OmniHallu is a unified framework for detecting hallucinations in multimodal large language models across both comprehension and generation tasks involving image, video, and audio modalities. It introduces OmniHallu-Bench, a 10,000-sample benchmark with claim-level human annotations for six cross-modal tasks (I2T, V2T, A2T, T2I, T2V, T2A). The system uses a multi‑agent architecture that decomposes outputs into atomic claims, verifies them with modality‑specific experts, and aggregates evidence through structured reasoning, while a preference‑optimized verifier reduces expert calls by 66% with minimal performance loss.
By Jianjiang Yang, Peihang Li, Shanqing Xu, Mengchen Qian, Lu Zhang, Meng Luo
The paper presents SEAR, a system for Task 2 of the second Multilingual Conversational Speech Language Model Challenge. It adapts Qwen3-Omni-30B-A3B-Instruct with a segment‑evidence‑aware pipeline that converts timestamped ASR into event spans, expands and crops them, and then generates both semantic and acoustic multiple‑choice questions using Qwen3.6-27B and Gemini 3.1 Flash‑Lite. After rigorous checks, 359,825 verified MCQs across 21 languages and accents are produced, and the system achieves 90.92% accuracy on the official evaluation set.
By Huy Hoang Le, Long-Bao Nguyen, Minh Tri Dao
SWRouter is a new routing method for multi‑turn large language model conversations that uses a similarity‑based context segmentation mechanism to construct prompts and a dual‑metric evaluation framework to separate construction accuracy from router performance. The approach addresses two key challenges in multi‑turn dialogue: preventing information loss or confusion during context construction and evaluating routing quality independently of prompt quality. Experiments on multi‑turn dialogue benchmarks show that SWRouter outperforms strong baselines, improving evaluation accuracy by 16.26% over the best individual large language model and by 8.22% over the Conv‑ID Context baseline.
By Yu Wang, Yuchen Li, Rui Kong, Xinran Chen, Jiamin Chen, Hengyi Cai, Shuaiqiang Wang, Jiashu Zhao, Yulun Zhang, Zhonghao Lyu, Haoyi Xiong, Linghe Kong, Jimmy Xiangji Huang, Dawei Yin
The Eloquence team presents three methods for the Interspeech 2026 MLC‑SLM Task 2, a multilingual MCQA challenge covering 21 languages. They fine‑tune Voxtral‑Mini‑3B with LoRA and data augmentation, achieving 0.72 macro‑accuracy; they use multimodal in‑context learning on Voxtral‑24B to correct label bias, reaching 0.81; and they deploy a training‑free retrieval system with a voice‑anchored memory, scoring 0.68. All approaches surpass the official baseline.
By Jordi Luque, Lorenzo Concina, Marco Matassoni, Alessio Brutti, Filippo Vella
The article examines how medical research lags behind the rapid evolution of large language models (LLMs). From January 2023 to June 2026, PubMed records in fourteen clinical domains grew 45‑fold, yet only 2.5 % employed randomized, controlled, or prospective designs. The evaluation gap widened from 1.33 to 6.08 quarters, with randomized trials assessing models that were on average 4.6 quarters older than other studies, and 62 % of such trials evaluated discontinued model families.
By Raad Bin Tareaf, Murad Al-Rajab, Samia Loucif
IndicTriMix presents a new benchmark and models for token‑level language identification in tri‑language code‑mixed text involving Hindi, Gujarati, and Bengali. The authors reformulate the task as sequence labeling and fine‑tune transformer models MuRIL and XLM‑RoBERTa, evaluating them on manually annotated test sets. They also introduce two code‑mixing generation methods using parallel sentences and release the datasets and fine‑tuned models for public use.
By Pruthwik Mishra, Rudra Trivedi, Avi Patel, Ashok Urlana, Shrikant Malviya
Nuha‑Speech is a new initiative aimed at creating general‑purpose Arabic speech‑large language models (speech‑LLMs). It includes the construction of a large Arabic Speech Question‑Answering corpus with over 1.5 million samples for instruction tuning, supervised fine‑tuning of Qwen‑Omni model variants at various scales, and a systematic evaluation framework with diverse tasks and tailored metrics. The project seeks to establish foundational infrastructure for Arabic speech‑LLMs amid limited Arabic speech resources.
