This post shares some of the key architectural principles and lessons we’ve learned.
A standalone MLLM — however capable — is not enough for real-world deployment. The gap between a model that performs well on benchmarks and a system that reliably handles production traffic at scale is enormous.
The shift toward agentic architectures addresses this: instead of a single model call, an agent orchestrates multiple steps — perception, reasoning, tool use, verification — to solve complex tasks robustly.
1. Reliability under distribution shift
Real-world content is adversarial and ever-evolving. An agent that relies on a single monolithic model will degrade unpredictably. We design agents with explicit verification steps: intermediate outputs are checked before influencing downstream decisions.
2. Latency vs. reasoning depth
Deeper reasoning chains produce more accurate decisions but increase latency. We use adaptive depth: simple cases are resolved in one step, complex or ambiguous cases trigger multi-step reasoning. This keeps P99 latency acceptable while maintaining accuracy on hard cases.
3. Fine-grained multimodal understanding
Text-only LLMs cannot handle the rich visual content that characterizes most real-world policy violations. Our MLLMs are specifically fine-tuned on domain-relevant visual-language pairs, with specialized heads for tasks like OCR, object grounding, and visual relation understanding.
4. Alignment with evolving policies
Platform policies change frequently. We build agents with explicit policy-grounding: reasoning is tied to retrievable policy documents, so updating a policy doesn’t require full model retraining.
Several papers from our team formalize parts of this work:
The convergence of stronger foundation models, better alignment techniques (RLHF, GRPO), and scalable agentic architectures is opening up entirely new categories of industrial AI applications. We’re just at the beginning of understanding how to build trustworthy, adaptive, and efficient agentic systems at scale.
If you’re working on similar challenges, feel free to reach out — we’re actively hiring researchers and engineers in this space.
]]>Our CVPR 2021 paper Involution: Inverting the Inherence of Convolution for Visual Recognition proposes a novel atomic operation that inverts these principles.
Involution is spatial-specific and channel-agnostic:
This simple inversion leads to a surprisingly capable and efficient operator. Notably, self-attention — the driving force behind Vision Transformers — can be seen as a special case of involution.
Replacing convolution with involution in standard ResNet-50 backbones:
| Task | Dataset | Improvement |
|---|---|---|
| Classification | ImageNet | +1.6% top-1 accuracy |
| Detection | COCO | +2.5% box AP |
| Segmentation | COCO | +2.4% mask AP |
| Semantic Seg. | Cityscapes | +4.7% mIoU |
With lower FLOPs: 66%, 65%, 72%, and 57% of the convolutional baseline respectively.
Involution provides a unified framework connecting convolution, self-attention, and dynamic filtering — clarifying the design space of neural network operations for vision. It shows that the right inductive bias can yield both better accuracy and better efficiency simultaneously.
Code and pretrained models: github.com/d-li14/involution
arXiv: arxiv.org/abs/2103.06255
📢 [2020-12] Our workshop proposal “Continual Learnig in Computer Vision” has been accepted by CVPR 2021, which is selected out of 109 proposals.
📢 [2020-12] Reviewer service: ICML 2021; NeurIPS 2020 (Top 10% high-scoring reviewer).
📄 [2020-11] Our paper “Generative Adversarial Networks in Computer Vision: A Survey and Taxonomy” has been accepted by ACM Computing Surveys. The journal is ranked 3/221 in Computer Science Theory & Methods (Impact Factor: 7.99):
📢 [2020-08] I join the Bytedance AI Lab as a research scientist this month.
📢 [2020-06] We have successfully organized the 1st “Continual Learning in Computer Vision” workshop at CVPR 2020.
📄 [2020-06] Our paper “IROS 2019 Lifelong Robotic Vision: Object Recognition Challenge [Competitions]” has been published at IEEE Robotics & Automation Magazine.
📄 [2020-05] Our paper “Synthetic-Neuroscore: Using a neuro-AI interface for evaluating generative adversarial networks” has been accepted by Neurocomputing.
📄 [2020-01] 2 papers have been accepted by ICRA 2020: “OpenLORIS-Object: A Robotic Vision Dataset and Benchmark for Lifelong Deep Learning”, and “Are we ready for service robots? The OpenLORIS-scene datasets for lifelong SLAM”.
]]>