NUS and collaborators propose ViF to curb visual hallucination snowballing in multi-agent systems

终结多智能体视觉幻觉“滚雪球”！新国立等提出ViF：无需改造模型，即插即用

NUS LV-Lab and collaborators proposed ViF, accepted to ICLR 2026. Across 8 benchmarks, 4 MAS structures, and 10 VLMs, it reports 2.4%–3.8% average gains. ViF replaces text-only passing with visual relay tokens and layered attention redistribution, cutting HS by over 30% on average and nearly 40% in ring topology.

Why it matters: HKR-H/K/R all pass: the hook is concrete, the mechanism and eval grid are disclosed, and hallucination control matters to agent builders. Scope stays research-heavy, so it sits at the featured threshold, not same-day must-write.