13 成立, 不可 scale 到 BFCL 全 server (c) Game24 +48 实测 program-execution offload 而非 capability transfer, 确定性 puzzle 写 solver 就到顶 (d) 100 example box-gen 与 eval set 的 overlap 未明示, 可能 train=test (e) 跨 domain transfer 完全没测 — 每 box per-benchmark 训, "generalizable" 标题口号大于证据 (f) 缺 vs TOUCAN-SFT-8B / vs prompt-only baseline 的 head-to-head / <strong>定位</strong>: TOUCAN/ToolMind/ToolACE 是<em>权重派</em>(SFT 蒸馏), AgentDistill 是<em>库派</em>(工具运行时蒸馏), EnvTuning 是 env 派 — agent capability 应该放在哪是这一波 MCP 浪潮的根本二选一 / 链 #22 #27 #29 #31 #34 #35 + Alita | 46 KB |
| 39 | [MCP-Flow (TikTok + SJTU + ZJU + UIUC, ACL 2026 Main)](https://zhangshuibai.github.io/paper-notes/agents/notes/39_MCPFlow_2510.24284.html) | arXiv:2510.24284 v1 2025-10-28 → v3 2026-04-16 · **ACL 2026 Main 接收** / Wenhao Wang(TikTok 实习 / ZJU)+ SJTU(陈思衡组)+ UIUC / <strong>1,166 真实 MCP server</strong>(发现)/ 356 (可调通) <strong>· 11,536 tool · 68,733 instruction-fn-call pair · 6,439 trajectory</strong> — TOUCAN 的 5.8× tool / 2.4× server,但 traj 数砍到 TOUCAN 1.5M 的 0.4% / <strong>6 marketplace 跨平台爬取</strong>(Smithery 40.2% + MCP.so 19.5% + Glama 13.6% + PulseMCP 9.3% + PipeDream 9.3% + MCPHub 7.2% + Manual)用 <strong>Playwright web-agent</strong> 增量抓 JSON 配置 → 本地 npm/uvx/SSE 起服(dolphin-mcp client)→ 排除需 API key / 专有软件的 server / <strong>合成 5 阶段</strong>: GPT-4o few-shot(每 tool 5 ins)→ slot-fill regex 修补 → <strong>WizardLM evolve depth=2</strong>(concretization/reasoning)→ GPT-4o fn-call → 本地 server 真调拿 response → 三重过滤(embed sim &gt;0.8 丢 + GPT-4o+DeepSeek-V3 双投票 tool selection + DeepSeek-V3 quality &lt;6 丢)/ <strong>96-99% 标签正确率</strong>(300 sample 人工抽查)/ <strong>主表 Smithery</strong>: MCP-Flow-Qwen3-0.6B unseen-server AST 58.0 vs Claude-4-Sonnet 48.4(+9.6pp)· Llama3.1-8B 65.2 vs GPT-4o 50.8(+14.4pp)· MCP-Flow-Qwen-4B 候选 tool=100 时 AST 67.0 vs Claude 51.6 · ToolACE-8B Tool 89.4 高但 Param 49.8 低(真 server 数据 vs schema 合成的清晰对照)/ <strong>BFCL V3</strong>: Qwen3-4B+MCP-Flow Multi-Turn 31.38 / Non-Live 85.58 / Live 78.09 — Single-turn 跑赢 TOUCAN-7B(78.52/74.50)但 Multi-Turn 输 ToolACE-2-8B(38.38)/ <strong>GAIA</strong>: Qwen3-4B 10.68→<strong>21.36 (+100%)</strong> · GPT-4o 29.13→33.98 (+17%) · Claude-4-Sonnet 55.34→57.28 (+4%) 把 MCP-Flow 当 initial fn-call generator / <strong>训练</strong>: LoRA + LLaMA-Factory + 候选 tool=10 + 1 epoch / Qwen3-0.6B/4B + Llama3.1-8B 三档 / <strong>开源</strong>: GitHub <code>wwh0411/MCP-Flow</code> + HF <code>wwh0411/MCP-Flow</code> dataset(<strong>无 LICENSE 文件,HF cardData=None,license 未声明</strong> 2026-05 实测)· 训练 ckpt <strong>未发布</strong>(论文宣传 contribution 但 HF 上没有)/ <strong>诚实点</strong>: (a) 1,166 是 discovery scale,executable scale 仅 356 = TOUCAN 495 的 71% (b) 论文 Table 1 立项<strong>没把 TOUCAN 列出来</strong>(同期 paper,2025-10) (c) 没在 MCP-Universe / MCP-Atlas / Toolathlon / MCPMark 上报数 (d) Multi-Turn 弱是结构性问题(traj 数砍 235×)/ <strong>定位</strong>: TOUCAN 深×MCP-Flow 广 = 真 MCP 数据栈两极 · 都 from Smithery 上游(#24) · 推荐 SFT recipe: MCP-Flow 68K(广 fn-call) + TOUCAN 119K(深 multi-turn) 合并 → ART MCP•RL / EnvTuning 再 RL / 链 #18 #20 #22 #24 #25 #27 #28 #29 #34 #35 #38 | 59 KB |
