multimodal-agents

Official codebase for the paper "WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction"

Official Repository for NeurIPS'25 Paper "Tool-Augmented Spatiotemporal Reasoning for Streamlining Video Question Answering Task"

MedCTA: A Benchmark for Clinical Tool Agents

PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs

A modular framework for benchmarking multimodal AI agents in a reproducible, full-OS environment. Using and adaption of the Smolagents's CodeAgent, Docker containers to run the VM in, VM's created using Qemu.

Legacy PosterMELD project-page history; the production site now ships with the main PosterMELD repository.

Evaluation of GPT-4o-mini on OSWorld desktop automation benchmark. Compares screenshot-only vs accessibility tree-enhanced approaches across 10 tasks (Chrome, LibreOffice, file ops, etc). Documents critical coordinate extraction failures and provides architectural recommendations for GUI agents.

几何可信具身导航

Adaptive Recursive Image Grid for Multimodal Agents.

Fully local AI desktop application that uses multi-agent orchestration to analyze short videos (~1 min) through natural language queries. All AI inference runs offline with no cloud dependencies.

CoSee is a research prototype for diagnosing shared-state collaboration in resource-constrained visual agents using an auditable Board-based multimodal VQA workflow.