The paper introduces a steering‑vector‑based causal attribution framework to study how large vision‑language models (LVLMs) translate visual input into emotional narratives. By creating a specialized dataset, the authors uncover a functional decoupling in the LVLM’s three‑stage Adapt‑Aggregate‑Execute mechanism: visual emotional cues are first aggregated in middle layers via sentiment‑specific attention heads, then translated into narrative generation in deeper layers through emotion‑general pathways. Using these insights, they regulate emotional information routing to strengthen attention flow and amplify semantic activation, achieving significant performance gains on the MER‑UniBench and reducing emotional hallucinations through inference‑time intervention.
By Chengsheng Zhang, Chenghao Sun, Zhining Xie, Xinmei Tian
The paper evaluates how different design choices—document feeding strategy, retrieval method, and execution mode—affect Vision‑Language Models (VLMs) on long‑document question answering. Experiments on two benchmarks show that a multi‑tool agent only outperforms static input when the VLM is large, that retrieval modality (image vs text) is more critical than the specific retriever, and that combining the best pipelines per question can significantly boost performance. The study highlights the trade‑offs between token efficiency, model size, and pipeline complexity for deploying VLMs on complex documents.
By Kenan E. Ak, Jay Mohta, Gwang Gook Lee, Yan Xu, Dimitrios Dimitriadis
World Action Agent (WAA) is a multi‑agent framework that lets vision‑language models (VLMs) directly pilot robots by operating within a visual action workspace. The workspace provides automatically selected contact views, editable action rehearsals, and in‑view correction to refine decisions before low‑level execution. WAA learns procedural skills from expert videos and human teaching, and its interaction traces can train smaller VLMs, achieving state‑of‑the‑art success on LIBERO‑Pro and improving out‑of‑domain performance on robosuite and Qwen3.5‑9B.
By Yehang Zhang, Haojian Huang, Yifan Chang, Jianchong Su, Bohan Zhou, Yingjie Xu, Wosong Chen, Tianhao Zhou, Chenxu Wang, Tianyi Zhang, Yangkai Wei, Wenqian Li, Shiyuan Deng, Yinchuan Li, Ying-Cong Chen, Zexi Li
The paper introduces FlipDir, a training‑free inference‑time technique that mitigates answer flips in vision‑language models by steering hidden states along a low‑rank subspace derived from contrastive image pairs. It employs a margin‑based gate to attenuate steering only during uncertain decoding steps, thereby restoring original predictions while keeping stable ones unchanged. The authors also present VisFlip, a benchmark framework that evaluates models across nine dataset‑variation combinations in scientific reasoning, robot‑scene understanding, and medical VQA, showing that FlipDir consistently outperforms existing methods on recovery and preservation metrics.
By Yeonsung Jung, Joonhyun Jeong, Hoang Pham, Joowon Kim, Yoonsik Park, Viet Dac Lai, Eunho Yang
AV‑GRPO introduces a modality‑anchored diffusion reinforcement learning framework for joint audio‑video generation, addressing limitations in fidelity, text‑modality alignment, and cross‑modal synchronization. It decouples learning signals through modality‑anchored rollouts, employs trajectory‑locked frozen‑tower optimization to reduce computational cost, and adapts objectives to each modality’s dynamics. The accompanying 5DAV dataset provides difficulty‑controllable, decoupled training samples, and experiments on JavisBench and VABench show AV‑GRPO surpasses LTX‑2.3 in generation quality, semantic alignment, and synchronization.
By Zhiyu Xu, Weilong Yan, Yufei Shi, Shiyang Li, Yihao Liu, Kin-Man Lam, Yuewen Cao
AgenticCADedit introduces a stateful, tool‑mediated approach to multimodal 3D CAD editing, transforming the process from generating a single complete program to executing a sequence of incremental, verifiable actions on a persistent CAD state. By committing each step, inspecting geometry, and selectively reverting faulty operations, the method preserves partial progress and builds upon earlier edits. Experiments across three large language models show substantial gains in validity and acceptance, with the weakest baseline model’s validity rising from 51.0% to 94.8% and a token‑cost reduction of 66.7% compared to neuralCAD‑Edit.
By Saptarshi Neil Sinha, Mika Silvan Goschke, Paul Julius K\"uhn, Arjan Kuijper, Michael Weinmann
arXiv:2503. 02781v3 Announce Type: replace-cross Abstract: Predicting clinical outcomes from preclinical data is essential for selecting safe and effective drug combinations and for reducing late-stage failures.
