Lightweight Structured Multimodal Reasoning for Clinical Scene Understanding in Robotics

Healthcare robotics requires robust multimodal perception and reasoning to ensure safety in dynamic clinical environments. Current Vision-Language Models (VLMs) demonstrate strong general-purpose capabilities but remain limited in temporal reasoning, uncertainty estimation, and structured outputs needed for robotic planning. We present a lightweight agentic multimodal framework for video-based scene understanding. Combining the Qwen2.5-VL-3B-Instruct model with a SmolAgent-based orchestration layer, it supports chain-of-thought reasoning, speech-vision fusion, and dynamic tool invocation. The framework generates structured scene graphs and leverages a hybrid retrieval module for interpretable and adaptive reasoning. Evaluations on the Video-MME benchmark and a custom clinical dataset show competitive accuracy and improved robustness compared to state-of-the-art VLMs, demonstrating its potential for applications in robot-assisted surgery, patient monitoring, and decision support.

Semantic-VAE: Semantic-Alignment Latent Representation for Better Speech Synthesis

While mel-spectrograms have been widely utilized as intermediate representations in zero-shot text-to-speech (TTS), their inherent redundancy leads to inefficiency in learning text-speech alignment. Compact VAE-based latent representations have recently emerged as a stronger alternative, but they also face a fundamental optimization dilemma: higher-dimensional latent spaces improve reconstruction quality and speaker similarity, but degrade intelligibility, while lower-dimensional spaces improve intelligibility at the expense of reconstruction fidelity. To overcome this dilemma, we propose Semantic-VAE, a novel VAE framework that utilizes semantic alignment regularization in the latent space. This design alleviates the reconstruction-generation trade-off by capturing semantic structure in high-dimensional latent representations. Extensive experiments demonstrate that Semantic-VAE significantly improves synthesis quality and training efficiency. When integrated into F5-TTS, our method achieves 2.10% WER and 0.64 speaker similarity on LibriSpeech-PC, outperforming mel-based systems (2.23%, 0.60) and vanilla acoustic VAE baselines (2.65%, 0.59). We also release the code and models to facilitate further research.

Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?

Natural language processing (NLP) models often replicate or amplify social bias from training data, raising concerns about fairness. At the same time, their black-box nature makes it difficult for users to recognize biased predictions and for developers to effectively mitigate them. While some studies suggest that input-based explanations can help detect and mitigate bias, others question their reliability in ensuring fairness. Existing research on explainability in fair NLP has been predominantly qualitative, with limited large-scale quantitative analysis. In this work, we conduct the first systematic study of the relationship between explainability and fairness in hate speech detection, focusing on both encoder- and decoder-only models. We examine three key dimensions: (1) identifying biased predictions, (2) selecting fair models, and (3) mitigating bias during model training. Our findings show that input-based explanations can effectively detect biased predictions and serve as useful supervision for reducing bias during training, but they are unreliable for selecting fair models among candidates.

FLEXI: Benchmarking Full-duplex Human-LLM Speech Interaction

Full-Duplex Speech-to-Speech Large Language Models (LLMs) are foundational to natural human-computer interaction, enabling real-time spoken dialogue systems. However, benchmarking and modeling these models remains a fundamental challenge. We introduce FLEXI, the first benchmark for full-duplex LLM-human spoken interaction that explicitly incorporates model interruption in emergency scenarios. FLEXI systematically evaluates the latency, quality, and conversational effectiveness of real-time dialogue through six diverse human-LLM interaction scenarios, revealing significant gaps between open source and commercial models in emergency awareness, turn terminating, and interaction latency. Finally, we suggest that next token-pair prediction offers a promising path toward achieving truly seamless and human-like full-duplex interaction.

StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs

Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks.

Pushing Toward the Simplex Vertices: A Simple Remedy for Code Collapse in Smoothed Vector Quantization

Vector quantization, which discretizes a continuous vector space into a finite set of representative vectors (a codebook), has been widely adopted in modern machine learning. Despite its effectiveness, vector quantization poses a fundamental challenge: the non-differentiable quantization step blocks gradient backpropagation. Smoothed vector quantization addresses this issue by relaxing the hard assignment of a codebook vector into a weighted combination of codebook entries, represented as the matrix product of a simplex vector and the codebook. Effective smoothing requires two properties: (1) smoothed quantizers should remain close to a onehot vector, ensuring tight approximation, and (2) all codebook entries should be utilized, preventing code collapse. Existing methods typically address these desiderata separately. By contrast, the present study introduces a simple and intuitive regularization that promotes both simultaneously by minimizing the distance between each simplex vertex and its $K$-nearest smoothed quantizers. Experiments on representative benchmarks, including discrete image autoencoding and contrastive speech representation learning, demonstrate that the proposed method achieves more reliable codebook utilization and improves performance compared to prior approaches.