By Yingzhi Wang, Reem Alhazzani, Muhammad Alqurishi
The paper presents the KuaiRP series of role‑playing models, detailing a multi‑stage training pipeline that balances deep domain knowledge injection with the preservation of general agent capabilities. The approach includes a standardized character template, a supervised fine‑tuning (SFT) data pipeline, a rule‑based reward function for reinforcement learning, and a novel two‑stage on‑policy distillation (OPD) with cumulative‑divergence decay (CDD) to recover general skills. Experimental results show that the models achieve state‑of‑the‑art role‑playing fidelity in target domains while maintaining low deployment costs.
By Yipeng Wang, Ziwei Zhang, Jiahui Zhang, Qi Gan, Kai Sheng
RetroThinker is a multi-stage post‑training framework that enhances SpeechLLMs by enabling them to self‑verify and forward‑correct Chain‑of‑Thought reasoning steps during inference. It combines supervised fine‑tuning on curated retrospective thinking data with length‑based direct preference optimization to improve reasoning while the user speaks. On the GSM8K benchmark, RetroThinker achieves an 11% absolute accuracy gain over non‑retrospective baselines while maintaining comparable latency.
By Yi-Jen Shih, Puyuan Peng, Abdelrahman Mohamed, David Harwath
The paper surveys how large language models (LLMs) are being applied in legal tasks such as judgement prediction, document analysis, and drafting. It reviews the benefits of automation while highlighting legal challenges like privacy, bias, and explainability. The authors also discuss data resources for legal domain specialization and outline future research directions.
By Zhongxiang Sun
The paper explores using a large language model (LLM) to generate background context for social media posts and tests four ways to integrate this context into a Sentence-BERT (SBERT) hate‑speech detection classifier. The methods include text concatenation, embedding concatenation, hierarchical transformer fusion, and LLM‑driven text enhancement. Experiments on the Latent Hatred dataset of implicitly hateful tweets and the MAMI dataset of misogynous memes show that adding generated context can raise F1 scores by up to 3 points in textual and 6 points in multimodal settings compared to a zero‑context baseline, with embedding concatenation yielding the best results.
By Joshua Wolfe Brook, Ilia Markov
Deep Research Bench II is a new benchmark designed to evaluate Deep Research Agents (DRAs) by requiring them to produce research reports for 132 grounded tasks across 22 domains. Each report is assessed using 9,430 fine‑grained binary rubrics that cover information recall, analysis, and presentation, all derived from expert‑written investigative articles through a rigorous LLM‑plus‑human pipeline. Evaluation of current state‑of‑the‑art DRAs shows that even the best models satisfy fewer than 50% of these rubrics, highlighting a significant gap between automated agents and human experts.
By Ruizhe Li, Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, Zhendong Mao
The paper introduces CoCoEval, a framework for evaluating large language model (LLM)–simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level. Using CoCoEval, the authors compare human conversations with those generated by GPT‑4.1, GPT‑5.1, and Claude Opus 4, finding that LLMs produce far fewer such behaviors under vanilla prompting and that prompt engineering or fine‑tuning often over‑produces specific behaviors. The study highlights gaps between human and LLM‑simulated interactions that conventional Likert‑scale evaluations miss, raising concerns about using LLMs as proxies for human social interaction.
By Ryo Kamoi, Ameya Godbole, Binglin Zhou, Xiaoxin Lu, Longqi Yang, Rui Zhang, Mengting Wan, Pei Zhou
The study examined whether brain-language model alignment reflects shared computational mechanisms or merely stable lexical‑semantic correspondences. Using whole‑brain encoding across Mandarin, English, and French, transformer representations predicted activity in a distributed network that overlapped across languages and remained stable across layers. Contextual embeddings and measures of prediction or compression did not outperform static lexical embeddings, suggesting that alignment is robust but not informative about shared computational processes.