By Yepeng Huang, Xiaorui Su, Varun Ullanat, Intae Moon, Ivy Liang, Lindsay Clegg, Damilola Olabode, Ruthie Johnson, Nicholas Ho, Megan Gibbs, Alexander Gusev, Bino John, Marinka Zitnik
The paper investigates task collapse—a failure mode where online RL fine‑tuning of a pretrained flow‑matching vision‑language‑action policy erodes performance on individual tasks—using a 450M‑parameter SmolVLA policy on LIBERO‑10. Three exploration‑noise strategies are compared: a fixed noise scale, a learned noise network, and an uncertainty‑gated controller that reallocates exploration based on novelty and competence signals without task labels. The uncertainty‑gated controller prevents task collapse across all tested seeds, whereas the other two approaches consistently cause collapse, demonstrating its effectiveness in preserving task performance during fine‑tuning.
By Mehmet Turan Yard{\i}mc{\i}, Yunus Emre \c{C}o\u{g}urcu
Multimodal Routing and Region Refinement for Language-Guided Medical Image Segmentation (MRSeg) is a parameter‑efficient framework that uses frozen ConvNeXt‑Tiny and PubMedBERT encoders to extract multiscale visual features and clinical text tokens. A joint router predicts a sparse mixture over low‑rank adapter bases, enabling separate adaptation for two visual scales and text while keeping feature‑specific parameters distinct. Region Bridge aggregates dense visual tokens into latent regions using text‑derived queries, refines them via self‑attention and text cross‑attention, and redistributes the refined information back to the feature maps, culminating in a multiscale decoder that combines refined semantic features with shallow image evidence. MRSeg achieves state‑of‑the‑art Dice/mIoU scores on QaTa‑COV19 and MosMedData+ with only 7.11 M trainable parameters and 7.60 GFLOPs.
By Md Maklachur Rahman, Md Hasan Al Banna, Saraf Anjum, Assame Arnob, Tracy Hammond
The paper evaluates post‑training quantization (PTQ) for text‑to‑speech (TTS) models across multiple architectures using a unified protocol. It shows that reducing weights to 4‑bit per‑channel can significantly lower predicted mean opinion scores (UTMOS) and that even 8‑bit per‑tensor scaling can cause severe degradation, with the impact varying by model. A staged ablation identifies the sensitive components, and per‑layer GPTQ can recover performance to within 0.1 UTMOS, while real int8 and int4 kernels confirm the simulated results on hardware, demonstrating that each configuration must be validated on the target runtime.
By Se Un Park, Yutae Kim, Junyoung Park
The paper investigates why vision‑language models that tokenize images with vector‑quantized (VQ) codebooks frequently hallucinate objects on grounded yes/no tasks. By applying activation patching across 25 models from eight large‑language‑model families, the authors uncover an early‑layer attention routing circuit shared by VQ‑tokenized VLMs. They develop a three‑gate diagnostic that isolates ten models carrying this circuit, show that swapping a single architectural component (VQ+Linear) introduces the circuit, and demonstrate that ablating the early‑layer ($L_0$) component reduces hallucinations in open‑ended generation by 31 % while other decoding‑time fixes do not.
By Shamanthak Hegde, Xiangrui Liu, Maitreya Patel, Yezhou Yang
The paper investigates whether layer-wise visual‑text similarity in multimodal large language models (MLLMs) truly reflects content‑level cross‑modal interaction. By injecting Gaussian noise into the visual stream of 13 MLLMs, the authors show that task accuracy drops sharply while traditional scalar alignment metrics (CKA, SVCCA, MIR, principal‑angle cosine) fail to distinguish corrupted from clean inputs, a phenomenon they term the alignment illusion. They propose the principal‑angle gap (PA gap) as a more reliable geometric diagnostic that correlates better with task performance and reveals when internal geometry diverges from accuracy.
By Hong-Han Wang, Yuntao Wang, Hu Ding
GOMA (Graph-Optimized Multimodal Alignment) introduces a dual-embedding approach for multimodal retrieval, separating content embeddings supervised for paired identity from semantic embeddings trained with cross-modal pairs and observed relationships. The method fuses these embeddings, applies semantic agreement to weight graph edges, and uses restart graph propagation to reinforce the initial signal, enabling both single-modality and dual-attribute retrieval. Across six datasets and four tasks, GOMA outperforms 14 external methods on 14 primary metrics, with controlled experiments highlighting the impact of separate supervision, graph regularization, and semantic-guided propagation.
