Multimodal AI News and AI Updates
Follow Multimodal AI developments across AI companies, labs, and open-source projects.
Latest Multimodal AI news, research, benchmarks, product releases, and industry adoption updates.
Latest Articles
- HTCE: Homeomorphism-inspired Topological Cross-Scene Experience Canonicalization and Reuse for Object Navigation — HTCE lets embodied navigation agents reuse past successful routes across different homes by canonicalizing their topology, topping SR and SPL on HM3D and MP3D with a frozen vision-language policy.
- JHiMA-WebRL: A Joint Hierarchical Multi-Agent Reinforcement Learning Framework for Long-Horizon Web Tasks — JHiMA-WebRL jointly trains a 7B task decomposer and a 7B actor with RL and reaches 37.2% success on WebArena-Lite-v2, versus 24% for a 32B single-agent baseline trained on the same tasks.
- DiMuRe: Diagnosis-Guided Multilevel Recovery for Robotic Task Execution — DiMuRe lets robot agents diagnose mid-task disturbances and pick a recovery level, raising complete-task success from 58.7% to 77.7% on a new disturbance benchmark, and to 81.9% with replanning.
- BtsPruner: Role-Aware Visual Token Pruning for Large Vision-Language Models — BtsPruner, a training-free visual token pruning method, keeps 91.6% of LLaVA-1.5-7B performance with just 32 of 576 visual tokens and gives up to 5.95x CUDA-latency speedup on LLaVA-NeXT-7B.
- Controlling Action-Head Co-Adaptation to VLM Representation Anisotropy — A study of vision-language-action models finds action heads over-focus on high-variance VLM dimensions, and a parameter-free fix called AdaIE adds an average 19 points of success across five real-world manipulation tasks.
- Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs — The H2RB framework adapts pretrained vision-language-action robot policies using unpaired human demonstrations, raising real-world object transfer scores from 8.0 to 51.8.
- DualWorldBench: Can Agents Plan Deliveries across Symbolic and Grounded Worlds? — DualWorldBench tests agents on pickup-and-delivery planning in symbolic and grounded worlds and finds they understand local scenes but struggle to build globally feasible multi-goal plans.
- GlucoAgent: Personalized Glucose Forecasting from Multimodal Personal Context — GlucoAgent, an LLM agent that writes and runs code over a person's CGM, wearable and meal data, beat Chronos-2 on glucose forecasting with an average RMSE of 22.5 vs 23.5 mg/dL on CGMacros.
- AceAction-5M: Hierarchical Goals for Generalist Action Learning Across 3D Game Worlds — AceAction is a dataset of human gameplay across multiple 3D games with hierarchical task, subgoal and skill annotations aligned to keyboard-and-mouse actions, aimed at training generalist vision-language-action agents.
- DGC-Net: Cross-Modal Dynamic Routing with Global-Guided Compensation in Multimodal Sentiment Analysis — DGC-Net improves multimodal sentiment analysis by routing text, audio and visual cues through global-guided expert selection, aiming to fix inconsistent signals across modalities in user videos.