arxiv arXiv cs.CL · just now · research

ClinFusion: vision-centric MLLM sets new state of the art in medical benchmarks

from English

Researchers introduce ClinFusion, a vision-centric multimodal large language model designed for holistic medical understanding that unifies 2D and native 3D image analysis. The system utilizes a compositional cascaded vision encoder with a Cascade Spatial-Aware Locality Fusion operator to address the challenges of heterogeneous medical imaging.

ClinFusion outperforms leading open-source models like Hulu-Med and Lingshu on 20 out of 24 benchmarks, including visual question answering and report generation.
It demonstrates multimodal capabilities superior to proprietary models such as GPT-5.2 and Gemini-3-Flash on 13 out of 16 benchmarks.
The authors propose MedIF-Bench for instruction-following assessment and a region-of-interest-grounded method for factualness-driven report evaluation.
A blinded evaluation by board-certified radiologists confirmed ClinFusion produced the highest-ranked reports, validating the new RoI-grounded metric's correlation with expert judgment.

ClinFusion establishes a new state of the art across comprehensive medical benchmarks and can be augmented with agentic tool use for retrieval-augmented clinical workflows.

Importance 2/3 arXiv cs.CL Multimodal Research paper

Related articles

lab NVIDIA Research · 8h ago

Multimodal AI analysis of non-verbal communication in Delta Force

Researchers analyze gameplay videos from the competitive shooter game Delta Force to extract and understand emergent non-verbal communication, such as gestures and movement patterns. This work addresses a gap in prior studies that have largely focused on MOBA games or verbal coordination in other shooters.

arxiv arXiv cs.CL · 12d ago

CSSEL-P2P enables simultaneous speech translation without architectural changes

The article presents CSSEL-P2P, a data-driven approach for simultaneous speech translation that avoids the brittle architectural changes and explicit read/write policies often required by decoder-only LLM systems. The method utilizes fixed-length chunks for cumulative streaming decoding with a rewind-based committed prefix, combined with teacher-labeled prefix-to-prefix targets and bounded waiting for fine-tuning.

arxiv arXiv cs.CL · 12d ago

Groc-PO introduces grounded context preference optimization to reduce multimodal hallucinations

Researchers propose Grounded Context Preference Optimization (Groc-PO), a framework designed to address untruthfulness issues like visual hallucinations and content fabrication in Multimodal Large Language Models (MLLMs). The method tackles the credit-assignment challenge where standard Direct Preference Optimization (DPO) fails to suppress error propagation from early grounding stages by applying supervision only at the final answer.

arxiv arXiv cs.CL · 13d ago

Interactive multi-feature fusion improves semantic reconstruction from non-invasive brain recordings

A new framework addresses the representational mismatch in continuous semantic reconstruction by fusing static lexical representations (W2V) with dynamic contextualized representations (GPT). The study benchmarks two integration approaches: linear Naive Concatenation and non-linear Multi-Head Cross-Attention.

arxiv arXiv cs.CL · 13d ago

Audio-language embedding models fail to encode negated sound concepts

A study reveals that audio-language embedding models like CLAP struggle with negation, mapping affirmative and negated captions to nearly identical representations. To expose this limitation, the authors introduce NegEval-Audio, a framework converting datasets into Retrieval-Neg and Multiple-Choice Negation tasks.