| 38 | [🗺️ MCP Post-Training Landscape (2026 H1 synthesis)](https://zhangshuibai.github.io/paper-notes/agents/notes/38_MCPPostTraining_2026_landscape.html) | **2026-05 landscape synthesis 注(非 paper 精读)** / cross-link #03/#18/#19/#20/#21/#22/#23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37 共 20+ MCP-adjacent deep-read · 系统化 2026 H1 新 12 个 entries 不在已有笔记内: <strong>MCP-Flow</strong>(2510.24284, 1,166 server / 11,536 tool / 68K traj) · <strong>FireFly</strong>(2605.17558, verify-first 反向 pipeline) · <strong>ComplexMCP</strong>(2605.10787, tool 依赖+env noise) · <strong>MCPHunt</strong>(2604.27819, 第一个 MCP 安全 bench) · <strong>EnvFactory</strong>(2605.18703, robust RL) · <strong>ATLAS</strong>(2603.06713, large toolspace SLM RL) · <strong>InfTool</strong>(2512.23611, 自演化合成) · <strong>TDScaling</strong>(2602.03219, trajectory diversity) · <strong>CUBE</strong>(2603.15798, 元 bench 统一 schema) · <strong>TPS-Bench</strong>(2511.01527) · <strong>TOBench</strong>(2605.16909, vision+tool 三合一) · <strong>EvoMemBench/CHI-Bench/Herculean</strong> 等 domain-specific bench / **10 章结构**: §1 SVG L1-L5 全景 → §2 数据(真 MCP/合成/schema-only/license 净化表) → §3 RL infra(MCP-Universe verl/ART 单卡/verifiers/OpenEnv 四档原生 + NeMo-RL/rLLM/OpenRLHF 通用) → §4 评测(5 大已 covered + 9 个 2026 新进) → §5 模型 + HF 社区 2026 创建 MCP-tag 表 → §6 leaderboard 2026-05 snapshot(BFCL V4 Top-10 + 4 MCP bench top-1) → §7 8 条 trend(A 真 MCP vs 合成 / B SFT vs RL / C cold-start 两阶段 / D frontier 引用 / E tokenizer 添 tool_call token / F frontier gap 缩 vs 扩 / G mobile MCP / G' vertical bench / H training-free 库派) → §8 5 recipe 决策树 SVG + 单 H100 单周 / 8H100 单月命令行步骤 → §9 #03-#37 cross-link 一覧表 → §10 open question / **关键校订**: ToolACE 是 Apache-2.0 不是 CC-BY-NC(实地 HF curl 验证 #35,#22/#31/#34 此前误判) / **诚实警示**: ≤40B 开源模型在 4 大 MCP long-horizon bench (Universe/Atlas/Mark/Toolathlon) 至今 0 top-10 record · 真实 sys card 引用只 MCP-Atlas (Claude Opus 4.7 77.3%) + BFCL (GPT-5.5) 两家 · Smithery 7,300+ server (2026-05) 是上游 registry / **本 landscape 不是 paper**,与 #16 (DesktopAgents) / #21 (4 MCP Bench 横评) / #24 (Smithery industry) / #26 (4 bench codedive) / #27 (≤40B MCP) / #31 (Prime Intellect hub) / #36 (mobile) 同性质 industry-survey style | 70 KB |

## 论文清单 (按时间)

### Speculative Decoding × RL

| 论文 | 团队 | 日期 | arXiv |
|---|---|---|---|
| EAGLE-1 | SJTU et al | 2024-01 | [2401.15077](https://arxiv.org/abs/2401.15077) |
| EAGLE-2 | SJTU et al | 2024-06 | [2406.16858](https://arxiv.org/abs/2406.16858) |
