Event Focus

AICon Global Artificial Intelligence Development and Application Conference will be held in Shenzhen on August 21–22. The program has been fully released, with sessions covering Agent engineering, large-model infrastructure, AI-native development, embodied intelligence, and related engineering practices.

One highlighted talk will be delivered by Bohan Zhuang, a ZJU Hundred Talents Program researcher and doctoral supervisor at Zhejiang University. His session, titled “Efficient Long-Context Modeling for Multimodal Reasoning,” will appear under the track “Large Model Efficiency Engineering and Agent System Practice.”

Why Efficiency Matters

The talk centers on a practical shift in AI development: model capability alone is no longer enough. As multimodal models move toward longer context windows, video understanding, generation, and Agent-style execution, inference cost, memory use, and latency become major deployment constraints.

Zhuang’s presentation will discuss algorithm–system co-design, meaning that model algorithms and inference infrastructure are optimized together. The agenda includes three efficiency directions:

• Efficient attention, including sparse and linear attention, to address the quadratic cost growth of standard attention over long sequences;
• Efficient memory, including KV-cache compression and cache management, to reduce GPU memory pressure during long-context inference;
• Efficient decoding, including parallel decoding strategies, to improve throughput and reduce generation latency.

KV-cache refers to stored key-value tensors reused during inference, while prefill is the initial stage where a model processes the prompt and builds that cache. Both become expensive when context grows.

Multimodal Agents and World Models

The session will connect these techniques to core multimodal Agent capabilities: understanding and generation. On the understanding side, the focus includes long-video comprehension, spatial reasoning, and decision-making under limited compute. On the generation side, the emphasis is on lower-cost world model generation, including 3D world reconstruction and long-video generation.

The article lists several research examples from Zhuang’s team:

• FPSAttention improves three-dimensional attention for video DiT. On NVIDIA H20 running Wan2.1-14B for 720p video generation, it achieved up to 7.09× attention-operator acceleration and 4.96× end-to-end video generation speedup while largely preserving quality;
• FlashBlock explores caching in the block diffusion paradigm, reporting up to 1.44× token throughput improvement and 1.6× attention-time reduction in long-text and video generation experiments;
• Mirage stores spatial memory for video models in latent space, reducing three-dimensional cache memory use by up to 55×, achieving up to 10× end-to-end acceleration, and obtaining the best result on WorldScore.

These system and algorithm efforts are being integrated into Inferix, a unified framework described as a Block Diffusion inference engine for World Models, supporting KV-cache management, streaming video generation, and interactive world rollout.

Agent Loops and Long Context

The article also highlights efficiency challenges in Agent loops, where a system repeatedly observes, reasons, acts, and updates context. In frameworks such as Claude Code, Codex, and OpenClaw, context length can accumulate over time, increasing both first-round prefill time and KV-cache memory consumption.

To address this, the team proposed TriAttention for KV-cache compression in long reasoning. In a 32K-token generation experiment on AIME25, it achieved 2.5× throughput improvement or 10.7× KV-cache memory compression while maintaining reasoning accuracy comparable to full attention.

The team also explored collaboration between large and small models. R-Stitch switches models based on entropy and achieved 3–4× acceleration across reasoning tasks with different model sizes while keeping accuracy close to full large-model decoding. Agent-as-a-Router turns model selection itself into an Agent Loop; in tests on 2,900 coding tasks, the router-based framework achieved a higher average task score than opus while costing less than half as much.

Outlook

AICon Shenzhen reflects a broader industry trend: the next phase of AI competition is increasingly about reliable systems, not only stronger base models. Multimodal reasoning, long-context Agents, video generation, and embodied AI all require faster inference, better cache management, and lower latency.

At the same time, the article notes an important trade-off. Sparse or linear attention and KV-cache compression can significantly reduce cost and improve speed, but they may introduce controllable, non-zero quality loss. The best configuration can also vary by task, modality, and sequence length. The likely direction is therefore not a single universal acceleration trick, but diagnostic and tunable efficiency frameworks that can be adapted to real Agent and world-model workloads.