External Papers &Reproductions
Tracking, reviewing, and reproducing important research papers from the broader scientific community. Our commitment to validating and building upon existing research.
Browse Papers by Research Area
Select one or more research areas to filter papers. Papers must match ALL selected areas.
Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows
By Yilun Yao, Xinyu Tan, Chao-Hsuan Liu, Yaoming Li, Zhengyang Wang, Wenhan Yu, Zhewen Tan, Yuxuan Tian, Guangxiang Zhao, Lin Sun, Xiangzheng Zhang, Tong Yang
Published in: arXiv preprint (2026)
Citation: Yao et al., 2026
Notes: Operationalizes 'Agent = Model + Harness' by evaluating 6 harnesses x 8 models on 106 tasks. Directly encodes AJL's core thesis; recommends reporting model-harness pairs.
Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses
By Jiahang Lin, Shichun Liu, Chengjun Pan, Lizhi Lin, Shihan Dou, Xuanjing Huang, Hang Yan, Zhenhua Han, Tao Gui
Published in: arXiv preprint (2026)
Citation: Lin et al., 2026
Notes: Self-evolving harness (tools/middleware/memory/prompts) via observability; lifts Terminal-Bench 2 69.7 -> 77.0% with cross-model transfer. Open code; directly reproducible on the building-agents stack.
Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation
By Sayash Kapoor, Benedikt Stroebl, Peter Kirgis, et al.
Published in: arXiv preprint (2025)
Citation: Kapoor, Stroebl, Kirgis et al., 2025
Notes: Standardized eval harness across models x scaffolds x benchmarks; 21,730 rollouts and 2.5B tokens of released logs. Most reproducible target (open harness + logs).
On the Biology of a Large Language Model
By Lindsey J et al
Published in: Transformer Circuits
Citation: Lindsey J et al. (2025). On the Biology of a Large Language Model. Transformer Circuits
Notes: Mechanistic interpretability research
Tracing the thoughts of a large language model
By Anthropic Team
Published in: Anthropic
Citation: Anthropic Team. (2025). Tracing the thoughts of a large language model. Anthropic Research
Notes: Monitor for opportunities
XAttention: Block Sparse Attention with Antidiagonal Scoring
By Xu R, Xiao G, Huang H, Guo J, Han S
Published in: arXiv
Citation: Xu R et al. (2025). XAttention: Block Sparse Attention with Antidiagonal Scoring. arXiv:2503.16428
Notes: High priority reproduction target
EM-LLM: Human-inspired Episodic Memory for Infinite Context LLMs
By Zafeirios Fountas, Martin A Benfeghoul, Adnan Oomerjee, Fenia Christopoulou, Gerasimos Lampouras, Haitham Bou-Ammar, Jun Wang
Published in: ICLR 2025
Citation: Fountas et al., ICLR 2025
Notes: Human-inspired episodic memory: event segmentation via Bayesian surprise + two-stage (similarity + temporal-contiguity) retrieval; event boundaries align with human-perceived events. Most cognitively-grounded episodic mechanism -- key reference for the human-like memory project.
Effective Prompt Extraction from Language Models
By Yiming Zhang, Nicholas Carlini, Daphne Ippolito
Published in: COLM 2024
Citation: Zhang Y, Carlini N, Ippolito D. (2024). Effective Prompt Extraction from Language Models. COLM 2024. arXiv:2307.06865
Notes: Prompt-extraction attack framework across 11 LLMs; relevant to agent/harness security (system-prompt and tool-description leakage).
MemoryBank: Enhancing Large Language Models with Long-Term Memory
By Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, Yanlin Wang
Published in: AAAI 2024
Citation: Zhong et al., AAAI 2024
Notes: Long-term memory with Ebbinghaus forgetting curve (time + relevance based updating), demoed via the SiliconFriend companion. Reference for human-like forgetting/decay in the memory system.
Reflexion: Language Agents with Verbal Reinforcement Learning
By Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao
Published in: NeurIPS 2023
Citation: Shinn et al., NeurIPS 2023
Notes: Verbal reinforcement learning: agent reflects on failures and stores episodic insights in memory for later attempts (91% pass@1 on HumanEval). Reference for episodic self-reflection memory.
Generative Agents: Interactive Simulacra of Human Behavior
By Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein
Published in: UIST 2023
Citation: Park et al., UIST 2023
Notes: Canonical 'memory stream': observations stored with timestamp + importance + embedding, retrieved by recency + importance + relevance, plus reflection. Closest prior art to the nanoagent episodic-memory project -- differentiate against this.
MemGPT: Towards LLMs as Operating Systems
By Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez
Published in: arXiv (2023)
Citation: Packer et al., 2023
Notes: Virtual context management / tiered memory (main context vs external store) with paging, inspired by OS memory hierarchy. Reference for the working-memory + retrieval layer of the nanoagent memory system.
