Indraneil Paul
PhD Researcher, Language Models & Code
Berlin, Germany
I'm a PhD researcher at the UKP Lab, TU Darmstadt, advised by Iryna Gurevych and Goran Glavaš. I work on the mid- and post-training of language models, with an emphasis on reasoning, agentic coding, and tool use. Lately, I have been exploring the role of mid-training in instilling deeper alignment and world modeling capabilities in LMs.
My longer-term aim is to extend LMs' capabilities in long-horizon operation by improving how they reason, offload computation, and learn from environment or agent feedback. To this end, I also study scalable supervision via verifiers that improve models along hard-to-verify axes like security and efficiency.
Previously I was an Applied Scientist at Amazon, and before that a dual-degree student at IIIT Hyderabad. I've contributed to several open LM training and evaluation releases, including StarCoder2 and BigCodeBench.
🔬Research Interests
-
Code LMs & tool use
Developing and evaluating capable code models and extending them for long-horizon operation — tool use and learning from environment feedback. This spans the mid-training that stretches models beyond repository-scale context, and the pre-training corpora and benchmarks that ground everyday tool use.
-
Verifiers & scalable supervision
Building the scalable supervision that post-training leans on — pinning down what actually makes RLVR and code verifiers effective, and training reward models that score generations along hard-to-verify axes like security and efficiency, across languages and criteria.
-
Pre-training efficiency & grounding
The pre-training foundation the rest builds on — getting more out of code-LM training by grounding models in code obfuscation and compiler intermediate representations, strengthening multilingual transfer, and keeping adaptation modular and parameter-efficient.
📚Publications
- 2027
- 2026
- 2026
- 2026
- 2025
- 2025
- 2025
- 2024
- 2024
- 2023
- 2022
No publications in this topic.
📰News And Updates
-
Aletheia, on what makes RLVR for code verifiers tick, accepted at TMLR.
-
Co-organizing the SemEval 2026 Task on GenAI Code Detection & Attribution.
-
AICD Bench presented at EACL 2026 (Rabat).
-
Droid, a resource suite for AI-generated code detection, presented at EMNLP 2025 (Suzhou).
-
Started an Applied Scientist PhD internship at Amazon (AWS) in Berlin, working on RL for cloud tool-calling in Amazon Q Developer.
-
BigCodeBench (Oral) and ObscuraCoder (Poster) presented at ICLR 2025 (Singapore).
-
Invited talk — Challenges in Code LMs at IIIT Hyderabad.
-
Invited talk — Code Generation: Challenges and Solutions at BHT Berlin.
-
IRCoder received an Outstanding Paper Award at ACL 2024 (Bangkok).
-
StarCoder 2 and The Stack v2 released.
✉️Contact
I'm always glad to talk about code models, verifiers, and long-horizon agents — or to hear about roles and collaborations. Reach me by email, or find me on the profiles linked at the top of the page.