← Back to research

Research Brief · 2026-10-27 Core

Live-OPSD: On-Policy Self-Distillation for Proactive Streaming Large Video Models

The idea, in brief

Generate from causal context, then score the same continuation with training-only evidence hints. Distill those distributions back into the unhinted response/silence policy.

Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.
Live-OPSD privileged-context self-distillation framework
Live-OPSD privileged-context self-distillation framework Original figure · paper authors ↗

Release evidence

Paper describes training and evaluation; no method-specific code, weights or data release was verified.

Last checked · 2026-10-31