Postdoctoral Researcher Position, Knowledge and Data Engineering Laboratory

About this position

[Background of the recruitment and description of the project]

• Background of the recruitment and description of the project

We are seeking researchers to work on foundational technologies for new language generation AI, such as diffusion language models and flow-based models, within the NEDO project "Project to Develop a Multimodal Foundation Model for AI Robots and Physical AI."

This project is a large-scale initiative aimed at establishing the foundational technologies of next-generation language generation models. The successful candidate will work in the following environment:

• Dedicated access to a substantial allocation of GPUs on ABCI (AI Bridging Cloud Infrastructure, AIST). Training runs using several tens of nodes, each equipped with eight H200 GPUs, are feasible, providing an environment where pre-training, distillation, and evaluation of large-scale models can actually be carried out.

• Research is conducted on the premise of submitting and presenting papers at top international conferences, with support for conference travel expenses.

• Opportunities for joint research with the institutions and companies participating in the project.

In this call, we are recruiting both researchers specializing in model analysis and training algorithms, and researchers specializing in large-scale training/inference systems and performance optimization. Strength in either direction is sufficient to apply.

Please note that this is a fully funded position dedicated to this NEDO project, and the appointee is expected to devote their working time to the project.

https://www.toyota-ti.ac.jp/Lab/kde/en/

[Work content and job description]

Research on foundational technologies for large-scale diffusion language models and flow models (pre-training, distillation, acceleration, and optimization), targeting publications at top international conferences. Candidates can focus on either of the following tracks (handling both is not required). Specific research topics will be decided based on your interests and strengths.

A. Model Analysis & Training Algorithms

Analysis of generation processes/internal mechanisms and model improvements based on these insights.

Application of related techniques (parallel/non-autoregressive decoding, RL/post-training, inference-time scaling, sampling schedules, data synthesis, evaluation/benchmarks, distillation, architecture improvement, etc.).

Bringing expertise from diffusion/flow models in non-language domains (vision, speech, video) into language generation is also welcome (hands-on large-scale training experience is not required; theoretical/analysis-driven approaches are valid).

B. Training/Inference Systems & Performance Optimization

Optimization and implementation of large-scale parallel GPU training (data/tensor/pipeline parallelism, communication optimization) and inference acceleration (KV caching, quantization, custom kernels, pipeline construction). Open-source release of code is actively encouraged alongside paper publishing.

• Assigned department

Existing departments

[Work location]

• Address

468-8511 Aichi Toyota Technological Institute Hisakata 2-12-1, Tempaku-ku, Nagoya

[Number of hired]

Number of hired:4 person(s)

Number of hired description:A few positions (up to 4)

Job content supplemental explanation:Hiring date, Starting date: As soon as possible

[Research environment]

• The project has a dedicated allocation of GPUs on ABCI (AI Bridging Cloud Infrastructure, AIST). Training runs using several tens of nodes, each equipped with eight H200 GPUs, are feasible, providing an environment where pre-training, distillation, and evaluation of large-scale models can actually be carried out.

• Research is conducted as a team of faculty members, doctoral students, and researchers, with opportunities for joint research with participating institutions and companies.

[Publication]

• Research is conducted on the premise of submission to and presentation at top international conferences in NLP, machine learning, and ML systems (ACL/EMNLP/COLM/NeurIPS/ICLR/MLSys, etc.). Travel expenses for conference attendance are supported.

• Publishing results as peer-reviewed papers is the basis of our research activity; in addition, within the constraints of the project, we actively release models and code.

• Because you will be working hands-on with substantial compute, the environment allows you to engage with both model research and systems research.

[Career development]

• We support building connections with collaborators and the international research community during the appointment, with both academic and industry career paths in mind after the term.

[Informal inquiries before applying]

• We welcome informal online meetings at any stage while you are considering applying. We can explain the research topics, research environment, and compensation, so that you can confirm the fit before preparing a full application. Please feel free to contact us with your CV (any format).