Plan4MC: Skill Reinforcement Learning and Planning for Open-World Long-Horizon Tasks
By Haoqi Yuan, Chi Zhang, Hongcheng Wang, Feiyang Xie, Penglin Cai, Hao Dong, Zongqing Lu
Published in: NeurIPS 2023 FMDM Workshop
Citation: Yuan et al., 2023
Notes: Hybrid: RL learns basic skills, LLM builds a skill graph to plan over them. 40 tasks, many requiring >10 sequential skills; strong sample efficiency. Bridges the RL and LLM-agent camps. Open code.
Voyager: An Open-Ended Embodied Agent with Large Language Models
By Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, Anima Anandkumar
Published in: arXiv (2023)
Citation: Wang et al., 2023
Notes: Flagship LLM lifelong-learning agent in Minecraft: automatic curriculum + ever-growing code skill library + iterative prompting loop with self-verification. Black-box GPT-4, no fine-tuning. Maps directly onto building-agents (memory/loop/self-verification). Open code.
JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models
By Zihao Wang, Shaofei Cai, Anji Liu, Yonggang Jin, Jinbing Hou, Bowei Zhang, Haowei Lin, Zhaofeng He, Zilong Zheng, Yaodong Yang, Xiaojian Ma, Yitao Liang
Published in: arXiv (2023)
Citation: Wang et al., 2023
Notes: Memory-augmented multimodal agent: perceives visual input + instructions, plans, and performs embodied control. 200+ tasks; 5x reliability on ObtainDiamondPickaxe. Successor line to GITM. Open code.
Ghost in the Minecraft (GITM): Generally Capable Agents via LLMs with Text-based Knowledge and Memory
By Xizhou Zhu, Yuntao Chen, Hao Tian, Chenxin Tao, Weijie Su, Chenyu Yang, Gao Huang, Bin Li, Lewei Lu, Xiaogang Wang, Yu Qiao, Zhaoxiang Zhang, Jifeng Dai
Published in: arXiv (2023)
Citation: Zhu et al., 2023
Notes: Hierarchical LLM agent (goal -> sub-goal -> action) with text knowledge + memory. Unlocks 100% of Overworld tech tree, +47.5% on ObtainDiamond, ~10,000x more compute-efficient than VPT/DreamerV3. Open code.
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
By Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, Karthik Narasimhan
Published in: ICLR 2024
Citation: Jimenez et al., ICLR 2024
Notes: Canonical agentic coding evaluation harness (2,294 real GitHub issues). Reproduction target for agent evaluation infrastructure.
Cognitive Architectures for Language Agents (CoALA)
By Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, Thomas L. Griffiths
Published in: TMLR 2024
Citation: Sumers et al., TMLR 2024
Notes: Framework connecting modern LLM agents to classical cognitive architectures (SOAR/ACT-R). Best reference for the agentic-systems lineage; reading/reference rather than code reproduction.
MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge
By Linxi Fan, Guanzhi Wang, Yunfan Jiang, Ajay Mandlekar, Yuncong Yang, Haoyi Zhu, Andrew Tang, De-An Huang, Yuke Zhu, Anima Anandkumar
Published in: NeurIPS 2022 (Outstanding Paper)
Citation: Fan et al., NeurIPS 2022
Notes: Foundational Minecraft benchmark + simulator: 1000s of open-ended tasks, internet-scale knowledge base (videos/wiki/forums), and the MineCLIP learned reward. The environment most Minecraft agents (incl. Voyager) run in. Open source.
ReAct: Synergizing Reasoning and Acting in Language Models
By Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao
Published in: ICLR 2023
Citation: Yao et al., ICLR 2023
Notes: Foundational Think-Act-Observe (TAO) loop for LLM agents; the primitive the building-agents curriculum is built on. Baseline reproduction target.
Inner Monologue: Embodied Reasoning through Planning with Language Models
By Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, Brian Ichter
Published in: CoRL 2022
Citation: Huang et al., CoRL 2022
Notes: Closed-loop environment-feedback ('observe') precursor to ReAct, in robotics. Conceptual reference for the origin of the observe step.
The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
By Frank Rosenblatt
Published in: Psychological Review (1958)
Citation: Rosenblatt, Psychological Review, 65(6):386-408, 1958
Notes: Foundational. Takes the McCulloch-Pitts threshold neuron and adds adaptive weights + an error-driven learning rule, making it trainable from data -- the first precisely specified, trainable neural network and the birth of machine learning from neural models. Replication target: single-line Python perceptron predict + weight-update rule.
A Logical Calculus of the Ideas Immanent in Nervous Activity
By Warren S. McCulloch, Walter Pitts
Published in: Bulletin of Mathematical Biophysics (1943)
Citation: McCulloch & Pitts, Bulletin of Mathematical Biophysics, 5:115-133, 1943
Notes: Foundational. Models neurons as binary all-or-none threshold logic units and proves networks of them can compute any propositional/logical function -- the artificial neuron as a logic gate (fixed weights, no learning). Origin of the McCulloch-Pitts neuron; the direct ancestor of Rosenblatt's perceptron. Replication target: single-line Python threshold neuron.
Suggest Papers for Reproduction
Found an interesting paper that should be reproduced? Help us validate and build upon existing research by suggesting papers for our reproduction pipeline.