IJCAI-ECAI 2026 runs from 15 to 21 August in Bremen, Germany. As the comprehensive top conference in AI, IJCAI has long been the exam hall where the industry’s big labs show what they really achieved in the past year. Who made real progress on which front, which technical routes are consolidating into consensus, the papers are the most direct answer.
As the conference opens, a clear trend is visible. AI compute cost stays high, and the marginal return of scaling parameters no longer keeps pace with the marginal cost. The economic reality is reshaping the logic of innovation, from “can we make it bigger” to “can we use it more efficiently.”
Huawei’s four accepted papers at IJCAI 2026 offer a technical path for that turn: use smarter architecture design and training strategy to offset data scarcity and extract more intelligence per unit of compute.
The first paper attacks a hidden ceiling in the scale race for vision foundation models. Plain ViTs scaled from 6B to 22B, but hierarchical ViTs stayed stuck below 2B, because they demand far more data and training craft. Huawei’s Distilling and Scaling Hierarchical Vision Transformer to 30B Parameters pushes the ceiling from 2B to 30B. Its Efficient Hierarchical ViT architecture keeps multi-scale extraction while staying compute-efficient. The team trained dense models from 200M to 5B parameters and added a sparse mixture-of-experts variant to reach 30B total, the largest public hierarchical ViT to date. A two-stage recipe, MAE self-supervised pretraining on ImageNet-21K then distillation from leading foundation models on a 27-million-image set, delivered the proof: the 30B MoE model activates only 6.7B parameters at inference yet hits 89.0 per cent accuracy on ImageNet-1K, beating the larger EVA-CLIP-18B, and holds up on video and dense-prediction tasks.
The second paper targets where Video-LLMs burn compute: frame encoding. Uniform sampling wastes it, missing a one-second key event while re-encoding long static stretches; query-driven selection fails at detailed captioning, which has no query. Huawei’s AI data team proposed the Learnable Frame Selector (LFS), which learns which frames let the model write a better caption. It scores each frame for event importance, selects per segment rather than globally, and is trained end to end: selected frames go to a frozen Video-LLM, the caption is compared with a human one, and the loss updates only the selector. The LLM never moves; LFS is a plug-in module. Across models and benchmarks it lifted results uniformly, and the team built a new benchmark, ICH-CC, from real Chinese intangible-heritage cooking scenes.
The third paper tackles cross-task generalisation in large language models. Per-token dynamic routing works but is slow and memory-heavy; representation fine-tuning is efficient but imprecise. Huawei Cloud, with several universities, proposed RaMod (Representation-Aware Modularity). It edits a strategy-filtered subset of middle-layer hidden representations for precise, self-guided adaptation, and an asynchronous scheduler allocates memory only when needed. Versus the original LLM, RaMod cut extra prefill time by 83 per cent, roughly halved generation latency, and reduced memory footprint by 79 per cent, while improving cross-task generalisation. The economics of deployment shift: one base model plus a few light intervention modules can serve very different tasks without a full copy per scenario.
The fourth paper faces the hardest case: not enough training data. Code-switching speech translation mixes languages and suffers acute data scarcity. Huawei’s Translation Service Center, with Xiamen University and the University of Macau, proposed aligning semantic representations per language, building a dedicated “expert team” for each. A mixture-of-experts speech projector gives each language a dedicated set of experts; a routing mechanism sends speech features to the right expert. Two auxiliary losses keep routing honest: a language-specific loss and an intra-group load-balancing loss. Training runs in stages, from ASR pretraining per language, to assembling the MoE, to a smooth transition into single-language speech translation, then into code-switching data. Against Whisper, SeamlessM4T and LLaST, it beat the strongest baseline, SeamlessM4T, on four Fisher and NTUML2021 test sets, reaching a BLEU of 39.52 and a COMET of 81.33. The signal is clear: in data-scarce cross-lingual settings, refined architecture and gradual training can beat piling on data. The caveat: it is a sharp knife for a limited set of languages; scaling to dozens remains unproven.
Put together, the four papers point one way. A 30B hierarchical ViT that beats an 18B rival while activating 6.7B. A frame selector that subtracts wasted compute before it happens. RaMod that compresses full fine-tuning into targeted edits on hidden states. A speech translator that proves architecture can compensate for thin data. From basic research, AI data, cloud services and translation, the same logic runs through all of them: Huawei is replacing scale density with design density.
If the last two years of AI were a contest of mining ore, 2026 is showing the turn to smelting. Sparse activation, smart selection, modular adaptation, gradual training: none of these need more chips, only a deeper read of the problem. The parameter arms race is over. AI’s second half will be won by whoever understands the real deployment problems and has the engineering to solve them systematically.
Source: Leiphone (leiphone.com). Translated and adapted for The Robot Belt.