Agent-Native Research Artifacts (ARA) Benchmark Collection
收藏资源简介:
该仓库收集了32个Agent-Native Research Artifacts (ARAs),这是一种机器可读、以可重复性为先的研究论文替代格式。每个ARA将研究论文分解为四个基础层,使人类和智能体都能阅读声明、运行代码、检查证据并重放研究探索过程(包括死胡同)。
This repository collects 32 Agent-Native Research Artifacts (ARAs), a machine-readable, reproducibility-first alternative format for research papers. Each ARA decomposes a research paper into four foundational layers, enabling both humans and AI agents to read claims, execute code, verify evidence, and replay the entire research exploration process, including dead ends.
数据集概述
该仓库收集了 32 个 Agent-Native Research Artifacts (ARA),这是一组以机器可读、可复现为优先的论文替代格式。每个 ARA 将一篇研究论文分解为四个基础层,使得人类和智能体都能阅读论文声明、运行代码、检查证据,并重放研究的实际探索过程(包括失败的路径)。
ARA 结构
每个 ARA 工件的目录结构如下:
| 层级 | 路径 | 内容 |
|---|---|---|
| 摘要 | PAPER.md |
论文的人类可读概述。 |
| 认知层 | logic/ |
包含 claims.md(声明,附有可证伪标准和证明指针)、concepts.md、experiments.md、problem.md、related_work.md,以及 solution/ 子目录下的 algorithm.md、architecture.md、constraints.md、heuristics.md。 |
| 工件层 | src/ |
可运行代码、配置文件和 environment.md。 |
| 证据层 | evidence/ |
包含 figures/ 和 tables/,每个都与所支持的声明相关联。 |
| 轨迹层 | trajectory.html |
自包含的交互式查看器:左侧为可点击的过程图,右侧为每一步的详细展开,将每一步与对应的声明、结果和代码链接起来。 |
| 跟踪层 | trace/ |
包含 exploration_tree.yaml(研究 DAG 源)和 exploration_tree.html(仅树状视图)。 |
如何浏览
- 阅读任意目录下的
artifacts/<benchmark>/<name>/PAPER.md获取概述,然后深入logic/和evidence/目录。 - 在浏览器中打开工件的
trajectory.html文件,可逐步重放研究过程:左侧为过程图(包括死胡同分支),右侧为每步详细展开(显示该步骤所做的事情、关联的声明、结果和代码指针)。由于 GitHub 会内联显示 HTML 源码,需克隆仓库或下载文件以查看渲染后的页面。
数据集构成
仓库包含以下四个子集:
Paperbench (23 个)
| 工件 | 论文标题 |
|---|---|
| adaptive-pruning | APT: Adaptive Pruning and Tuning of Pretrained Language Models for Efficient Training and Inference |
| all-in-one | All-in-one Simulation-Based Inference (Simformer) |
| bam | Batch and Match: Black-Box Variational Inference with a Score-Based Divergence |
| bbox | BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models |
| bridging-data-gaps | Efficient Transfer Learning in Diffusion Models via Adversarial Noise |
| fre | Unsupervised Zero-Shot Reinforcement Learning via Functional Reward Encodings |
| ftrl | Fine-Tuning RL Models is Secretly a Forgetting Mitigation Problem |
| lbcs | Refined Coreset Selection: Minimal Coreset Size under Model Performance Constraints |
| lca-on-the-line | LCA-on-the-Line: Benchmarking Out-of-Distribution Generalization with Class Taxonomies |
| mechanistic-understanding | A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO |
| pinn | Challenges in Training PINNs: A Loss Landscape Perspective |
| rice | RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation |
| robust-clip | Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Vision-Language Models |
| sample-specific-masks | Sample-specific Masks for Visual Reprogramming-based Prompting |
| sapg | SAPG: Split and Aggregate Policy Gradients |
| self-composing-policies | Self-Composing Policies for Scalable Continual Reinforcement Learning |
| self-expansion | Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning |
| semantic-self-consistency | Semantic Self-Consistency: Enhancing Language Model Reasoning via Semantic Weighting |
| sequential-neural-score-estimation | Sequential Neural Posterior Score Estimation (NPSE) |
| stay-on-topic-with-classifier-free-guidance | Stay on Topic with Classifier-Free Guidance |
| stochastic-interpolants | Stochastic Interpolants with Data-Dependent Couplings |
| test-time-model-adaptation | Test-Time Model Adaptation with Only Forward Passes (FOA) |
| what-will-my-model-forget | What Will My Model Forget? Forecasting Forgotten Examples in Language Model Refinement |
ReBench (5 个)
| 工件 | 论文标题 |
|---|---|
| rebench-fix_embedding | Fix Embedding (RE-Bench task) |
| rebench-nanogpt_chat_rl | nanoGPT Chat RL (RE-Bench task) |
| rebench-restricted_mlm | Restricted-Architecture MLM (RE-Bench task) |
| rebench-rust_codecontests | Rust CodeContests Inference (RE-Bench task) |
| rebench-triton_cumsum | Triton Cumsum Kernel (RE-Bench task) |
Speedrun (1 个)
| 工件 | 论文标题 |
|---|---|
| nanogpt-speedrun | NanoGPT Speedrun |
Extra (3 个)
| 工件 | 论文标题 |
|---|---|
| andes | Andes: Defining and Enhancing Quality-of-Experience in LLM-Based Text Streaming Services |
| venn | Venn: Resource Management for Collaborative Learning Jobs |
| expbench | EXP-Bench: Can AI Conduct AI Research Experiments? |
来源与质量
每个 ARA 均根据源论文及其附带的代码汇编而成,内容(声明、图表、表格)均追溯到原始论文。其中一部分还通过了 SEAL 的结构和跨层验证。这些工件是用于研究和评估目的的重构,权威结果请参考原始论文及其许可协议。
基准说明
paperbench、rebench 和 speedrun 中的工件对应于 ARA 评估基准,而 extra 工件是基准之外额外的完整 ARA。
许可协议
内容采用 Creative Commons Attribution 4.0 International (CC BY 4.0) 许可协议发布。