| EAGLE-3 | SJTU et al | 2025-03 | [2503.01840](https://arxiv.org/abs/2503.01840) |
| SPEC-RL | Shopee | 2025-09 | [2509.23232](https://arxiv.org/abs/2509.23232) |
| ReSpec | NTU + 阿里 | 2025-10 | [2510.26475](https://arxiv.org/abs/2510.26475) |
| DAS | Together AI + UIUC | 2025-11 | [2511.13841](https://arxiv.org/abs/2511.13841) |
| SRT | — | 2026-01 | [2601.09083](https://arxiv.org/abs/2601.09083) |
| DFlash | — | 2026-02 | [2602.06036](https://arxiv.org/abs/2602.06036) |
| NVIDIA NeMo-RL + EAGLE-3 | NVIDIA | 2026-04 | [2604.26779](https://arxiv.org/abs/2604.26779) |
| Hidden States Drift (KVShot) | — | 2026-04 | [2604.26412](https://arxiv.org/abs/2604.26412) |
| Mirror-SD | — | 2025-10 | [2510.13161](https://arxiv.org/abs/2510.13161) |
| SSD (Speculative Speculative Decoding) | Tri Dao 等 | 2026-03 | [2603.03251](https://arxiv.org/abs/2603.03251) |
| Performance or Illusion? | — | 2026-01 | [2601.11580](https://arxiv.org/abs/2601.11580) |
| DropMatch | — | 2026-03 | [2603.03333](https://arxiv.org/abs/2603.03333) |
| Speculative Verification (SV) | — | 2025-09 | [2509.24328](https://arxiv.org/abs/2509.24328) |
| LK Losses | — | 2026-02 | [2602.23881](https://arxiv.org/abs/2602.23881) |
| DDTree | — | 2026-04 | [2604.12989](https://arxiv.org/abs/2604.12989) |
| Orthrus | Adobe Research + Univ of Oregon + Google DeepMind | 2026-05 | [2605.12825](https://arxiv.org/abs/2605.12825) · [GitHub](https://github.com/chiennv2000/orthrus) · [HF Orthrus-Qwen3-8B](https://huggingface.co/chiennv/Orthrus-Qwen3-8B) |
| (经典) Leviathan et al. — Fast Inference via SD | Google | 2022-11 | [2211.17192](https://arxiv.org/abs/2211.17192) |
| (经典) Chen et al. — Accelerating LLM Decoding | DeepMind | 2023-02 | [2302.01318](https://arxiv.org/abs/2302.01318) |

### 前沿模型 / 训练系统 / LM 方法

| 论文 | 团队 | 日期 | arXiv |
|---|---|---|---|
| INTELLECT-1 | Prime Intellect | 2024-12 | [2412.01152](https://arxiv.org/abs/2412.01152) |
| MiniMax-01 | MiniMax | 2025-01 | [2501.08313](https://arxiv.org/abs/2501.08313) |
| INTELLECT-2 | Prime Intellect | 2025-05 | [2505.07291](https://arxiv.org/abs/2505.07291) |
| MiniMax-M1 | MiniMax | 2025-06 | [2506.13585](https://arxiv.org/abs/2506.13585) |
| LongLLaDA | OpenMOSS / SJTU | 2025-06 | [2506.14429](https://arxiv.org/abs/2506.14429) |
| OmniHuman 1.5 | ByteDance | 2025-08 | [2508.19209](https://arxiv.org/abs/2508.19209) |
| TOP | Zuhri et al | 2025-08 | [2508.19228](https://arxiv.org/abs/2508.19228) |
| INTELLECT-3 | Prime Intellect | 2025-12 | [2512.16144](https://arxiv.org/abs/2512.16144) |
| MiMo-V2-Flash | Xiaomi LLM-Core | 2026-01 | [2601.02780](https://arxiv.org/abs/2601.02780) |
| MARS | gradient + 港大 + McGill 等 | 2026-01 | [2601.15498](https://arxiv.org/abs/2601.15498) |
