ResearcherSenior Researcher Infinite Context LLMs

Our company

Huawei is a leading global information and communications technology (ICT) solutions provider. Through our constant dedication to customer-centric innovation and strong partnerships we have established leading end-to-end capabilities and strengths across the carrier networks enterprise consumer and cloud computing fields. Our products and solutions have been deployed in over 170 countries serving more than one third of the worlds population.

Our Research Center

With 20 sites across Europe and 1500 researchers Huaweis European Research Institute (ERI) oversees fundamental and applied technology research academic research cooperation projects and strategic technical planning across our network of European R&D facilities. Huaweis ERI includes our Research Center in Zurich Switzerland. A major element of it is a rapidly growing research laboratory focused on fundamental research in the area of computing systems (new hardware new software new algorithms).

The research work of the lab will be carried out not only by Huaweis internal research staff but also by our academic research partners in universities across Europe and beyond. The lab will provide an open research environment where academics will be encouraged to visit and work on fundamental long-term research alongside Huawei staff in an environment that like the best universities and research institutes is open and conducive to such world-leading scientific work.

The research direction

Large language models are increasingly asked to work over far more information than fits in their context window: entire code repositories years of documents the full history of a long-running agent. Todays answer is a patchwork around the model: retrievers chunkers summarizers and prompt templates each a fix for a model that cannot natively hold and organize that much information. We want to replace the patchwork and make non-stop interaction a reality: a context window that is practically unbounded with the model itself organizing what it accumulates in a meaningful way.

Memory-augmented models are one such direction. The model encodes a large corpus once into its own key/value memory and at every step routes its attention to the few pieces of memory that matter so that retrieval happens inside the model in the models own representation space. Recent work has shown that this scales to hundreds of millions of tokens on a single node and can be added to existing models with lightweight adaptation. The next step is to show where it matters: real long-horizon tasks solved by a model that carries its memory with it. Getting there raises open questions on multi-hop reasoning on information that changes over time on the right agent loop around such a model and on how to evaluate it fairly against todays retrieval and agentic pipelines.

For our Computing Systems Laboratory in Zürich we are looking for an outstanding researcher at postdoc or senior level to tackle these challenges. You will work in a small motivated and multicultural team in close collaboration with leading academic groups and with Huaweis product teams and you will take a leading role in one or more of the research threads below. The level of the position will be matched to your experience.

Your responsibilities

Identify the long-horizon tasks where memory inside the model has the clearest advantage over retrieval and agentic harnesses outside it (e.g. repository-scale software engineering auto-research long-running assistants) and build the benchmarks baselines and evaluation to demonstrate it.

Design the agent loop around a memory-augmented model: what the model decides from its own internal state what is left to tools and how new or changed information is written back into memory.

Being an investigator: quickly familiarizing with a emerging large model architectures dissect their components through series of interpretable experiments recognize bottlenecks identify algorithmic opportunities for improvement.

Adapt and train open-weight LLMs with memory-augmented attention (e.g. with LoRA) and study what their routing space has learned.

Extend memory-augmented models with structure: multi-hop reasoning associative and graph-based memory and the notion of time for information that evolves.

Work with the team on scaling inference to billions of tokens across the GPU host memory and storage hierarchy including approximate nearest-neighbor routing inside attention.

Publish at top-tier venues (e.g. ICLR NeurIPS ICML ACL MLSys) file patents and contribute to shared & open-source tooling and benchmarks.

Contribute to shaping the research agenda own work packages end to end and represent the work towards internal product teams.

Your background

PhD in Computer Science Electrical Engineering or a related field or equivalent research experience. For the senior level several years of post-PhD experience in research or industry.

Strong expertise in at least one of the following and working knowledge of the others:

o Large language models: attention mechanisms (e.g. GQA MLA GDN) long-context methods sparse and memory-augmented attention fine-tuning (e.g. LoRA).

o LLM-based agents: tool use context engineering agent harnesses and benchmarks for long-horizon tasks (e.g. SWE-bench-style software engineering deep research).

o Information retrieval: dense retrieval retrieval-augmented generation (RAG) multi-hop question answering graph-based retrieval (HNSW DiskANN).

o Efficient inference: KV-cache management approximate nearest-neighbor search CPU and GPU memory hierarchies prefill and decode phase performance optimization throughput and latency considerations.

Hands-on experience training and evaluating LLMs in PyTorch with a good grasp of the surrounding ecosystem (e.g. Hugging Face Transformers vLLM SGLang) and the engineering discipline to run large experiments reproducibly.

Demonstrated ability to open up a transformer and localize a behavior inside it - modifying attention internals in custom research codebases applying interpretability or model - analysis methods and designing controlled ablation experiments whose small effects are interpreted with appropriate statistical care.

Strong programming skills in Python.

A track record of publications at top-tier venues in machine learning natural language processing or information retrieval.

Excellent written and verbal communication skills with a proven ability to present complex technical information clearly and concisely to a variety of audiences. Collaborative work style with the ability to work in a multicultural environment.

Why join us

Collaborate with world-class scientists and engineers in an open curiosity-driven environment.

Push the research frontier and make a global impact through top-tier publications and open-source software.

Opportunities for professional growth and development.

Competitive salary and a high quality of life in Zurich at the center of Europe.

Last but certainly not least: be part of innovative projects that make a difference.

Applications need to include a CV and a motivation letter both merged in one document. Please indicate a range of possible starting dates in your motivation letter.