A research team led by Professor Chang D. Yoo at KAIST has introduced **VOTP (Video-based Optimal TransPort Preference)**, a technique that lets AI systems learn human intentions and judgment criteria from just a few preference videos rather than the thousands to tens of thousands of human evaluations typically required. The work was selected as an oral presentation at ICML 2026, an honor reserved for roughly the top 0.7% of submissions — 168 of 23,918 papers.
The central problem VOTP tackles is the cost of preference data. Aligning models to human values usually depends on large volumes of human feedback, where people rank or rate outputs so the system can infer what "good" looks like. That process is slow, expensive, and hard to scale. VOTP reframes the task around short videos that demonstrate preferences, using optimal transport — a mathematical framework for efficiently matching one distribution to another — to extract the underlying judgment criteria from a sparse set of examples.
If the approach generalizes, the implications are significant for reinforcement learning from human feedback and for robotics, where collecting dense human labels is especially burdensome. Learning robust preferences from a handful of demonstrations could lower the barrier to aligning agents with nuanced human goals, particularly in domains where each evaluation is costly to obtain.
The recognition at ICML 2026 underscores growing academic momentum around data-efficient alignment, a counterweight to the prevailing trend of scaling up data and compute. As frontier systems push deeper into agentic and embodied applications, methods that teach machines what humans actually want — cheaply and from limited signal — are becoming a research priority in their own right.
VOTP joins a wave of 2026 work, spanning institutions like KAIST and the University of Pennsylvania, aimed at making AI both more capable and more efficiently aligned with human judgment.
Source: [Mirage News — KAIST Breakthrough](https://www.miragenews.com/kaist-breakthrough-robots-now-mimic-human-1689467/)