AI Researcher

About VinSmart Future
VinSmart Future (VSF) is the technology company of Vingroup, formed by merging the group's technology ecosystem - VinApp, VinIT, VinBigdata, and other tech units. AI is our foundation. With nearly 4,000 local and international technology experts, we build technologies that connect data, models, and infrastructure at national scale.
AI Researcher
Work Location: Vincom Dong Khoi, District 1, HCMC
Job overview
We are hiring AI Researchers to build large language models that compete with the best in the world. That is the explicit goal, and we will state the obvious: we are not there yet. Vietnamese is where we start and where we intend to be uncontested - but the bar we measure ourselves against is global, not regional.
This is a research position, not an applied-engineering one. You will own open problems, run the experiments that settle them, publish the results, and work with our engineers to fold what wins back into the model.
We optimize for research depth first. A strong publication record at A* venues is the primary signal we look for. Engineering ability is a hard floor, not the differentiator.
Key Responsibilities:
1. Own a research agenda (primary)
Identify and attack open problems in LLMs, from first principles rather than from the last paper you read.
Design and run rigorous experiments: clear hypotheses, controlled ablations, honest baselines, error bars. Scaling-law-informed small-scale studies before you spend a large training run.
Publish at top venues (NeurIPS, ICML, ICLR, ACL, EMNLP, NAACL, AAAI, CVPR, ECCV/ICCV, etc.).
Release models, code, and benchmarks where it makes sense. We want VSF's work to be reproducible and used.
2. Research areas
Post-training and alignment. SFT, instruction tuning, RLHF, DPO, GRPO and successors; reward modeling; preference data design; the failure modes nobody advertises - alignment tax, the see-saw effect across tasks, reward hacking, mode collapse.
Reasoning and agents. RL from verifiable rewards; long chain-of-thought and test-time compute; self-verification and process supervision; tool use, multi-step planning, and long-horizon agentic behaviour.
Long context, memory, and retrieval. Context extension and its failure curve; efficient attention; retrieval-augmented and memory-augmented architectures; what the model should keep in weights versus look up.
Efficiency and inference. Distillation, quantization, sparsity, speculative decoding, KV-cache design - treated as research problems in the accuracy/latency/cost trade-off, not as an engineering afterthought.
Multilingual and Vietnamese. Cross-lingual transfer, low-resource adaptation, and the specific ways English-first design choices - tokenization, diacritics, compounding - break on Vietnamese. Understanding why they break is part of the work.
3. Evaluation as research
Build the evaluation suites the field lacks, especially for Vietnamese and multilingual generation, reasoning, and long-context behaviour.
Design automated evaluation that catches regressions and cross-task interference early in a training run, not after it.
Be skeptical of your own numbers. Contamination, leakage, and benchmark overfitting are real, and we would rather find them ourselves than have a reviewer or a user find them for us.
4. Transfer to product
Work with research engineers to move validated methods into the production model and serving stack (vLLM, TGI, Triton).
You are not expected to own the serving stack or an on-call rotation. You are expected to know enough about latency, throughput, and cost that your methods can survive contact with production.
Mentor junior researchers and interns.
Minimum Requirements
Research (primary)
PhD in CS / AI / ML, or a Master's with an equivalent research record.
First-author publications at A*-ranked venues.
Demonstrated ability to take a problem from question to result to paper independently.
Real depth in at least one of the research areas above. Breadth across several is a plus, but we would rather have someone who is genuinely deep in one thing.
Engineering (required floor)
Strong Python and deep PyTorch - you write your own training and evaluation code.
Solid grasp of transformer internals, tokenization, optimization, and modern training recipes.
Hands-on experience with distributed training (FSDP, DeepSpeed, Megatron, or equivalent) and with fine-tuning at scale (LoRA, QLoRA, full-parameter).
Some experience turning research code into something other people can run - an open-source release, an internal service, or a clean handoff to an engineering team. We do not require years of industry engineering, but we do require that your work has left your laptop.
Preferred Qualifications
Sustained record at A* venues rather than a single paper, or a first-author paper with clear influence on later work.
Direct experience on a frontier or near-frontier model effort - pre-training, post-training, or evaluation.
Area chair, reviewer, or workshop organizer at a major venue.
Widely used open-source models, datasets, or libraries.
Experience with large multi-node training runs and the debugging that comes with them.
Research on low-resource or multilingual settings.
Why You'll Love Working Here
Flexible working hours and attendance policy (Work from Home on working Saturdays).
Attractive compensation and bonus packages, highly competitive in the market.
Exclusive employee benefits across the Group's ecosystem in accordance with company policies.
Opportunity to work on large-scale and strategic technology projects.
Professional technology environment with leading scientists, experts, and engineers from top technology companies in Vietnam and around the world.
Full statutory insurance coverage in accordance with Vietnamese Labor Law (Social Insurance, Health Insurance, Unemployment Insurance), along with private healthcare insurance based on job grade and annual health check-ups at reputable hospitals and healthcare centers nationwide.
Participation in internal activities, team-building programs, and annual company events.