Beyond Structural Unification: A Perspective and Roadmap on Unified Multimodal Models
A position paper outlining five layers of functional unification for multimodal understanding and generation: inputs, representations, reasoning, optimization, and evaluation.
Sep 30, 2026
MASCOT: Multi-Agent Socio-Collaborative Companion Systems
MASCOT is a multi-agent framework for socio-collaborative companions that uses bi-level optimization—persona-aware behavioral alignment and collaborative dialogue optimization—to counter persona collapse and social sycophancy, improving role consistency and reducing redundant dialogue.
Sep 1, 2026
CultureVLM: Characterizing and Improving Cultural Understanding of Vision-Language Models for over 100 Countries
CultureVLM characterizes and improves cultural understanding of vision-language models across more than 100 countries using culturally-grounded benchmarks and training procedures.
Sep 1, 2026
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
SARA is a unified RAG framework that balances local factual precision with global coverage by combining natural-language spans with compact semantic compression vectors, achieving consistent gains under strict context budgets.
Jul 1, 2026
Reasoning Is Not All You Need: Examining LLMs for Multi-Turn Mental Health Conversations
A systematic study showing that reasoning capabilities alone are insufficient for LLMs in multi-turn mental health conversations, isolating failure modes that demand additional safety and empathy-aware design.
Jul 1, 2026
MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A
Structure-aware multimodal RAG for enterprise Q&A, combining layout-aware report parsing, page-level slide representations, and inference-time context assembly. Includes FastRAGEval for efficient answer evaluation.
Jul 1, 2026
MedHalu: Hallucinations in Responses to Healthcare Queries by Large Language Models
MedHalu is a fine-grained benchmark for studying hallucinations in LLM responses to consumer healthcare queries, analyzing hallucination patterns across models, query types, and medical specialties.
Jun 1, 2026
TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization
Regularizes prompt optimization to limit prompt bloat and narrow, sample-specific rules. Combines gradient purification, semantic edit regularization, and regularization-guided updates to improve out-of-distribution generalization.
May 20, 2026
UniSD: Towards a Unified Self-Distillation Framework for Large Language Models
A unified study of LLM self-distillation that combines teacher agreement, EMA stabilization, contrastive learning, feature matching, and divergence clipping to improve adaptation without stronger external teachers.
May 7, 2026
Sysformer: Safeguarding Frozen Large Language Models with Adaptive System Prompts
Sysformer learns adaptive, query-conditioned system prompts to safeguard frozen large language models, providing fine-grained safety control without modifying model weights.
Apr 23, 2026