Live intelligence

Latest AI updates

1 updates
Research

ReFrain Adds Abstention to Multimodal Models Without Losing Correct Answers

Researchers propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with a displayed "none of the above" option. On Qwen2.5-VL-3B at a zero-budget target, ReFrain retained all 57 SFT-gained answers with no losses while improving correct abstention by 7.30 percentage points. At a 0.5% calibration budget it reached 74.05% correct abstention with 8 losses, versus CLIP-UP EmbLoRA at 64.65% with 62 losses. Confirmation on 228 independent question clusters yielded 66.05% correct abstention with no losses at the 0.5% point.

multimodal modelsabstention and refusalmodel calibration