Anmol Kabra

I am a Computer Science PhD student at Cornell University, working on LLM post-training and AI for Science. I am fortunate to be advised by Kilian Weinberger and affiliated with AI Materials Institute.

Scientific discovery requires AI agents to generalize beyond their training distribution, which demands two things: transferable reasoning skills and specialized domain knowledge. Because curating training data for either is expensive and hard to scale, I train LLMs on synthetic data and RL environments that teach transferable skills such as knowledge composition, question decomposition, and retrieval. Along the second axis of specializing agents, I study fine-tuning and context engineering to ground LLMs in scientific domains underrepresented in pretraining.

Recently as an intern at Snorkel AI, I worked on Terminal-Bench-Science and synthetic RL training environments for scientific AI agents. Previously, I was an AI/ML Quant Intern at Bloomberg’s AI Engineering team, where I prototyped tool-use agents for the ASKB<GO> function. I was also a Research Engineer at ASAPP, working on LLMs, privacy, and anomaly detection with Ethan Elenberg and Kilian Weinberger.

I received my BS in Computer Science from Cornell University and MS from Toyota Technological Institute at Chicago (TTIC). At Cornell, I was named a Merrill Presidential Scholar for my undergraduate research with Carla Gomes and Kilian Weinberger. I was generously supported by the Tata Scholarship and Telluride Scholarship.

Fun fact: I juggle more hobbies than I can juggle number of balls 🤹‍♂️

news

2026 Oct	Awarded NVIDIA GPU Academic Grant for project: RL Training of Tool-Use Agents on Nemotron-3 with Nemo RL. Paper on arxiv: PhantomEnvironments: Training LLM Agents in Fictional Worlds with code on github. Oral award at the COLM Workshop on Learning from Situated and Embodied Interaction. Poster at COLM Workshop on Lifelong Agents: Learning, Aligning, and Evolving and Snorkel Frontier Data Summit.
2026 May	Wrote a blog with NVIDIA on RL fine-tuning: Putting the NVIDIA DGX Station to Work at Cornell. Joined Snorkel AI as AI Research Intern in San Francisco.
2026 Mar	ICLR 2026 paper invited for Spotlight talk at the NSF-NAIRR annual meeting. Outstanding Reviewer Award at the ICLR DATA-FM Workshop. 3 papers accepted to ICLR 2026 Workshops: Learning from Synthetic Data Improves Multi-Hop Reasoning: DATA-FM and VerifAI. The Reliability Gap in Agentic Evidence Verification for Materials Science: FM4Science and AI-WILD Large Multimodal Models Enable Scalable Monitoring of Aquaculture Ponds: ML for Remote Sensing
2026 Jan	Paper accepted to ICLR 2026: Learning from Synthetic Data Improves Multi-Hop Reasoning with code on github.
2025	Paper accepted to ICML 2025: PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation with code on github. Oral award at the ICML Workshop on Long-Context Foundation Models.
2024	Paper on arxiv: AiSciVision: A Framework for Specializing Large Multimodal Models in Scientific Image Classification with code on github. Started PhD at Cornell Computer Science. Paper accepted to ICML Workshop on Humans, Algorithmic Decision-Making and Society: The Limitations of Model Retraining in the Face of Performativity. Joined Bloomberg as an AI/ML Quant Intern in New York City. Paper accepted to Foundations of Responsible Computing (FORC) 2024: Score Design for Multi-Criteria Incentivization.

selected papers

* equal contributions

PhantomEnvironments: Training LLM Agents in Fictional Worlds

Anmol Kabra, Swathi Saravana Selvam, Albert Gong, Chao Wan, Christian Belardi, Dongyoung Go, Katie Z. Luo, and Kilian Q. Weinberger

Under Review (2026). Oral presentation at COLM Workshop on Learning from Situated and Embodied Interaction.

paper bibtex code

@unpublished{2026kabra_phantomenvs,
  title = {{PhantomEnvironments: Training LLM Agents in Fictional Worlds}},
  author = {Kabra, Anmol and Selvam, Swathi Saravana and Gong, Albert and Wan, Chao and Belardi, Christian and Go, Dongyoung and Luo, Katie Z. and Weinberger, Kilian Q.},
  year = {2026},
  note = {Under Review},
}

Learning from Synthetic Data Improves Multi-hop Reasoning

Anmol Kabra, Yilun Yin, Albert Gong, Kamilė Stankevičiūtė, Dongyoung Go, Johann Lee, Katie Z. Luo, Carla P. Gomes, and Kilian Q. Weinberger

In ICLR (2026). Spotlight talk at NSF-NAIRR annual meeting.

paper bibtex code poster

@inproceedings{2026kabra_phantomreasoning,
  title = {{Learning from Synthetic Data Improves Multi-hop Reasoning}},
  author = {Kabra, Anmol and Yin, Yilun and Gong, Albert and Stankevičiūtė, Kamilė and Go, Dongyoung and Lee, Johann and Luo, Katie Z. and Gomes, Carla P. and Weinberger, Kilian Q.},
  year = {2026},
  booktitle = {The Fourteenth International Conference on Learning Representations},
}

PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation

Albert Gong*, Kamilė Stankevičiūtė*, Chao Wan*, Anmol Kabra*, Raphael Thesmar, Johann Lee, Julius Klenke, Carla P. Gomes, and Kilian Q. Weinberger

In ICML (2025). Oral presentation at ICML Workshop on Long Context Foundation Models.

paper bibtex code poster

@inproceedings{2025gong_phantomwiki,
  title = {{PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation}},
  author = {Gong*, Albert and Stankevičiūtė*, Kamilė and Wan*, Chao and Kabra*, Anmol and Thesmar, Raphael and Lee, Johann and Klenke, Julius and Gomes, Carla P. and Weinberger, Kilian Q.},
  year = {2025},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
}

Exponential Family Model-Based Reinforcement Learning via Score Matching

Gene Li, Junbo Li, Anmol Kabra, Nathan Srebro, Zhaoran Wang, and Zhuoran Yang

In NeurIPS (2022). Oral presentation.

paper bibtex code poster

@inproceedings{2022li_score_matching,
  title = {Exponential Family Model-Based Reinforcement Learning via Score Matching},
  author = {Li, Gene and Li, Junbo and Kabra, Anmol and Srebro, Nathan and Wang, Zhaoran and Yang, Zhuoran},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {35},
  pages = {28474--28487},
  year = {2022},
}

Characterizing the Loss Landscape in Non-Negative Matrix Factorization

Johan Bjorck, Anmol Kabra, Kilian Q. Weinberger, and Carla P. Gomes

In AAAI (2021).

paper bibtex

@inproceedings{2021bjorck_nmf,
  title = {Characterizing the Loss Landscape in Non-Negative Matrix Factorization},
  author = {Bjorck, Johan and Kabra, Anmol and Weinberger, Kilian Q. and Gomes, Carla P.},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {35},
  number = {8},
  pages = {6768--6776},
  year = {2021},
  doi = {10.1609/aaai.v35i8.16836},
  url = {https://doi.org/10.1609/aaai.v35i8.16836},
}

also known for

(sorted ascending by number of characters per item)

biking
cooking
running
juggling
being outdoors
playing Table Tennis
following Formula 1 and motorsports
walking fast so that my legs heat up
reading books, newspapers, and research papers