SIVA-RL: Sensitivity-Invariance Visual Alignment for Multimodal Reinforcement Learning
Merged summary
TL;DR — SIVA-RL is a reinforcement-learning framework that makes vision-language models actually ground their answers in visual evidence, replacing intervention-type supervision with outcome-based, sample-wise supervision. It matters because RLVR can produce correct answers that ignore the image, undermining trustworthy multimodal reasoning.
- Uses token-aligned, distance-constrained within-image "PatchSwap" to build localized interventions; a frozen audit policy scores each clean-vs-intervention pair, and the observed reward drop becomes soft routing weights.
- Routes pairs by effect: large-drop pairs drive sensitivity alignment, low-drop pairs drive clean-anchored invariance alignment, ambiguous pairs are down-weighted — decoupling intervention construction from supervision assignment.
- Backbone-agnostic (compatible with GRPO and DAPO); improves 3B and 7B models over matched RL baselines across all nine benchmarks tested.
- Reports an 8.79 percentage-point gain on vision-dependent reasoning and up to 14.9% relative overall improvement across the four configurations.
Sources (1)
SIVA-RL: Sensitivity-Invariance Visual Alignment for Multimodal Reinforcement Learning
TL;DR — SIVA-RL is a reinforcement-learning framework that makes vision-language models actually ground their answers in visual evidence, replacing intervention-type supervision with outcome-based, sample-wise supervision. It matters because RLVR can produce correct answers that ignore the image, undermining trustworthy multimodal reasoning.
- Uses token-aligned, distance-constrained within-image "PatchSwap" to build localized interventions; a frozen audit policy scores each clean-vs-intervention pair, and the observed reward drop becomes soft routing weights.
- Routes pairs by effect: large-drop pairs drive sensitivity alignment, low-drop pairs drive clean-anchored invariance alignment, ambiguous pairs are down-weighted — decoupling intervention construction from supervision assignment.
- Backbone-agnostic (compatible with GRPO and DAPO); improves 3B and 7B models over matched RL baselines across all nine benchmarks tested.
- Reports an 8.79 percentage-point gain on vision-dependent reasoning and up to 14.9% relative overall improvement across the four configurations.