StarTV-Agent pairs fast workflow reuse with slow reasoning for traffic scene analysis
Researchers propose an agent that reuses proven workflows for routine traffic-analysis tasks and switches to deliberate step-by-step reasoning when no suitable workflow exists, reporting accuracy gains and lower token costs.
In plain words
Understanding traffic scenes from images needs both sharp visual analysis and reasoning about space and time. The described agent keeps a memory of workflows that worked before, reusing them quickly on familiar tasks, and falls back to slower, step-by-step planning for new ones. The authors say this cuts computing cost while improving results.
What was announced
Tool-augmented vision-language models can call specialized perception tools, but the paper identifies two obstacles: terminal rewards give little guidance for intermediate decisions and error recovery, and repeated planning adds cost across similar tasks. StarTV-Agent addresses this by echoing the fast-slow split of human cognition, combining policy learning with experience evolution so the agent reasons adaptively when needed and reuses prior experience when possible.
Training uses two components: Role-Sensitive Process Optimization, which blends process advantages with terminal advantages, and Weakness-Driven Demonstration Learning, which supplies validated teacher demonstrations aimed at the student's weak questions. At inference, Lightweight Adaptive Workflow Reuse offers a fast path that selects suitable workflows and binds task-specific arguments and resource references, while a slower ReAct-based path handles tasks without a matching workflow. Label-free feedback based on execution completion, workflow adoption, and cost savings guides reuse and memory updates. Across three benchmarks, the authors report up to 8.0% relative performance improvement and token consumption reduced by up to 52.1%.
What it does not show
- Evidence is limited to the paper's abstract; benchmark names, baselines, and evaluation details are not stated, and the reported figures are the authors' own claims without independent confirmation.
For security and IT teams
- Scope
- The source is an arXiv abstract describing the method and headline results only; it does not cover deployment details, model availability, or full experimental setup.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.