Research map / Vision-language and multimodal / Multimodal LLMs

EcoAlign: An Economically Rational Framework for Efficient LVLM Alignment

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 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 EcoAlign, an inference-time, budget-constrained thought-graph search that scores actions by a forward-looking economic value and uses a weakest-link safety rule to improve LVLM safety and utility at lower token cost.

Problem

LVLMs are vulnerable to jailbreaks, and existing alignment methods trade off safety, utility, and computational cost; output-only or additive safety scores miss harmful reasoning processes and waste compute.

Approach

Treats LVLM inference as budget-constrained optimal path search over a dynamic DAG of thoughts; actions are valued by a net-present-value-style lookahead whose horizon shrinks as budget is consumed, and path safety is the minimum node safety (weakest link), with nodes scoring below zero pruned.

Key results

• On GPT-4o, EcoAlign reaches 96.5 MMSafety and 91.8 MSSBench with Avg. Cost 21.2, versus CoT at 104.3 Avg. Cost.
• On Gemini-2.5-Flash, EcoAlign scores 97.7 on MMSafetyBench vs. 87.8 for the next-best baseline (VLM-Guard) and 64.2 for base.
• Removing cost control (w/o Cost Control) raises GPT-4o Avg. Cost from 21.2 to 79.6 with no safety gain.

Surprising findings

• Removing cost normalization from the value function increases cost sharply while yielding negligible safety change, and can lower OCRBench (GPT-4o 86.0 to 76.1).
• Low budget (B=500) paradoxically increased average cost while dropping utility, due to early termination.

Limitations

• Relies on the LVLM's own self-evaluation scores for safety, utility, and cost, with no external verifier.
• Evaluation uses GPT-4o as judge, and the paper's own tables show inconsistent column headers and numbers across sections.
• Cost is measured in tokens and API pricing rather than wall-clock or hardware compute for closed models.

Experimental setup

Datasets: MMSafetyBench MSSBench SIUO OCRBench MathVista MMStar

Benchmarks: MMSafetyBench MSSBench SIUO OCRBench MathVista MMStar

Metrics: safety score utility accuracy average normalized cost latency

Baselines: Base CoT CoD VLM-Guard Best-of-N GoT

Builds on: chain-of-thought chain-of-draft graph of thoughts VLM-Guard

Closest papers by meaning

• ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.89)
• Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models (ICLR 2025 · similarity 0.88)
• SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.88)
• DAVSP: Safety Alignment for Large Vision-Language Models via Deep Aligned Visual Safety Prompt (AAAI 2026 · similarity 0.88)
• LVLM-Aided Alignment of Task-Specific Vision Models (CVPR 2026 · similarity 0.88)
• Turning Drift into Constraint: Robust Reasoning Alignment in Non-Stationary Multi-Stream Environments (ICML 2026 · similarity 0.88)