Senior Agentic Reasoning Researcher

Huawei Finland R&D; Helsinki; On-site; EUR 90,000 - 130,000; Full time

Job summary

Huawei Finland R&D is seeking a passionate researcher to advance AI safety and alignment within Harmony OS and related ecosystems. You will work in a Helsinki cloud content security lab, tackling multimodal data challenges and building scalable, interpretable AI systems.

The role emphasizes cutting-edge research with potential production impact, collaborating with international teams to protect mobile AI users and deliver robust detection capabilities.

Qualifications

• PhD in Computer Science, Deep Learning, ML, Mathematics or related field.
• Strong track record in Explainable AI and Safety domain.
• Experience with LLMs/VLMs and neural-symbolic AI.
• Publications in top conferences encouraged.

Responsibilities

• Design architectures for AI detection systems over multimodal data.
• Explore visual QA, visual reasoning, and multimodal grounding.
• Develop interpretable and aligned AI models with human values.
• Implement alignment techniques for LLMs and agentic workflows.
• Reverse-engineer neural networks to understand internal workings.
• Develop data attribution methods to quantify influence of data.
• Build benchmarks/datasets to measure reasoning and reliability.
• Defend against jailbreaking, prompt injections, and adversarial exploits.
• Communicate insights to stakeholders.
• Turn research into production-ready tools and frameworks.

Skills

LLMs & VLMs

Neural networks

Explainable AI

Python/Java/C++

Graph reasoning

Education

PhD in CS/ML

Tools

PyTorch

TensorFlow

Job description

Huawei Consumer Business Group is the leader in the all-scenario AI life. It covers smartphones, PC and tablets, wearables, mobile broadband devices, family devices and device cloud services. Huawei Consumer Business Group is dedicated to delivering the latest technologies to consumers. We believe in creating safe, privacy-friendly, and high-quality products with great user experiences for our consumers.

The mission of Cloud Content Security Lab at our Helsinki Research Center is to protect the online safety of Huawei mobile users within the Harmony OS ecosystem through technological advancements. We build advance AI detection solutions to ensure content safety, safety in LLMs & agentic AI such as alignment and machine unlearning. We are looking for a passionate and motivated researcher with a solid track record of solving challenging problems, advancing state-of-the-art, and demonstrated passion for ground-breaking research that could be scaled to production environment with an emphasis on the intersection of AI and Safety.

Your responsibilities:

• Design and implement effective, efficient and novel architecture to enable AI detection system that understands and able to reason over complex multimodal data, such as images, videos, text and cultural knowledge.

• Explore novel techniques for visual question answering, visual reasoning, text-based reasoning, and multimodal knowledge grounding. Explore the use of knowledge graphs and other neural-symbolic techniques to enhance reasoning capabilities.

• Create new techniques to make AI models/Agentic AI system more interpretable, unbiased, transparent and aligned with human values for real-world scenarios.

• Create new techniques to ensure consistent alignment between user-centric explanations and intrinsic model behavior. Develop and implement state-of-the-art alignment techniques (e.g., RLHF, RLAIF, Constitutional AI) specifically tailored for LLMs CoT, agentic workflows and multi-step reasoning.

• Reverse-engineer on neural networks to understand internal workings of AI models rather than treating them as "black boxes". Implement and integrate AI explainbility techniques and tools.

• Develop data attribution methods to quantify the influence of specific data points (text and image) on a model's prediction and implement solution.

• Build rigorous benchmarks and datasets to measure the quality of reasoning capabilities and reliability, quality of user-centric explanations, model transparency, and effectiveness of treatments.

• Develop robust defenses against jailbreaking, prompt injections, and adversarial exploits that target a model’s planning and tool-use capabilities.

• Communicate insights to stakeholders.

• Turn cutting-edge AI detections and AI safety papers into high-performance, scalable code, transforming theoretical breakthroughs into production-ready tools and frameworks.

Requirements:

• Ph.D. in Computer Science, Deep Learning, Machine Learning, Mathematics or other related fields.

• Focused on research in the AI field with good track records and high motivation in Explainable AI and Safety domain.

• Strong proficiency in Large Language Models (LLMs)/VLM, neural networks, neuro-symbolic AI, graph reasoning architectures.

• Strong background in Python, Java, or C++, with deep knowledge of ML frameworks such as PyTorch, TensorFlow.

• Successful experience in AI alignment to human values and expectations, model robustness improvement, controlled and continual learning, neural network interpretability and editing techniques is highly valued.

• Pioneering novel methods and neural networks that revolutionized machine learning or the AI field, or revolutionized the industry, is a big bonus.

• Strong publication record in top conferences (i.e. NeurIPS/ICLR/ICML/ AAAI/ACL/CVPR/ICCV/EMNLP/NAACL).

• Candidates with ICPC, IOI/IMO, IOAI and other international competition in computer science, machine learning, and AI, etc. are highly preferred.

• Good teamwork, enjoy working with multi‑culture teams, passionate in challenging the status quo.

This is a full‑time consultant position based in Ruoholahti, Helsinki, where you will be employed by a Huawei external supplier.

Why join us?

Join a high‑caliber lab: Work in our Cloud Content Security Lab in Ruoholahti, Helsinki, alongside experienced international researchers and engineers in a focused environment that values scientific excellence, depth, and real impact to protect millions of Huawei Mobile users.