华为 IJCAI 2026 论文盘点：从「规模密度」转向「设计密度」

TL;DR - A media roundup of Huawei's four IJCAI 2026 accepted papers, framed as a shift from "scale density" (bigger parameter counts) to "design density" (smarter architecture and training strategies) as compute costs outpace scaling returns. It matters as a signal that a major vendor is prioritizing efficiency-per-FLOP across its research, cloud, data, and translation units.

• Hierarchical ViT to 30B: An "Efficient Hierarchical ViT" architecture plus SMoE variant scales hierarchical ViTs past the ~2B ceiling to 30B total params; trained via MAE self-supervised pretraining on ImageNet-21K then distillation from multiple foundation models on 27M images. EHV-5B-MoE activates only 6.7B params yet hits 89.0% ImageNet-1K linear probe, beating the larger EVA-CLIP-18B, with gains also on video and dense prediction.
• LFS (Learnable Frame Selector): Replaces uniform frame sampling for Video-LLMs with a scoring network trained end-to-end against caption quality — segment-wise selection (top frame per time window) rather than global top-K, backpropagating through a frozen Video-LLM as a plug-in module. Also introduces ICH-CC, a human-written detailed-captioning benchmark from real Chinese intangible-heritage cooking scenes.
• RaMod (Representation-Aware Modularity): Extends ReFT beyond single-task adaptation by intervening on a policy-selected subset of mid-layer hidden representations instead of just prefix/suffix tokens, with an asynchronous scheduler that allocates/releases memory per intervention. Reported vs. base LLMs: 83% less extra prefill time, 79% lower memory; the article's "100% lower generation latency" claim is stated as-is and looks dubious.
• Code-switching speech translation: A per-language MoE speech projector with language-specific and intra-group load-balancing losses, plus four-stage progressive training (ASR pretraining → MoE assembly → ASR→ST transition → CS-ST adaptation) to work around CS data scarcity; beats SeamlessM4T on Fisher and NTUML2021 with up to 39.52 BLEU / 81.33 COMET. The piece notes open questions on scaling this to dozens of languages.