| NCP / ConceptLM | — | 2026-02 | [2602.08984](https://arxiv.org/abs/2602.08984) |
| Kimi K2.5 | Kimi Team | 2026-02 | [2602.02276](https://arxiv.org/abs/2602.02276) |
| TIDE | PKU YuanGroup | 2026-04 | [2604.26951](https://arxiv.org/abs/2604.26951) |
| MTP (Gloeckle, Meta) | Meta FAIR | 2024-04 | [2404.19737](https://arxiv.org/abs/2404.19737) |
| MuToR (Registers) | — | 2025-05 | [2505.10518](https://arxiv.org/abs/2505.10518) |
| L-MTP (Leap MTP) | NeurIPS'25 | 2025-05 | [2505.17505](https://arxiv.org/abs/2505.17505) |
| FastMTP | Tencent BAC | 2025-09 | [2509.18362](https://arxiv.org/abs/2509.18362) |
| FSP (Beyond MTP) | — | 2025-10 | [2510.14751](https://arxiv.org/abs/2510.14751) |
| CoPD (Co-Evolving Policy Distillation) | — | 2026-04 | [2604.27083](https://arxiv.org/abs/2604.27083) |
| Prism (test-time scaling for dLLM) | — | 2026-02 | [2602.01842](https://arxiv.org/abs/2602.01842) |
| veRL/HybridFlow | ByteDance | 2024-09 | [2409.19256](https://arxiv.org/abs/2409.19256) |
| AReaL | 清华 + 蚂蚁 | 2025-05 | [2505.24298](https://arxiv.org/abs/2505.24298) |
| LlamaRL | Meta | 2025-05 | [2505.24034](https://arxiv.org/abs/2505.24034) |
| Laminar | — | 2025-10 | [2510.12633](https://arxiv.org/abs/2510.12633) |
| Retrofitting Recurrence | UMD / NYU / LLNL et al. | 2025-11 | [2511.07384](https://arxiv.org/abs/2511.07384) |
| RePlaid (continuous DLM scaling) **(本人参与)** | NVIDIA + Cornell + UW–Madison + MBZUAI + GaTech | 2026-05 | [2605.18530](https://arxiv.org/abs/2605.18530) |
| Nemotron-Labs-Diffusion (tri-mode LM) `Tech Report` | NVIDIA + GaTech + HKU + MIT | 2026-05 | [PDF](https://d1qx31qr3h6wln.cloudfront.net/publications/Nemotron_Diffusion_Tech_Report_v1.pdf?VersionId=db8_EMO8B.vmU26.jr7Le9pN3MqcUDNL) |

### Agent 训练 / 环境 / 内存

| 论文 | 团队 | 日期 | arXiv |
|---|---|---|---|
| LiteResearcher | — | 2026-04 | [2604.17931](https://arxiv.org/abs/2604.17931) |
| Agent-World | — | 2026-04 | [2604.18292](https://arxiv.org/abs/2604.18292) |
| DR-Venus | inclusionAI | 2026-04 | [2604.19859](https://arxiv.org/abs/2604.19859) |
| MemPalace 审视 | — | 2026-04 | [2604.21284](https://arxiv.org/abs/2604.21284) |
| Search-R1 | UIUC + Meta | 2025-03 | [2503.09516](https://arxiv.org/abs/2503.09516) |
| ReTool | ByteDance Seed | 2025-04 | [2504.11536](https://arxiv.org/abs/2504.11536) |
| MemoryBank | Sun Yat-Sen U + HIT + KTH | 2023-05 | [2305.10250](https://arxiv.org/abs/2305.10250) |
| Mem0 | mem0.ai | 2025-04 | [2504.19413](https://arxiv.org/abs/2504.19413) |
| MEM1 | MIT + NUS + SMART | 2025-06 | [2506.15841](https://arxiv.org/abs/2506.15841) |
| Memory-R1 | LMU Munich + TUM | 2025-08 | [2508.19828](https://arxiv.org/abs/2508.19828) |
| RLAnything | Gen-Verse | 2026-02 | [2602.02488](https://arxiv.org/abs/2602.02488) |
