DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation
The paper titled 'DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation' addresses inefficiencies in the training of Vision-Language Models. It highlights that a significant portion of prompts used in standard datasets are ineffective due to minimal learning signals. The authors propose a new dataset, DeltaPrompts, which aims to improve model performance by focusing on prompts that reveal functional capability gaps.
- ▪Up to 69% of prompts in standard reasoning datasets are effectively zero-delta, providing minimal learning signal.
- ▪The proposed DeltaPrompts dataset consists of 200k synthetic, high-divergence reasoning problems.
- ▪DeltaPrompts yields up to 15% relative improvement on various benchmarks, even with highly-optimized reasoning models.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.15532 (cs) [Submitted on 15 May 2026] Title:DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation Authors:Jaehun Jung, Hyunwoo Kim, Brandon Cui, Ximing Lu, David Acuna, Prithviraj Ammanabrolu, Yejin Choi View a PDF of the paper titled DeltaPrompts: Escaping the Zero-Delta Trap in Multimodal Distillation, by Jaehun Jung and 6 other authors View PDF HTML (experimental) Abstract:Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
