Visual Token Routing cuts VLM agent visual tokens up to 5.8x at same accuracy
wellecks · x · 2026-11-05
Visual Token Routing (VTR) is a new method for faster, cheaper VLM agents: it jointly trains a lightweight router that drops temporally redundant visual tokens — or tokens the agent doesn't need to act on — cutting visual tokens by up to 5.8x with no accuracy loss.
Researcher Akari Asai uses it to raise a bigger question: why can't one computer-use agent handle all agentic tasks? Do visual web navigation and deep research really need separate scaffolds and training pipelines?
She favors a unified "look at the screen and decide what to click, type, or read next" approach, but flags the bottleneck: re-evaluating the screen after every action. BrowseComp runs often exceed 50 searches, so doing it all through a browser gets expensive fast. VTR learning which visual tokens to skip suggests a research agent could go much further on the same compute budget.
More from coding & agent
- OpenAI hires Pierre, whose open-source dev tools already power parts of Codex — gdb · 2026-11-05
- Agent ships hf_dataset() in DuckDB: query Hugging Face datasets with SQL — lhoestq · 2026-11-05
- French search engine ibou launches API to ground LLMs and AI agents in the real web — IgorCarron · 2026-11-05
- YC's Weave raises $13.5M to measure whether AI coding tools actually pay off — ycombinator · 2026-11-05
- Team places 7th of 5,000+ in ARC-AGI-3 Challenge: a good harness beats a bigger model — gklambauer · 2026-11-05
- Intent mobile adds PR views, syntax highlighting and private agent secret requests — LukeW · 2026-11-05