Open Source Project
A fast, scalable, and easy-to-use Python interface to Lean 4. LeanFlow lets you run Lean code, interact with proofs, and evaluate formal statements directly from Python.
Philipp Borchert
Reasoning 路 AI for Math 路 Multilingual NLP
I'm a NLP Researcher fascinated by how we can teach LLMs to reason and tackle complex problems. My research spans from multilingual NLP over information extraction to my current work on reasoning in AI for Math at Huawei in London. I completed my PhD at KU Leuven, where my research focused on multilinguality and NLP for Business applications.
What excites me
Selected Publications & Projects
Preprint
We identify conjecturing as a critical bottleneck in formal mathematical reasoning. We introduce ConjectureBench and LEAN-FIRE, an inference-time method that improves autoformalisation.
Preprint
TopoAlign addresses data scarcity for Math LLMs by aligning code to mathematical structures, improving performance on MiniF2F and ProofNet benchmarks.
ICLR 2026 路 Rio de Janeiro
DRIFT decomposes informal statements, retrieves dependent premises, and selects illustrative examples before formalization, achieving state-of-the-art on ProofNet.
ACL 2025 路 Vienna
FLARE fuses source and target language representations within low-rank adapters, enhancing cross-lingual transfer while maintaining parameter efficiency.
EMNLP (Findings) 2024 路 Miami
MT-LLM integrates machine translation encoders into LLM backbones using self-distillation, unlocking NLU capabilities for over 127 languages.
NAACL 2024 路 Mexico City
MultiRep improves few-shot relation classification by combining multiple sentence representations using contrastive learning.
EMNLP 2023 路 Singapore
CORE is a dataset for few-shot relation classification focused on company relations, challenging models with high contextual complexity.
EMNLP 2023 路 Singapore
This study investigates whether debiasing techniques can be effectively transferred across different languages within multilingual LLMs.
EMNLP 2023 路 Singapore
SEER introduces a novel method for selecting diverse and representative examples for in-context learning in complex QA tasks.