By Xu Wang, Xunkai Li, Yinlin Zhu, Rong-Hua Li, Guoren Wang
The paper investigates how post‑training compression techniques—such as pruning, quantization, and distillation—affect demographic fairness in Whisper speech‑recognition models. It finds that pruning and INT4 quantization significantly widen word‑error‑rate gaps between demographic groups, especially for Black/AA and Asian speakers, while distillation tends to reduce these gaps. The study introduces a temporal‑taxation metric to quantify the increased correction effort required for marginalized speakers after compression.
By Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
BanglaTurn is a new corpus of 35,374 Bangla podcast speech samples, each 3 to 15 seconds long, labeled for end‑of‑turn detection through speaker diarization, an LLM pass, and human verification. A Whisper‑based model with task‑specific classification heads achieves 84.33 % accuracy on a balanced test set, outperforming the Smart‑Turn v3 baseline (69.28 %) and reducing the false‑negative rate from 51.57 % to 7.55 %, though with a higher false‑positive rate. The study also details the contributions of encoder‑layer fine‑tuning, multi‑scale pooling, INT8 quantization, and reports inference latency of 165–191 ms on CPU.
By Mizbaul Haque Maruf
The paper investigates whether the Voxtral audio‑language model can detect speech spoofing. It shows that without task‑specific adaptation, the model’s language‑model layers prioritize semantic content, making spoof‑discriminative acoustic cues less separable. By applying lightweight weight‑decomposed low‑rank adaptation (DoRA), the authors create Spooftral, which achieves an equal error rate of 4.25% on the ASVspoof5 evaluation set.
By Avishai Weizman, Yehuda Ben-Shimol, Itshak Lapidot
The paper introduces a method that allows automatic speech recognition systems to learn new words during test time using unlabeled data. It combines a frozen CTC acoustic model for spellings, a frozen language model for detecting out‑of‑vocabulary words, and an adaptation module that expands the vocabulary by learning lexical token representations from CTC-generated candidates. Experiments on LibriSpeech and dysarthric speech data show relative character‑error‑rate reductions of up to 14.97% and 6.67% for recurring OOV words, respectively.
By Mengqi Wang, Mark A. Hasegawa-Johnson, Haolong Zheng, Chang D. Yoo
STRAND is a new benchmark that tests multimodal large language models’ ability to track objects, their states, and relationships over time in videos. It evaluates intermediate reasoning by breaking queries into sub‑questions and uses Faithful Accuracy to ensure all parts of an answer are correct. The authors also propose an object‑centric framework that builds structured trajectories and shows reduced hallucinations and better temporal consistency compared to existing models.
By Thong Nguyen, Tri Cao, Khoi Le, Cong-Duy Nguyen, Quynh Vo, See-Kiong Ng, Bryan Hooi Kuen-Yew
The paper introduces an LLM-assisted system for correcting speaker attribution errors during meetings. It combines streaming ASR, diarization, and concise LLM-generated summaries to guide users in providing brief corrective feedback, which updates the transcript and adds online speaker enrollments. The approach includes mechanisms to accurately interpret user corrections and a simulation for large-scale evaluation, achieving significant reductions in DER and speaker substitution error on the AMI headset test set, with a pilot usability study highlighting further improvements.
By Xinlu He, Yiwen Guan, Badrivishal Paurana, Pitipat Kongsomjit, Zilin Dai, Jacob Whitehill
The paper introduces a cause-aware error recovery framework for cascaded Automatic Speech Recognition – Large Language Model (ASR‑LLM) pipelines in Spoken Dialogue Systems. It replaces simple ASR confidence filtering with precision‑focused detectors that use deep ASR latent representations to classify token‑level errors into perception, comprehension, and deletion failures. This fine‑grained diagnosis enables the LLM to execute targeted, multi‑turn clarification strategies, leading to a more than two‑fold increase in recall on domain‑shift errors and significant reductions in word error rate and downstream task errors across varied accents, distortions, and domains.
By Yizhou Peng, Ziyang Ma, Changsong Liu, Yi-Wen Chao, Xie Chen, Eng Siong Chng