CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

Hugging Face Daily Papers Papers

Summary

CORE introduces a distillation method that transfers compositional ranking judgments from a reranker to an embedding model using a Rank-KL objective, enhancing compositional retrieval performance across benchmarks without compromising standard tasks.

MLLM-based embedding models remain limited in compositional retrieval, often failing to distinguish scenes containing the same concepts but different attribute-object bindings. Yet the same backbone can resolve such distinctions when used as a cross-attentive reranker, motivating us to distill its compositional judgments into the embedding model. We propose CORE, which synthesizes candidate lists spanning five compositional matching levels and introduces a Rank-KL objective that trains the embedding model to reproduce the reranker's fine-grained ranking. We further introduce a graded evaluation protocol and compare contrastive learning, pairwise CoSENT, and listwise Rank-KL under the same data and tuning budget. Our comparison shows that both CoSENT and Rank-KL use the multi-level supervision more effectively than contrastive learning, with Rank-KL achieving the strongest overall performance. Across three compositional reasoning benchmarks (COLA, SUGARCREPE++, NEGBENCH), CORE-RERANKER-8B achieves an 82.7% total average, outperforming Jina-Reranker by 10.7 points, while CORE-EMBED-8B achieves the best total average (0.666) among all evaluated embedding models. The improvements transfer to the MCMR benchmark without sacrificing retrieval performance on COCO and Flickr30K.
Original Article
View Cached Full Text

Cached at: 09/04/26, 03:56 AM

Paper page - CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

Source: https://huggingface.co/papers/2609.04083

Abstract

CORE distills compositional ranking judgments from a cross-attentive reranker into an embedding model via synthesized multi-level candidates and a Rank-KL objective, improving compositional retrieval without degrading standard performance.

MLLM-based embedding modelsremain limited incompositional retrieval, often failing to distinguish scenes containing the same concepts but different attribute-object bindings. Yet the same backbone can resolve such distinctions when used as across-attentive reranker, motivating us to distill its compositional judgments into the embedding model. We propose CORE, which synthesizes candidate lists spanning five compositional matching levels and introduces aRank-KL objectivethat trains the embedding model to reproduce the reranker’s fine-grained ranking. We further introduce a graded evaluation protocol and comparecontrastive learning, pairwiseCoSENT, and listwise Rank-KL under the same data and tuning budget. Our comparison shows that bothCoSENTand Rank-KL use the multi-level supervision more effectively thancontrastive learning, with Rank-KL achieving the strongest overall performance. Across threecompositional reasoning benchmarks(COLA, SUGARCREPE++, NEGBENCH), CORE-RERANKER-8B achieves an 82.7% total average, outperforming Jina-Reranker by 10.7 points, while CORE-EMBED-8B achieves the best total average (0.666) among all evaluated embedding models. The improvements transfer to the MCMR benchmark without sacrificing retrieval performance on COCO and Flickr30K.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2609\.04083

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

Models citing this paper4

#### Alibaba-NLP/core-reranker-2b Feature Extraction• 2B• Updatedabout 2 hours ago • 1 #### Alibaba-NLP/core-emb-2b Feature Extraction• Updatedabout 2 hours ago • 1 #### Alibaba-NLP/core-reranker-8b Feature Extraction• 9B• Updatedabout 2 hours ago #### Alibaba-NLP/core-emb-8b Feature Extraction• Updatedabout 2 hours ago

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.04083 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.04083 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Enhancing Rubric-based RL via Self-Distillation

Hugging Face Daily Papers

Introduces CriPO, a method that enhances rubric-based reinforcement learning for LLMs via on-policy self-distillation, addressing unexplored and suppressed criteria. Experiments on medicine and science benchmarks show stronger final performance with roughly 2x fewer optimization steps.

Learning to Refine Hidden States for Reliable LLM Reasoning

arXiv cs.LG

Proposes ReLAR, a reinforcement-guided latent refinement framework that iteratively updates hidden representations in LLMs before decoding, improving reasoning reliability and efficiency compared to chain-of-thought methods.