| ClawGUI | — | 2026-04 | [2604.11784](https://arxiv.org/abs/2604.11784) |
| UI-TARS-2 | ByteDance Seed | 2025-09 | [2509.02544](https://arxiv.org/abs/2509.02544) |
| Agent World Model | Snowflake AI Research | 2026-02 | [2602.10090](https://arxiv.org/abs/2602.10090) |
| MCP-Atlas | Scale AI + NUS | 2026-02 | [2602.00933](https://arxiv.org/abs/2602.00933) |
| SETA | CAMEL-AI + Eigent.AI + SambaNova + KAUST 等 | 2026-01 | [CAMEL Blog](https://www.camel-ai.org/blogs/seta-scaling-environments-for-terminal-agents) / [GitHub](https://github.com/camel-ai/seta) |
| TOUCAN | MIT-IBM Watson AI Lab + UW | 2025-10 | [2510.01179](https://arxiv.org/abs/2510.01179) |
| EnvScaler | RUC NLPIR (Gaoling 高瓴) | 2026-01 | [2601.05808](https://arxiv.org/abs/2601.05808) |
| Smithery (industry analysis, not a paper) | Smithery AI (Henry Mao et al.) | 2024-12 至今 | [smithery.ai](https://smithery.ai/) · [GitHub](https://github.com/smithery-ai) |
| MCP-Universe | Salesforce AI Research | 2025-08 | [2508.14704](https://arxiv.org/abs/2508.14704) |
| BFCL (Berkeley Function-Calling Leaderboard) | UC Berkeley Sky Computing Lab (Patil, Mao, Yan, Ji, Suresh, Stoica, Gonzalez) | 2024-02 (V1) → 2025-07 (V4) · ICML 2025 paper | [OpenReview](https://openreview.net/forum?id=2GmDdhBdDk) · [Leaderboard](https://gorilla.cs.berkeley.edu/leaderboard.html) · [GitHub](https://github.com/ShishirPatil/gorilla/tree/main/berkeley-function-call-leaderboard) |
| Don't Just Fine-tune the Agent, Tune the Environment | Zhejiang U + Shanghai Innovation Inst + Westlake U + AWorld Team Inclusion AI + Nanjing U | 2025-10 (v1) · 2026-01 (v2) · **ICLR 2026** | [2510.10197](https://arxiv.org/abs/2510.10197) · [code (AWorld-RL/EnvTuning)](https://github.com/inclusionAI/AWorld-RL/tree/main/EnvTuning) |
| ACEBench: A Comprehensive Evaluation of LLM Tool Usage | USTC + Huawei Noah's Ark Lab (Chen Chen, Xinlong Hao, Weiwen Liu, Xu Huang, Xingshan Zeng et al.) | 2025-01 (v1) → 2025-11 (v8) · **EMNLP 2025 Findings** | [2501.12851](https://arxiv.org/abs/2501.12851) · [GitHub](https://github.com/ACEBench/ACEBench) · [ACL Anthology](https://aclanthology.org/2025.findings-emnlp.697/) · [Leaderboard](https://chenchen0103.github.io/ACEBench/) |
| Prime Intellect Environments Hub (industry analysis, not a paper) | Prime Intellect (Vincent Weisser CEO / Johannes Hagemann CTO / Will Brown 研究) | 2025-08-27 launch · 2026-02-10 Lab GA | [Hub](https://app.primeintellect.ai/dashboard/environments) · [Blog](https://www.primeintellect.ai/blog/environments) · [verifiers](https://github.com/PrimeIntellect-ai/verifiers) · [prime-rl](https://github.com/PrimeIntellect-ai/prime-rl) · [prime CLI](https://github.com/PrimeIntellect-ai/prime) |
