Research map / Vision-language and multimodal / Multimodal LLMs

Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Pragma-VL, an end-to-end MLLM alignment pipeline combining a risk-aware contrastive+SFT cold start with a parallel multi-head reward model trained on query-dependent weighted preference labels (PragmaSafe) and optimized with GRPO.

Problem

MLLMs face a safety-helpfulness trade-off: they over-refuse benign queries or miss latent cross-modal risks, and existing alignment uses static, context-agnostic trade-offs and weak visual risk perception.

Approach

Combine a risk-aware supervised contrastive cold start on the visual encoder (using BeaverTails-V severity tags) with a jointly trained parallel reward model that outputs helpfulness, harmlessness and a context-weighted scalar, trained on PragmaSafe labels with variance-aware query-dependent weights, then align via GRPO.

Key results

• On Qwen2.5-VL-7B, Pragma-VL reaches 31.66% ASR on MM-SafetyBench (vs. 48.75% base) and 63.47% SIUO safety rate (vs. 38.78% base).
• On Llava-1.5-7B, SIUO safety rises from 14.37% to 55.42%, and SPA-VL helpfulness win rate reaches 97.93%.
• Parallel reward architecture achieves 96.3% weighted accuracy at Δ≥2 vs. 85.5% for sequential and 79.1% for single-head on the PragmaSafe validation set.

Surprising findings

• Safe-RLHF-V reproduction and several specialized-data baselines (BeaverTails-V, SPA-VL) lose general-capability scores relative to the base model on GQA, MathVista and others.
• DPO on BeaverTails-V gains harmlessness (78.87%) at the cost of helpfulness (52.47%) on SPA-VL for Qwen.

Limitations

• Evaluation relies on GPT-4o as judge for win rates, with no human evaluation reported.
• Only two base models (Qwen2.5-VL-7B and Llava-1.5-7B) at 7B scale are evaluated.
• Theoretical ordering relies on asymptotic MLE and Fisher information arguments with assumptions (e.g., positively correlated gradients) that are asserted rather than empirically verified.

Experimental setup

Datasets: BeaverTails-V PragmaSafe SPA-VL MM-RLHF

Benchmarks: BeaverTails-V SPA-VL MM-SafetyBench SIUO MSSbench GQA ScienceQA TextVQA VizWiz VQAv2 MathVista

Metrics: win rate attack success rate effective rate safety rate

Baselines: SFT DPO Safe-RLHF-V SPA-VL MM-RLHF BeaverTails-V_harm BeaverTails-V_help MMSafe-PO

Builds on: GRPO Supervised Contrastive Learning Bradley-Terry reward modeling Safe RLHF-V BeaverTails-V

Closest papers by meaning

• Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.92)
• Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.91)
• SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models (CVPR 2025 · similarity 0.91)
• Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models (ICML 2024 · similarity 0.91)
• ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.90)
• SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.90)