The paper introduces ReTurn, a benchmark of 7,000 multimodal tasks designed to evaluate how models selectively use conversational history. ReTurn includes pairs that test task-carrying history (Reconfirm/Reground) and evidence-carrying history (Retrieve/Rebind), each preserving the same target question, media, and answer. Experiments on 13 vision‑language and audio‑language models show a drop in open‑ended accuracy from 93.7% to 72.3% when history is involved, revealing that high question recall does not guarantee effective task application and that competing media can misdirect responses.
By Shuoyang Sun, Kerui Gu, Hao Fang, Shaoli Huang, Bin Chen
arXiv:2610.29249v1 Announce Type: new
Abstract: Omni-modal assistants must remember not only what users say, but also what they show and convey. Existing long-term memory benchmarks are mostly text-a...
By Hong Ting Tsang, Boyi Zhang, Haoyu Huang, Jiaxin Bai, Yufei Li, Yisen Gao, Zhongwei Xie, Hongyu Luo, Jeff Z. Pan, Yangqiu Song
The article surveys multi‑turn conversational AI, highlighting its shift from isolated text prompts to sustained, multimodal interactions that involve clarifying goals, revising requests, and switching topics. It reviews literature across text‑only dialogue, AudioLLMs, multimodal and omni‑modal systems, and tool‑augmented agents, organizing findings around datasets, models, training, evaluation, and cross‑cutting challenges. The analysis reveals that while multimodal perception and action have progressed rapidly, systems still struggle with persistent memory, cross‑turn grounding, full‑duplex interaction, robust evaluation, and cultural alignment.
By Syeda Faiza Ahmed, Zien Sheikh Ali, Hunzalah Hassan Bhatti, Firoj Alam, Shammur Absar Chowdhury
VoiceLongMemEval (VLME) is a new benchmark that tests AI assistants on their ability to remember how users sounded by incorporating paralinguistic metadata—such as emotion labels, prosody descriptors, and voice events—into each conversational turn. The benchmark uses a three‑stage adversarial gate to ensure that models cannot succeed with transcript alone, revealing a significant affect gap: models gain 0.09 to 0.38 accuracy when provided with paralinguistic cues, and audio‑native models outperform standard ASR pipelines in extracting these signals. The dataset and code will be released upon acceptance.
By Ramit Pahwa, Parivesh Priye, Apoorva Beedu
arXiv:2610.27911v1 Announce Type: new
Abstract: Enterprise agents answer questions over workplace records that evolve through team collaboration. As information is shared, corrected, and revised acro...
By Seohyun Lee, Jongha Kim, Dohwan Ko, Ji Soo Lee, Seoung Choi, Jayun Ku, Hyunwoo J. Kim
UTILMEM is a new diagnostic benchmark that tests how conversational agents use long‑term memory, focusing on reasoning over dense histories, spotting implicitly relevant memories, synthesizing distributed evidence, and resisting interference from similar distractors. It contains 1,717 instances across five domains and evaluates a range of retrieval‑based and memory‑augmented systems. The study shows that strong performance on traditional factual recall does not guarantee effective memory utilization, highlighting a gap between retrieving information and integrating it into coherent, task‑oriented outputs.
By Peijun Qing, Fobo Shi, Soroush Vosoughi
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
By Qiji Zhou, Junyi Hu, Peiyu Wang, Yue Zhang
arXiv:2607. 02020v1 Announce Type: new Abstract: Multimodal large language models must continually adapt to evolving tasks and domains, yet standard continual learning metrics mainly measure whether old answers remain correct, leaving the stability of multimodal grounding largely unexamined.
By Qianyu Chen, Canran Xiao, Runxuan Tang
Large language model agents are increasingly deployed in human-human interaction settings, such as meeting assistants and clinical documentation systems, where they must observe conversations and retain information for downstream queries. Unlike traditional human-assistant settings, these environments are inherently multimodal, involve complex discourse phenomena such as anaphora and deixis, and contain asynchronous or conflicting information from multiple participants.
The paper introduces Omni Demand Understanding (ODU), a benchmark designed to test whether multimodal models can infer a user's underlying demand from complex audio‑visual interactions. ODU requires models to detect the presence of a demand and infer intent using multimodal and conversational context, evaluated across single‑turn and multi‑turn scenarios. The authors built ODU‑Bench through a taxonomy‑guided approach, agentic video generation, and human‑recorded interactions, and found that even top models like Gemini 3.1 Pro recover only 44.7% of key information, with many models exhibiting high false‑trigger rates.
By Qi Chen, Yunfei Chu, Haolin He, Yifan Yang, Zihan Liu, Yuxuan Wang, Ziyang Ma, Ruiyang Xu, Meng Gao, Yinsong Yan, Ling Wang, Hui Wang, Wen Huang, Yiheng Chen, Guanrou Yang, Qiuqiang Kong, Jin Xu, Xie Chen
arXiv:2607. 14499v1 Announce Type: new Abstract: Multi-modal Large Language Models (MLLMs) have made substantial advances on benchmarks, yet their real-world effectiveness remains uncertain.
By Yijiang Li, Huiqi Zou, Bingyang Wang, Ziang Xiao
The paper proposes an agentic framework for multimodal emotion reasoning that uses frozen models to avoid costly instruction tuning. It employs modality‑specific perception agents, an integration agent, and an MLLM critic to generate emotion predictions with evidence citations, followed by an evidence‑grounded verification stage that identifies errors and applies targeted prompt patches without updating model weights. On the MER2025 dataset, this approach improves F‑score across several large language‑vision models, outperforming static pipelines and matching or exceeding reflection baselines.
By Elyas Rashno, Farhana Zulkernine