Research
LLMs Vulnerable to Document-Level Prompt Injection in Peer Review
Study uncovers LLM vulnerabilities to hidden prompt injections, altering peer review outcomes in multiple languages.
LLMs Exposed: Hidden Prompts Threaten Academic Peer Review
Study reveals LLMs' vulnerability to prompt injections, altering review outcomes—except in Arabic.
Vis-CoT Framework Elevates AI Transparency Through Human Collaboration
Vis-CoT's reasoning graphs boost LLM accuracy by 24%, paving the way for more reliable AI systems.
New Framework Revolutionizes Context Management in Language Models
EDU-based Context Compressor boosts LLM efficiency, cuts costs, and sets new standards with StructBench dataset.
RAVEL Elevates Text-to-Image Models with Graph-Based Retrieval
RAVEL enhances T2I models like Stable Diffusion XL without extra data, using graph-based retrieval for more nuanced image generation.
New Framework Revolutionizes Context Management in LLMs
EDU-based Context Compressor enhances LLMs by preserving structure, reducing costs, and setting new industry benchmarks.
CubeBench Reveals LLMs' Physical-World Task Limitations
A new benchmark uncovers large language models' weaknesses in spatial reasoning and planning, offering insights for AI advancement.
UniCR: Calibrating AI Uncertainty for Enhanced Trust
UniCR framework refines AI decision-making by calibrating uncertainty and enforcing error budgets, without altering base models.
UniCR: Boosting AI Trustworthiness Without Model Changes
UniCR enhances AI reliability by calibrating uncertainty and managing error budgets, impacting decision-making across sectors.
RAVEL: Elevating Text-to-Image Models Without Extra Training
RAVEL enhances diffusion models using graph-based retrieval, improving rare and culturally nuanced image generation without extra data.
CubeBench Reveals LLMs' Struggles with Real-World Tasks
New benchmark highlights large language models' challenges in spatial reasoning and planning.
Dream-VL and Dream-VLA: Pioneering Vision-Language Models
Dream-VL and Dream-VLA, diffusion-based models, set new standards in visual planning and robotics, surpassing autoregressive models.
SoulX-LiveTalk Raises the Bar in Real-Time Avatar Creation
Discover a 14B-parameter model boosting VR and gaming with advanced bidirectional techniques.
New Benchmark Exposes Cognitive Gaps in Multimodal Models
MME-CC benchmark reveals vision-centric evaluation needs, with closed-source models outperforming open-source rivals.
Open-Source LLMs Revolutionize Clinical Note Processing
Researchers unveil a cost-effective pipeline using open-source LLMs to enhance entity recognition in clinical documentation.