Elamur Research⌕
☰
home / papers / 2025 / Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper

Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper

Paper · 2025 · Atsuyuki Miyai, Mashiro Toyooka, Takashi Otonari et al.

Summary

Existing AI Scientist (autoresearch) systems aim for fully automated science without a concrete target and work only with small single-file code, so the papers they produce are weak. The authors present Jr. AI Scientist, which imitates a junior lab member who receives a paper from a mentor. Given a baseline paper's PDF, LaTeX source and codebase, the system:

  1. analyses the paper's weaknesses;
  2. generates improvement ideas and checks them for novelty;
  3. runs experiments in three stages: implementation, refinement and ablations;
  4. drafts a manuscript and revises it using LLM reviews.

The system is built on AI Scientist-v2 and hands complex multi-file code editing to modern coding agents such as Claude Code. It was tested on three baseline papers whose authors gave permission. Its generated papers scored higher under DeepReviewer than those of earlier systems. Author checks and Agents4Science reviews still revealed serious flaws, and the paper documents development risks in detail. The work is useful as a realistic picture of where autonomous research agents still need human oversight.

Key points

  • Baselines: two out-of-distribution (OOD) detection papers (NeurIPS 2023, LoCoOp; IJCV 2025) and an ICLR 2025 spotlight paper on pre-training data detection for LLMs.
  • DeepReviewer scores: Jr. AI Scientist received 5.75. Earlier systems received 3.30 (AI Scientist-v1), 2.75 (AI Scientist-v2) and 3.25 (AI Researcher).
  • Evaluation: three complementary methods were used:
    • automated scoring with DeepReviewer;
    • an author review looking for hallucinated or fabricated content;
    • submission to Agents4Science, a venue for AI-generated research.
  • Example output: one generated OOD paper proposes NPT, a nuisance-prompt method on CLIP. It reports 93.0% AUROC and 31.5% FPR95. Removing entropy regularization lowers results to 87.2% AUROC and 55.4% FPR95.
  • Reported risks:
    • gaming AI-reviewer scores;
    • unreliable citations;
    • flawed interpretation of results;
    • fabricated descriptions that are hard to detect.
  • Code and discussion: github.com/Agent4Science-UTokyo/Jr.AI-Scientist.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓The University of TokyoTokyo · Japan98 works · 9,611 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 worksSocial Sciences · 2,009 worksSafety Research · 665 worksDecision Sciences · 306 worksInformation Systems and Management · 96 works

Cite

Cite as Atsuyuki Miyai, Mashiro Toyooka, Takashi Otonari, Zaiying Zhao and Kiyoharu Aizawa. 2025. Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper. arXiv:2511.04583.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2511.04583 · arXiv 2511.04583

@misc{miyai2025jr,
  title = {{Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper}},
  author = {Atsuyuki Miyai, Mashiro Toyooka, Takashi Otonari, Zaiying Zhao and Kiyoharu Aizawa},
  year = {2025},
  doi = {10.48550/arxiv.2511.04583},
  eprint = {2511.04583},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2511.04583}
}