By Ni Yang, Rui He, Philipp Homan, Iris Sommer, Davide Staub, Wolfram Hinzen
The paper presents a detailed examination of narrative elements—agency, setting, and events—within the Dolma web-scale pretraining corpus. Using a framework of 11 interpretable dimensions, the authors hand‑annotated 400 passages, expanded this to a 25,000‑passage LLM‑labeled dataset, and trained NarraBERT models to predict narrative features across 13 million passages, producing the NarraDolma dataset. The study reveals that narrative structure is measurable at scale and that narrative qualities vary unevenly across different data sources, topics, and formats, highlighting gaps in current data curation practices.
By Teagan Johnson, Elliott Ash, Andrew Piper, Maria Antoniak
The paper introduces MERIT, a training‑free agent that uses an online dual‑polarity memory of oracle‑verified corrections and failed attempts to improve LLM agent repair without updating parameters. MERIT retrieves only finalized memories, classifies failures, and employs a hybrid lexical‑dense retriever before the frozen model generates revisions. Experiments with Qwen2.5‑7B‑Instruct show that MERIT raises execution accuracy on the Spider benchmark from 66.34 % to 69.79 % and on BIRD from 47.35 % to 48.44 %, with varying evidence across datasets.
By Khang Nhat Hoang Vo, Tam Minh Chu, Anh Trac Duc Dinh, Thuyen Vinh Ha Bui, Tho Quan
HiPerViT is a compact vision-only architecture that injects an explicit second-order statistical prior into a transformer-based pipeline for texture recognition. It combines global and local image views with a compact bilinear descriptor encoded as a statistical token, and integrates this token with first-order spatial representations through Perceiver-style latent distillation. Across six texture recognition benchmarks, HiPerViT consistently outperforms strong vision-only baselines, achieving notable gains on DTD, GTOS-Mobile, and 1200Tex, and the improvements are largely independent of backbone depth or fusion topology.
By Jo\~ao Pedro C. A. de S\'a, Odemir Martinez Bruno
SCINTILLA‑SNN is a 3D spiking neural network designed to predict perineural invasion (PNI) in cholangiocarcinoma (CCA) from MRI scans. It uses a four‑stage hierarchical backbone and a Multi‑Scale Spike Aggregation module to focus on sparse, localized PNI cues while reducing computational load. In a 10‑year retrospective study of 182 patients, the model achieved an AUROC of 0.748 and cut inference energy by over 23× compared to dense MAC‑only approaches.
By Youngung Han, Yului Jeong, Kyeonghun Kim, Dohyun Kweon, Suah Park, Hyunsu Go, Sungha Park, Anna Jung, Jinyong Jun, Yunho Choe, Yunjin Seo, Ken Ying-Kai Liao, Hyuk-Jae Lee, Nam-Joon Kim
R4Tun is a large language model–driven framework that adapts an expert-designed segmentation pipeline (SAM4Tun) for segmental tunnel lining inspection from 3D point clouds. By tuning bounded parameters guided by structured context—memory, state, and knowledge—it improves mean Intersection-over-Union from 0.18 to 0.43–0.48 and overall accuracy from 0.42 to 0.59–0.65 across 30 tunnel subsets and three LLMs, with the best regular subsets achieving mIoU up to 0.796. The method consistently adjusts the same critical parameters across runs, offering a controlled, label‑free, cross‑LLM adaptation mechanism with auditable rationales.
By Xinghui Tao, Zehao Ye, Guangming Wang, Jelena Nini\'c, Brian Sheil
OmniKVQuant introduces a training‑free framework for quantizing the key‑value (KV) cache of omni‑modal large language models (Omni‑LLMs) that process audio, video, and text simultaneously. The method addresses two identified problems—temporal key drift and heterogeneous value geometry—by setting key quantization ranges over short input windows and rotating values separately for each modality. Applied to Qwen2.5‑Omni and Qwen3‑Omni, OmniKVQuant achieves 2‑bit KV caches while largely preserving performance across seven audio‑visual benchmarks, and includes a fused Triton decode kernel that eliminates the need for a dense FP16 cache.
By Suho Yoo, Hyunjong Ok, Jongmin Choi, Jihoo Jung, Joon Son Chung