📄Curriculum Vitae
Indraneil Paul
PhD Researcher, Language Models & Code · UKP Lab · TU Darmstadt
I am a researcher interested in optimizing the mid-training and post-training of Language Models (LMs), with an emphasis on agentic coding abilities and tool use. My long-term mission is to enhance LMs' long-horizon operation, unlocking their application beyond conventional settings to areas such as computer use and recursive workflows by improving their capabilities to reason, offload computation, and learn from environmental feedback. I also work on preference learning and verifiers, aiming to enhance LMs' capabilities beyond functional axes, such as security and efficiency. My interests span all facets of improving LM training efficacy, including data curation, context length extension, modularity, and reinforcement learning.
Education
- 09/22 - 10/26
ELLIS PhD Candidate in Informatics, TU Darmstadt, Germany
Enrolment - 07/17 - 07/19
Master's by Research in Computer Science, IIIT Hyderabad, India
Certificate - 08/13 - 05/17
Bachelor of Technology in Computer Science, IIIT Hyderabad, India
Certificate
Industry Experience
- 10/25 - 04/26
Applied Scientist PhD Intern, Amazon Inc. (Web Services), Berlin
- Researched RL methods to reduce the cloud tool-calling error rate in the Amazon Q Developer agent
- Explored asynchronous RL approaches for improved distributed training efficiency
- Automated RL environment creation for agents using infrastructure-as-code emulators
- 04/20 - 08/22
Applied Scientist, Amazon Inc. (Advertising), Bangalore
- Created text, image, and multi-modal models for improving EU ad moderation automation by 28%
- Researched multi-modal, multi-lingual, and multi-task pre-training objectives for ad catalog tagging
- Devised sample-efficient training methods for ViT models using self-labeling and sub-task distillation
- 07/19 - 03/20
Software Development Engineer, Amazon Inc. (Logistics), Hyderabad
- Implemented a planner enabling merchants to rank options and schedule last-mile package drop-offs
- Oversaw database tuning, JVM optimizations, and message queue setup for an event ingestion service
Contributor Experience
- 04/24 - 09/24
- 06/23 - 09/24
BigCode Project, ServiceNow and HuggingFace
- Contributed to StarCoder-2 pre-training data collection and training ablations
- Worked on containerization, evaluation framework, and annotation for BigCodeBench
- 05/17 - 07/17
Google Summer of Code, Green Navigation
- Implemented an LSTM forecaster for the EV-Charge-Prediction project to alleviate range anxiety
- Developed an ensemble solution that reduced absolute forecasting error by 39%
- Productionized the Bayesian Optimization service for optimal hyperparameter selection in training jobs
Summer Schools
- 07/23
Lisbon Machine Learning Summer School (LxMLS)
Certificate - 07/21
European Summer School in Logic, Language and Information (ESSLLI)
Certificate
Academic Service
- 07/26
Organizer, SemEval 2026 Task on GenAI Code Detection & Attribution
Event Page - 08/22
Program Committee, KDD 2022 Content Generation for e-Commerce Workshop
Event Page
Selected Publications
-
OctoLong: Mid-Training on Cross-Repository Code Contexts Enhances Long-Context Modeling
-
Themis: Training Robust Multilingual Code Reward Models for Flexible Multi-Criteria Scoring
-
Aletheia: What Makes RLVR for Code Verifiers Tick?
-
AICD Bench: A Challenging Benchmark for AI-Generated Code Detection
-
Droid: A Resource Suite for AI-Generated Code Detection
-
ObscuraCoder: Powering Efficient Code LM Pre-Training via Obfuscation Grounding
-
BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions
-
IRCoder: Intermediate Representations Make Language Models Robust Multilingual Code Generators
ACL 2024 Oral, Bangkok (Outstanding Paper) Slides Abstract PDF -
StarCoder 2 and The Stack v2: The Next Generation
-
Adapters: A Unified Library for Parameter-Efficient and Modular Transfer Learning
-
Sub-Task Imputation via Self-Labelling to Train Image Moderation Models on Sparse Noisy Data