| DockSmith: Scaling Reliable Coding Environments via an Agentic Docker Builder | StepFun (Jiaran Zhang, Luck Ma, Fanqi Wan, Di Qi, ..., Qi Han, Xiangyu Zhang) | 2026-01-31 (v1) · 2026-04-28 (v2) | [2602.00592](https://arxiv.org/abs/2602.00592) · [HF Model (anonymous)](https://huggingface.co/8sj7df9k8m5x8/DockSmith) · [HF Dataset (Apache-2.0)](https://huggingface.co/datasets/8sj7df9k8m5x8/docker_building_training) |
| ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset | Nanbeige Lab, BOSS Zhipin + Gaoling School of AI, Renmin University of China (Chen Yang, Ran Le, ..., Wayne Xin Zhao, Yang Song, Tao Zhang) | 2025-11-12 (v1) · 2025-12-05 (v2) | [2511.15718](https://arxiv.org/abs/2511.15718) · [HF Dataset (Apache-2.0)](https://huggingface.co/datasets/Nanbeige/ToolMind) · [HF Org Nanbeige](https://huggingface.co/Nanbeige) |
| ToolACE: Winning the Points of LLM Function Calling | Huawei Noah's Ark Lab + USTC + CUHK (Weiwen Liu, Xu Huang, Xingshan Zeng, ..., Yasheng Wang, Ruiming Tang, Defu Lian, Qun Liu, Enhong Chen, 27 作者) | 2024-09-02 (v1) · 2025-07-25 (v2) · **ICLR 2025** | [2409.00920](https://arxiv.org/abs/2409.00920) · [HF Org Team-ACE](https://huggingface.co/Team-ACE) · [HF Dataset (Apache-2.0)](https://huggingface.co/datasets/Team-ACE/ToolACE) · [HF Model ToolACE-8B (Apache-2.0)](https://huggingface.co/Team-ACE/ToolACE-8B) · [OpenReview](https://openreview.net/forum?id=8EB8k6DdCU) |
| AgentDistill: Training-Free Agent Distillation with Generalizable MCP Boxes | Princeton AI Lab + UMich + IIIS Tsinghua + SJTU + Columbia + CUHK + TCCI + UT Austin (Jiahao Qiu, Xinzhe Juan, Yimin Wang, Ling Yang, Xuan Qi, Tongcheng Zhang, Jiacheng Guo, Yifu Lu, Zixin Yao, Hongru Wang, Shilong Liu, Xun Jiang, Liu Leqi, **Mengdi Wang** 通讯) | 2025-06-17 (v1) | [2506.14728](https://arxiv.org/abs/2506.14728) · *no official code/data/model released as of 2026-05* |
| MCP-Flow: Facilitating LLM Agents to Master Real-World, Diverse and Scaling MCP Tools | TikTok + 上海交通大学 + 浙江大学 + UIUC (Wenhao Wang, Peizhi Niu, Zhao Xu, Zhaoyu Chen, Jian Du, Yaxin Du, Xianghe Pang, Keduan Huang, Yanfeng Wang, Qiang Yan, **Siheng Chen** 通讯) | 2025-10-28 (v1) · 2026-04-16 (v3) · **ACL 2026 Main** | [2510.24284](https://arxiv.org/abs/2510.24284) · [GitHub wwh0411/MCP-Flow](https://github.com/wwh0411/MCP-Flow) · [HF Dataset (license unstated)](https://huggingface.co/datasets/wwh0411/MCP-Flow) |

## 设计原则

- **单文件 HTML, 无外部依赖**——飞机/无网环境可读
- 中文为主,技术术语保留英文(Transformer / draft / target / acceptance length / KV cache / GRPO …)
- 每篇含: 动机 → 背景速查 → 核心方法 → 公式逐行翻译 → 实验亮点 → 局限 → 待验证问题
- 关键机制配 SVG 示意图或动画

## 目录结构

```
.
├── README.md
├── spec-rl/        # Speculative decoding × RL  — 10 PDFs / 8 HTML
├── models/         # 前沿模型 / 训练系统 / LM 方法 — 14 PDFs / 13 HTML
└── agents/         # Agent 训练 / 环境 / 内存    — 4 PDFs / 4 HTML
```