Beyond Scalar Distances: Semantic Attribute Gradients from Frozen MLLMs for Visual Embeddings

Hugging Face Daily Papers Papers
multimodal-llm vision-encoders attribute-aware zero-shot-retrieval group-relative-policy-optimization metric-learning

Summary

SAGA framework uses frozen multimodal large language models to provide attribute-aware supervision for vision encoders via Group Relative Policy Optimization, improving zero-shot image retrieval by 3–6 points on fine-grained benchmarks.

Vision encoders for retrieval are typically trained with class-label supervision: each training pair reduces to a scalar that uniformly pushes the embedding apart or pulls it together, as if every visual attribute either differed or matched. A multimodal large language model (MLLM), shown the same pair, can articulate those attributes and use them to predict whether the images share a class. We propose SAGA, a framework that turns this language-grounded, attribute-aware perception into a training signal for the encoder itself. Specifically, we use Group Relative Policy Optimization (GRPO) to reward the MLLM for correct predictions on the vision encoder's tokens. Since correct predictions require those tokens to expose the specific attributes that differ or match between the pair, the gradient pushes the encoder to encode them, replacing the uniform pair-level scalar with attribute-resolved supervision. An auxiliary attention-distillation loss anchors the encoder's embedding to tokens the MLLM attended to, and a standard metric-learning loss shapes the embedding geometry for nearest-neighbour retrieval. The MLLM is frozen throughout and discarded at inference, matching the deployment cost of a metric-learning baseline. SAGA improves Recall@1 by 3 to 6 points over state-of-the-art baselines on CUB-200-2011, Cars-196, FGVC-Aircraft, and iNaturalist Aves on zero-shot image retrieval.
Original Article
View Cached Full Text

Cached at: 06/17/26, 07:52 PM

Paper page - Beyond Scalar Distances: Semantic Attribute Gradients from Frozen MLLMs for Visual Embeddings

Source: https://huggingface.co/papers/2606.15134

Abstract

SAGA framework uses multimodal large language models to provide attribute-aware supervision for vision encoders through Group Relative Policy Optimization, improving zero-shot image retrieval performance.

Vision encodersfor retrieval are typically trained withclass-label supervision: each training pair reduces to a scalar that uniformly pushes the embedding apart or pulls it together, as if every visual attribute either differed or matched. Amultimodal large language model(MLLM), shown the same pair, can articulate those attributes and use them to predict whether the images share a class. We propose SAGA, a framework that turns this language-grounded,attribute-aware perceptioninto a training signal for the encoder itself. Specifically, we useGroup Relative Policy Optimization(GRPO) to reward the MLLM for correct predictions on the vision encoder’s tokens. Since correct predictions require those tokens to expose the specific attributes that differ or match between the pair, the gradient pushes the encoder to encode them, replacing the uniform pair-level scalar with attribute-resolved supervision. An auxiliaryattention-distillation lossanchors the encoder’s embedding to tokens the MLLM attended to, and a standardmetric-learning lossshapes the embedding geometry for nearest-neighbour retrieval. The MLLM is frozen throughout and discarded at inference, matching the deployment cost of a metric-learning baseline. SAGA improvesRecall@1by 3 to 6 points over state-of-the-art baselines on CUB-200-2011, Cars-196, FGVC-Aircraft, and iNaturalist Aves onzero-shot image retrieval.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2606\.15134

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.15134 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.15134 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.15134 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Stateful Visual Encoders for Vision-Language Models

Hugging Face Daily Papers

This paper introduces a stateful visual encoder for vision-language models that conditions visual representations on prior features, enabling better visual comparison in multi-image and agentic settings. The method shows consistent improvements across tasks such as cross-image spatial aggregation and longitudinal radiology.

Squeezing Capacity from Multimodal Large Language Models for Subject-driven Generation

Hugging Face Daily Papers

This paper proposes a novel approach that conditions diffusion models on Multimodal Large Language Models (MLLMs) for subject-driven image generation, using VAE-based identity conditioning and a Dual Layer Aggregation module to improve both semantic understanding and identity preservation while mitigating copy-paste artifacts.

LLaVA-UHD v4: What Makes Efficient Visual Encoding in MLLMs?

Hugging Face Daily Papers

This paper introduces LLaVA-UHD v4, which improves visual encoding efficiency in multimodal large language models by using slice-based encoding and intra-ViT early compression. It reduces computational costs by over 55% while maintaining or improving performance on high-resolution image tasks.