UCLA and Apple's OpenVLHarness adds multimodal memory to visual agentic tool-use
Yu Zhou · hf · 2026-10-20
A UCLA–Apple team introduces OpenVLHarness, a multimodal reasoning harness for real-world visual agentic tool-use. A controlled study on 22 benchmarks shows simply increasing inference budget yields only modest gains — tasks demand grounded perception, external knowledge, and numerical measurements via heterogeneous tools. OpenVLHarness combines a capability layer of task-level tool interfaces and evidence rendering with persistent multimodal memory that retains images and structured data across reasoning steps. Code is open-sourced.
More from coding & agent
- HSPR Recovers Full Agent Benchmark Scores From Small Subsets via Hierarchical Posterior Inference — SUDA · 2026-10-20
- AdaExploreBench Finds LLM Agents Struggle at Proactive Exploration in Evolving Environments — HKUST · 2026-10-20
- Our OAuth discovery doc and Python docs give agents opposite signup instructions — bshivarthy · 2026-10-20
- Fo launches MCP to let AI agents control and sense physical devices, 100+ devs already building — heyshrutimishra · 2026-10-20
- Open-source humanoid Asimov 1 ships docs: assembly guides, sim-to-real training, browser demo — Sentdex · 2026-10-20
- Cal.com Calendar MCP Server Lets AI Schedule, Update and Delete Events — modelcontextprotocol · 2026-10-20