References

62 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sumukh K Aithal, Anirudh Goyal, Alex Lamb, Y. Bengio, and M. Mozer. Leveraging the third dimension in contrastive learning. ArXiv, abs/2301.11790, 2023. — [arXiv:2301.11790](https://arxiv.org/abs/2301.11790) find
  2. Mina Ghadimi Atigh, Stephanie Nargang, Martin Keller-Ressel, and Pascal Mettes. Simzsl: Zero-shot learning beyond a pre-defined semantic embedding space. Int. J. Comput. Vis., 133:5161–5177, 2025. find
  3. Yoshua Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. pp. 41–48, 2009.
  4. Dexia Chen, Qianjie Zhu, Weibing Li, Yue Yu, Tong Zhang, and Ruixuan Wang. Preserve and sculpt: Manifold-aligned fine-tuning of vision-language models for few-shot learning. ArXiv, abs/2508.12877, 2025. — [arXiv:2508.12877](https://arxiv.org/abs/2508.12877) find
  5. Junkun Chen, Jilin Mei, Liang Chen, Fangzhou Zhao, and Yu Hu. Proto-ood: Enhancing ood object detection with prototype feature similarity. ArXiv, abs/2409.05466, 2024. — [arXiv:2409.05466](https://arxiv.org/abs/2409.05466) find
  6. Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. 2014 IEEE Conference on Computer Vision and Pattern Recognition, pp. 3606–3613, 2013. find
  7. Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. find
  8. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. find
  9. Jiankang Deng, J. Guo, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4685–4694, 2018. find
  10. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.
  11. Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don’t know by virtual outlier synthesis. In ICLR, 2022. find
  12. Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pretrained model clip. In AAAI, 2022. find
  13. Farzad Farhadzadeh, Debasmit Das, Shubhankar Borse, and F. Porikli. Lora-x: Bridging foundation models with training-free cross-model adaptation. ArXiv, abs/2501.16559, 2025. — [arXiv:2501.16559](https://arxiv.org/abs/2501.16559) find
  14. Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. Exploring the limits of out-of-distribution detection. In NeurIPS, 2021. find
  15. Robert Geirhos, Patricia Rubisch, Claudio Michaelis, M. Bethge, Felix Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. ArXiv, abs/1811.12231, 2018. — [arXiv:1811.12231](https://arxiv.org/abs/1811.12231) find
  16. Ting Gong, Tyler Lee, Cory Stephenson, Venkata Renduchintala, Suchismita Padhy, A. Ndirango, Gokce Keskin, and Oguz H. Elibol. A comparison of loss weighting strategies for multi task learning in deep neural networks. IEEE Access, 7:141627–141632, 2019. find
  17. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. ArXiv, abs/1706.04599, 2017. find
  18. Haonan Guo, Xin Su, Chen Wu, Bo Du, and L. Zhang. Saan: Similarity-aware attention flow network for change detection with vhr remote sensing images. IEEE Transactions on Image Processing, 33: 2599–2613, 2023. find
  19. Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and S. Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. ArXiv, abs/2306.13924, 2023. — [arXiv:2306.13924](https://arxiv.org/abs/2306.13924) find
  20. Calum Heggan, S. Budgett, Timothy M. Hospedales, and Mehrdad Yaghoobi. Metaaudio: A few-shot audio classification benchmark. pp. 219–230, 2022. find
  21. Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. find
  22. Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. find
  23. Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, 2021. find
  24. Galadrielle Humblot-Renaux, Sergio Escalera, and T. Moeslund. Beyond auroc co. for evaluating out-of-distribution detection performance. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 3881–3890, 2023. find
  25. Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. find
  26. Jan Kalina. From robust neural networks toward robust nonlinear quantile estimation. Sequential Analysis, 44:326 – 350, 2025. find
  27. Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NIPS, 2017. find
  28. Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. find
  29. Saebom Leem and Hyunseok Seo. Attention guided cam: Visual explanations of vision transformer guided by self-attention. pp. 2956–2964, 2024. find
  30. Feng Li, Hao Zhang, Yi-Fan Zhang, S. Liu, Jian Guo, L. Ni, Pengchuan Zhang, and Lei Zhang. Vision-language intelligence: Tasks, representation learning, and large models. ArXiv, abs/2203.01922, 2022. — [arXiv:2203.01922](https://arxiv.org/abs/2203.01922) find
  31. Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, 2021. find
  32. Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. find
