Research Assistant — Data Science Lab (dlab), EPFL
Sep 2024 – Sep 2025 · dlab.epfl.ch · Prof. Robert West
🧠 The Question
Both projects below come from the same place: language models are given far more context than they can use well, and context is the thing you pay for. One project attacks that from the demonstration side, the other from the compression side.
📄 GRAD — generating demonstrations instead of retrieving them
First author (equal contribution) — Findings of EMNLP 2025, presented in Suzhou.
Few-shot prompting usually means retrieving examples from a database you have to build, index and maintain. In GRAD we trained an LLM with GRPO to write an example for each specific question instead, which removes the database entirely.
What I ran:
- Reward design — a multi-objective, truncation-aware reward combining answer log-probability, accuracy and a demonstration count, with an explicit cap so the model cannot reward-hack by padding.
- PPO vs GRPO — compared both; kept GRPO for training stability.
- LoRA to fit training inside the memory available.
- GRADi, an SFT warm-started variant of the same recipe.
Result: beat RAG baselines on GSM8K, MATH and ARC across Qwen2.5-3B/7B/14B and Llama-3.2-3B — with shorter demonstrations and shorter answers. Trained only on maths, it generalised out of distribution to physics, chemistry and computer science. Code and models are public.
🗜️ LLMs as compressors — the benchmark programme
I designed the benchmark programme measuring LLMs as prompt compressors against LLMLingua-2, across mathematical reasoning, summarisation and multiple-choice QA at matched token budgets.
That programme became the evaluation backbone of Cmprsr (arXiv:2511.12281, under review at ACL 2026) — a question-agnostic token-level compressor, meaning one compressed prompt can serve whatever gets asked of it afterwards — and, a year later, the research basis for a production compression API at Compresr.
🛠️ Stack
PyTorch • HuggingFace Transformers • TRL • PEFT / LoRA • vLLM • Weights & Biases • SLURM / Run:ai