Machine Learning Researcher: Agentic Post-Training for On-Device Models

Apple, Inc.; United States; On-site; USD 180,000 - 260,000; Full time

Job summary

Apple Inc. is seeking a researcher to advance agentic post-training for on-device models.

You will develop supervised fine-tuning, reinforcement learning, and distillation methods to improve reasoning and multi-step tool use, while shaping training data and interactive environments for scalable evaluation. You will work across Apple’s hardware, software and ML teams to balance latency, memory and energy constraints with task success, publishing high-impact results and driving practical

Qualifications

• PhD or equivalent practical experience in a related field.
• Strong research track record with first-author publications or novel contributions.
• Proficient in Python and PyTorch for research-to-production workflows.
• Ability to formulate hypotheses, build training + evaluation pipelines, and run controlled experiments.

Responsibilities

• Develop post-training methods to convert foundation models into on-device agents.
• Explore supervised fine-tuning, reinforcement learning, preference optimization, and distillation.
• Design data, environments, reward signals, and evaluations for multi-step tool-use tasks.
• Balance performance with latency, memory, energy, and tool-use costs on-device.
• Collaborate across research and engineering teams and translate findings into Apple products.

Skills

Python

PyTorch

Research leadership

On-device ML

Agentic systems

Education

PhD in CS/ML/AI/CV/NLP or equivalent

Job description

The Video Computer Vision (VCV) organization is an applied research and engineering team developing next-generation multimodal intelligence technologies across Apple products. Within VCV, our team builds intelligent systems that combine multimodal understanding, reasoning, and tool use to deliver compelling user experiences. Our work is deeply integrated into the Apple ecosystem, partnering across hardware, software, and ML teams to deliver real-time, scalable, and privacy-preserving experiences reaching millions of users.

In this role, you will develop post-training methods that turn foundation models into capable, reliable, and efficient agents for on-device execution. You will develop supervised fine-tuning and reinforcement learning methods for agentic models, alongside the training data, interactive environments, reward signals, and evaluations needed to advance reasoning, tool use, and long-horizon tasks. A central research challenge is enabling smaller models to make effective decisions and complete multi-step tasks within tight latency, memory, and energy budgets. You will investigate how to improve reasoning, planning, and tool use while reducing reliance on lengthy generated reasoning and large inference budgets. Research directions may include structured reasoning, adaptive computation, and training objectives that balance task success with computational cost. You will own the projects end to end: identify capability gaps, formulate hypotheses, build training pipelines, conduct rigorous experiments, and translate frontier research into Apple products that deliver meaningful value to users.

Advance agentic post-training for on-device models. Develop supervised fine-tuning, reinforcement learning, preference optimization, and distillation methods that improve reasoning and multi-step tool use in on-device LLMs.

Design training data and interactive environments. Build scalable pipelines for trajectory generation, curriculum construction, and hard-example mining. Develop environments that support learning through tool use and multimodal interaction.

Design reward functions, verifiers, and graders for complex agent tasks, addressing long-horizon credit assignment, supervision quality, and gaps between reward scores and actual task success.

Improve capability under deployment constraints. Research methods that balance task success with latency, memory, energy, and tool-use costs, including structured or compressed reasoning and adaptive computation.

Build rigorous evaluations and diagnostics. Measure task success, generalization, reliability, and efficiency. Analyze model outputs and agent trajectories to distinguish modeling limitations from data, reward, environment, and infrastructure issues, and turn findings into targeted improvements.

Translate research into Apple experiences. Collaborate with software and hardware teams to integrate and validate agent capabilities under realistic on-device constraints.

Drive external research impact. Contribute to research projects from initial hypothesis through experimental validation, publish high-impact technical reports as a core author, and release open-source models that advance agentic post-training and support the broader research community. Example technical reports: "Internalized Visual Thinking for Video Reasoning, Structured Reasoning for Efficient Agentic Search."

MS in Computer Science, Machine Learning, Artificial Intelligence, Computer Vision, Natural Language Processing, or a related field, or equivalent practical experience.

A strong research track record demonstrated by either first-author publications at top-tier conferences (such as NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, ACL, or EMNLP) or substantial hands-on experience training frontier models, with contributions relevant to foundation models, reinforcement learning, multimodal learning, or agentic systems.

Strong programming skills in Python and proficiency with deep learning frameworks such as PyTorch.

Demonstrated ability to formulate research hypotheses, implement training and evaluation pipelines, conduct controlled experiments, and draw well-supported conclusions.

PhD in Computer Science, Machine Learning, Artificial Intelligence, Computer Vision, Natural Language Processing, or a related field, or equivalent practical experience. Research focus on foundation models, reinforcement learning, agentic systems, multimodal reasoning, or efficient machine learning.

Significant contributions to high-impact technical reports or open-source model releases in agentic post-training, with substantial ownership of research methodology, model training, experimental validation, and technical communication.

Experience developing or training agents for tool use and long-horizon tasks.

Experience designing interactive training environments, trajectory-generation pipelines, reward models, or scalable agent evaluations.

Experience improving smaller models through compute-aware post-training, with an understanding of on-device latency, memory, and energy constraints.

Ability to independently drive research projects and collaborate effectively across research and engineering teams.