Visual Token Routing cuts VLM agent visual tokens by 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: a lightweight router is jointly trained to drop visual tokens that are temporally redundant or irrelevant to the agent's actions.
- It achieves the same accuracy with up to 5.8x fewer visual tokens.
- Especially relevant for computer-use agents, where screenshots are high-resolution but frame-to-frame deltas are small (e.g. realtime/games).
- Researchers are highlighting it as promising work in the "smarter visual encoding" line of research.
More from coding & agent
- Agent Tool 'Dot' Lets Users Send Feature Requests Directly to the Dev Team — dkundel · 2026-11-05
- Model gets stuck in a loop spamming "it failed!!!" — burning 140K tokens — CtrlAltDwayne · 2026-11-05
- AX to support migrating agentic workloads between local and production clusters — rakyll · 2026-11-05
- OpenAI upgrades Codex: marked-up images now pinpoint edits via API — OpenAIDevs · 2026-11-05
- Baseten Router cuts coding-task costs 72% with only 0.8pt pass-rate drop vs Opus 5.5 — baseten · 2026-11-05
- BrowserForge: open-source MCP server gives agents your real browser with human approval gates — mostafazog · 2026-11-05