蚂蚁集团 IJCAI 2026 论文盘点：让 AI 学会「随机应变」

TL;DR - A media roundup of Ant Group's four IJCAI-ECAI 2026 accepted papers, framed around a common theme: moving AI from static solvers to systems that adapt to changing environments (data density, distribution shift, search dimensionality, evolving AIGC ecosystems). It matters as a snapshot of how an industrial lab aligns academic output with production constraints like real-time risk control and 1B+ user traffic swings.

• MSRGC-Net (with Chongqing Univ. of Posts & Telecom): training-free multi-scale echo state network reservoirs plus density-adaptive "granular ball" anchor graphs for time-series clustering; no preset cluster count, near-linear complexity instead of O(n²), and 12 best / 2 second-best across 15 metrics on 5 multivariate benchmarks.
• ROAD (with SJTU): treats the offline/online data mixing ratio in offline-to-online RL as a dynamic decision variable via bi-level optimization (inner Bellman error, outer expected online return) solved approximately with multi-armed bandits; plugs into IQL/PEX/CQL/Cal-QL, with PEX+ROAD scoring 71.12 average and 18/24 firsts on D4RL tasks.
• DSEBO (with ECNU and Nanjing Univ.): high-dimensional Bayesian optimization that grows the random-embedding subspace dimension on convergence triggers, using a shared embedding matrix so subspaces nest; adaptive expansion step size keyed to the best-value curve, evaluated on D=1000 synthetic functions and MSLR/Lasso-Hard/LIMO against REMBO, SIRBO, BAxUS, TuRBO.
• VGA-BenchV2 (with Beijing Film Academy and BIGAI): adds 36,000 task-level human annotations (13.46×/11.15×/1.55× over VGA-Bench) and a hybrid evaluator (VAQA-Net for continuous aesthetic scores; VTag-Net/VGQA-Net on Qwen VLMs), then reuses the learned aesthetic evaluator as an RL reward signal to fine-tune generators — closing the loop from scoring to optimization; dataset is open on Hugging Face.

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.