Applied Researcher: On-Device Multimodal Reasoning

Summary

The Video Computer Vision (VCV) organization is an applied research and engineering team developing real-time, on-device Computer Vision and Machine Perception technologies across Apple products. Within VCV, our team builds next-generation multimodal AI systems that combine on-device multimodal encoders, large language models, and foundation models to create intelligent systems capable of understanding, reasoning, and acting across language, vision, audio, and tools. 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.

Description

We are seeking an Applied Researcher with deep expertise in multimodal reasoning at small model scale — making vision-language models in the smaller regime (under ~10B parameters, down to sub-1B) think, plan, and act reliably under strict compute, memory, and latency constraints. In this role you will own the reasoning side of the on-device multimodal stack: designing compact VLMs that reason over images, video, and 3D scene content; compressing and distilling the reasoning process itself; and engineering the decoding and inference path that makes multi-step reasoning affordable on an Apple device. This role offers the unique opportunity to define what on-device intelligence looks like for hundreds of millions of users. You'll push the boundaries of what small models can achieve — enabling real-time multimodal understanding and multi-step reasoning without reliance on cloud connectivity. You'll collaborate with hardware teams, compiler engineers, and ML researchers to unlock capabilities that few organizations can deliver at Apple's scale and quality bar.
This role spans multiple dimensions of efficient on-device reasoning — including VLM architecture and connector design, reasoning post-training (SFT/RL), chain-of-thought compression, speculative and structured decoding, visual token reduction, quantization and distillation, and hardware-aware inference optimization.

A core focus of this role is efficient reasoning: compressed and latent chain-of-thought, reasoning distillation from frontier teachers, adaptive test-time compute (knowing when — and how long — to think), speculative and structured decoding, KV-cache compression, and visual-token efficiency. A second focus is reasoning over real-time visual perception experts. Rather than consuming pixels alone, the VLM should be able to invoke and reason over the outputs of specialist on-device vision models — feed-forward 3D scene and geometry estimators (VGGT-style reconstruction, depth, camera pose), human body and hand mesh/pose recovery, object detectors, localizers, and trackers — and fuse those structured, metric outputs into its reasoning about the scene. This raises real research questions: how to represent geometry, body parameters, and detections compactly in a token-budgeted context; how to schedule which experts run at which frame rate within a real-time budget; and how to train a small model to invoke, trust, and cross-check them. Efficiency is treated as a first-class metric here: reasoning quality is measured at a fixed latency, memory, and power budget.

Responsibilities

• Design, train, and post-train compact vision-language models (under ~10B, including sub-1B) that perform multi-step visual reasoning, grounded visual understanding, and language generation within on-device resource budgets
• Research and implement efficient reasoning techniques — compressed and latent chain-of-thought, reasoning-trace distillation, early-exit and budget-aware reasoning, adaptive compute allocation, and test-time scaling that maximizes reasoning quality per FLOP
• Own the decoding stack for on-device inference: speculative and self-speculative decoding, draft models and multi-token prediction, structured/constrained generation, KV-cache compression and quantization, prefill/decode scheduling, and streaming latency (TTFT, tokens/sec)
• Build reasoning over real-time perception experts: enable a small VLM to invoke and reason over on-device 3D scene reconstruction and geometry (VGGT-style feed-forward reconstruction, depth, camera pose), human body/hand pose and mesh recovery, object detection, localization, and tracking — designing the representations, interfaces, and training signals that make those outputs usable inside a limited context
• Develop structured-output fusion and expert scheduling: compact tokenizations for geometry, body parameters, and detections; policies for which perception models run at which resolution and frame rate; and mechanisms for the reasoner to resolve conflicts between experts and its own visual features
• Apply reasoning-focused post-training: supervised distillation from frontier teachers, preference and RL methods (GRPO, RLVR, STaR, rejection sampling), verifier- and reward-guided decoding, and process supervision for multimodal and spatially grounded chains
• Drive visual token efficiency and feature compression: token pruning, merging, and resampling, adaptive resolution and frame-rate policies, and learned connectors that preserve reasoning accuracy at a fraction of the visual token budget
• Develop distillation, pruning, and quantization strategies that preserve multimodal reasoning fidelity at reduced model sizes — including QAT and mixed-precision inference across the language decoder, vision encoder, and perception experts
• Explore hybrid architectures combining stateful components (SSM/Mamba, linear attention) with attention-based components to balance long-context visual reasoning — long video, multi-image, extended dialogue — against inference efficiency and bounded memory
• Optimize reasoning and encoder architectures for Apple silicon, including Neural Engine, GPU, and ANE-aware design patterns; profile and iterate on latency, memory footprint, thermal behavior, and power consumption across Apple's device portfolio
• Partner with our visual representation and video-encoder efforts (self-supervised and joint-embedding pretraining, streaming/stateful encoders, world models) to make representations reasoning-ready, and co-design the encoder-LLM interface
• Build evaluation frameworks that measure reasoning fidelity, visual and spatial grounding, hallucination, and robustness under device budgets, including accuracy-vs-latency and accuracy-vs-power trade-off curves
• Collaborate with hardware, compiler, and platform teams to co-design model architectures that exploit device-specific acceleration capabilities

Minimum Qualifications

• MS in Computer Science, Machine Learning, AI, Computer Vision, or a related field (or equivalent practical experience)
• Strong foundation in deep learning, with specific experience in LLM or VLM training, post-training, or inference optimization
• Demonstrated experience working with multimodal models (vision-language models, multimodal LLMs) in resource-constrained environments, including hands-on work with reasoning quality, decoding, or model compression
• Proficiency in Python and modern deep learning frameworks (PyTorch preferred), with familiarity with inference and optimization toolchains (quantization, distillation, pruning — e.g., vLLM/SGLang, llama.cpp, MLX, CoreML)

Preferred Qualifications

• PhD with research in efficient multimodal reasoning, LLM reasoning, model compression, efficient inference/decoding, or lightweight VLM architectures
• Experience training or post-training vision-language models end-to-end — connector/projector design, visual instruction tuning, resolution and token-budget trade-offs, small-model recipes
• Hands-on experience with reasoning techniques: chain-of-thought distillation and compression, latent/implicit reasoning, reward-guided decoding, RL for reasoning (GRPO, RLVR, STaR), or test-time compute allocation
• Expertise in decoding and serving optimizations: speculative decoding, structured/grammar-constrained generation, KV-cache quantization and eviction, continuous batching, long-context inference
• Experience combining LLMs with real-time perception models — 3D reconstruction and geometry (VGGT, DUSt3R/MASt3R-style, SLAM, monocular depth), human pose and body/hand mesh recovery (SMPL-family), detection, segmentation, or tracking — and with spatial or 3D-grounded reasoning and embodied/spatial VQA
• Experience deploying LLM or multimodal models on mobile or edge hardware (CoreML, MLX, TensorRT-LLM, or equivalent), with attention to ANE/GPU kernel and memory constraints
• Experience with quantization-aware training, mixed-precision inference, and knowledge distillation for vision-language models
• Familiarity with efficient vision encoders and self-supervised/joint-embedding pretraining (V-JEPA, I-JEPA, MAE, DINO/DINOv2, SigLIP, CLIP), Mamba/SSM vision backbones, or streaming architectures for real-time video with fixed memory budgets
• Interest in Video-LLMs, long-video reasoning, and world models for prediction and planning
• Publication record in top-tier venues is a plus (NeurIPS, ICML, ICLR, CVPR, ECCV, ACL, MLSys, ICRA, etc.)