  33. Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. find
  34. Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In ICLR, 2017. find
  35. Shuo Lu, Yingsheng Wang, Lijun Sheng, Lingxiao He, Aihua Zheng, and Jian Liang. Out-of-distribution detection: A task-oriented survey of recent advances. ACM Computing Surveys, 2024a. find
  36. Xintao Lu, Yonglong Ni, and Zuohua Ding. Cross-modal sentiment analysis based on clip image-text attention interaction. International Journal of Advanced Computer Science and Applications, 2024b. find
  37. Yihao Luo, Xiang Cao, Juntao Zhang, Peng Cheng, Tianjiang Wang, and Qi Feng. Dynamic multi-scale loss optimization for object detection. Multimedia Tools and Applications, 82:2349–2367, 2021. find
  38. Siladittya Manna, Soumitri Chattopadhyay, Rakesh Dey, Saumik Bhattacharya, and U. Pal. Dynamically scaled temperature in self-supervised contrastive learning. IEEE Transactions on Artificial Intelligence, 6:1502–1512, 2023. find
  39. Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. find
  40. Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In Thirty-Seventh Conference on Neural Information Processing Systems, 2023a. find
  41. Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Zero-shot in-distribution detection in multi-object settings using vision-language foundation models. arXiv preprint arXiv:2304.04521, 2023b. — [arXiv:2304.04521](https://arxiv.org/abs/2304.04521) find
  42. Bo Peng, Yadan Luo, Yonggang Zhang, Yixuan Li, and Zhen Fang. Conjnorm: Tractable density estimation for out-of-distribution detection. ArXiv, abs/2402.17888, 2024. — [arXiv:2402.17888](https://arxiv.org/abs/2402.17888) find
  43. Gabriel Pereyra, G. Tucker, J. Chorowski, Lukasz Kaiser, and Geoffrey E. Hinton. Regularizing neural networks by penalizing confident output distributions. ArXiv, abs/1701.06548, 2017.
  44. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021.
  45. Sina Sharifi, Taha Entesari, Bardia Safaei, Vishal M. Patel, and Mahyar Fazlyab. Gradient-regularized out-of-distribution detection. pp. 459–478, 2024. find
  46. Luísa Shimabucoro, Timothy M. Hospedales, and Henry Gouk. Evaluating the evaluators: Are current few-shot learning benchmarks fit for purpose? ArXiv, abs/2307.02732, 2023. — [arXiv:2307.02732](https://arxiv.org/abs/2307.02732) find
  47. Vaibhav Singh, Paul Janson, Paria Mehrbod, Adam Ibrahim, Irina Rish, Eugene Belilovsky, and Benjamin Th’erien. Beyond cosine decay: On the effectiveness of infinite learning rate schedule for continual pre-training. ArXiv, abs/2503.02844, 2025. — [arXiv:2503.02844](https://arxiv.org/abs/2503.02844) find
  48. Shreyas Vathul Subramanian and Vignesh Ganapathiraman. Zeroth order greedylr: An adaptive learning rate scheduler for deep neural network training. In 2023 IEEE 4th International Conference on Pattern Recognition and Machine Learning (PRML), pp. 593–601, 2023. find
  49. Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. find
  50. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.
  51. Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, 2022. find
  52. Jianqiao Wangni. Convolutional networks as extremely small foundation models: Visual prompting and theoretical perspective. ArXiv, abs/2409.10555, 2024. — [arXiv:2409.10555](https://arxiv.org/abs/2409.10555) find
  53. Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. find
  54. Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 26:2056–2068, 2022. find
  55. Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021. — [arXiv:2110.11334](https://arxiv.org/abs/2110.11334) find
  56. Jingkang Yang, Pengyun Wang, Dejian Zou, Zitang Zhou, Kunyuan Ding, Wenxuan Peng, Haoqi Wang, Guangyao Chen, Bo Li, Yiyou Sun, Xuefeng Du, Kaiyang Zhou, Wayne Zhang, Dan Hendrycks, Yixuan Li, and Ziwei Liu. Openood: Benchmarking generalized out-of-distribution detection. In NeurIPS Datasets and Benc… find
  57. Zonghan Yang and Yang Liu. On robust prefix-tuning for text classification. ArXiv, abs/2203.10378, 2022. — [arXiv:2203.10378](https://arxiv.org/abs/2203.10378) find
  58. Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. find
  59. Yanpeng Zhao, Yetian Chen, Kewei Tu, and Jin Tian. Curriculum learning of bayesian network structures. pp. 269–284, 2015. find
  60. Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452–1464, 2017. find
  61. Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022a. find
  62. Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022b. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/miyai-2025-jr-ai-scientist-risk/