EgoInstruct: An Egocentric Video Dataset of Face-to-face Instructional Interactions with Multi-modal LLM Benchmarking

Analyzing instructional interactions between an instructor and a learner who are co-present in the same physical space is a critical problem for educational support and skill transfer. Yet such face-to-face instructional scenes have not been systematically studied in computer vision. We identify two key reasons: i) the lack of suitable datasets and ii) limited analytical techniques. To address this gap, we present a new egocentric video dataset of face-to-face instruction and provide ground-truth annotations for two fundamental tasks that serve as a first step toward a comprehensive understanding of instructional interactions: procedural step segmentation and conversation-state classification. Using this dataset, we benchmark multimodal large language models (MLLMs) against conventional task-specific models. Since face-to-face instruction involves multiple modalities (speech content and prosody, gaze and body motion, and visual context), effective understanding requires methods that handle verbal and nonverbal communication in an integrated manner. Accordingly, we evaluate recently introduced MLLMs that jointly process images, audio, and text. This evaluation quantifies the extent to which current machine learning models understand face-to-face instructional scenes. In experiments, MLLMs outperform specialized baselines even without task-specific fine-tuning, suggesting their promise for holistic understanding of instructional interactions.

VoiceAssistant-Eval: Benchmarking AI Assistants across Listening, Speaking, and Viewing

The growing capabilities of large language models and multimodal systems have spurred interest in voice-first AI assistants, yet existing benchmarks are inadequate for evaluating the full range of these systems' capabilities. We introduce VoiceAssistant-Eval, a comprehensive benchmark designed to assess AI assistants across listening, speaking, and viewing. VoiceAssistant-Eval comprises 10,497 curated examples spanning 13 task categories. These tasks include natural sounds, music, and spoken dialogue for listening; multi-turn dialogue, role-play imitation, and various scenarios for speaking; and highly heterogeneous images for viewing. To demonstrate its utility, we evaluate 21 open-source models and GPT-4o-Audio, measuring the quality of the response content and speech, as well as their consistency. The results reveal three key findings: (1) proprietary models do not universally outperform open-source models; (2) most models excel at speaking tasks but lag in audio understanding; and (3) well-designed smaller models can rival much larger ones. Notably, the mid-sized Step-Audio-2-mini (7B) achieves more than double the listening accuracy of LLaMA-Omni2-32B-Bilingual. However, challenges remain: multimodal (audio plus visual) input and role-play voice imitation tasks are difficult for current models, and significant gaps persist in robustness and safety alignment. VoiceAssistant-Eval identifies these gaps and establishes a rigorous framework for evaluating and guiding the development of next-generation AI assistants. Code and data will be released at https://mathllm.github.io/VoiceAssistantEval/ .

MDAR: A Multi-scene Dynamic Audio Reasoning Benchmark

The ability to reason from audio, including speech, paralinguistic cues, environmental sounds, and music, is essential for AI agents to interact effectively in real-world scenarios. Existing benchmarks mainly focus on static or single-scene settings and do not fully capture scenarios where multiple speakers, unfolding events, and heterogeneous audio sources interact. To address these challenges, we introduce MDAR, a benchmark for evaluating models on complex, multi-scene, and dynamically evolving audio reasoning tasks. MDAR comprises 3,000 carefully curated question-answer pairs linked to diverse audio clips, covering five categories of complex reasoning and spanning three question types. We benchmark 26 state-of-the-art audio language models on MDAR and observe that they exhibit limitations in complex reasoning tasks. On single-choice questions, Qwen2.5-Omni (open-source) achieves 76.67% accuracy, whereas GPT-4o Audio (closed-source) reaches 68.47%; however, GPT-4o Audio substantially outperforms Qwen2.5-Omni on the more challenging multiple-choice and open-ended tasks. Across all three question types, no model achieves 80% performance. These findings underscore the unique challenges posed by MDAR and its value as a benchmark for advancing audio reasoning research.Code and benchmark can be found at https://github.com/luckyerr/MDAR.

SimulSense: Sense-Driven Interpreting for Efficient Simultaneous Speech Translation

How to make human-interpreter-like read/write decisions for simultaneous speech translation (SimulST) systems? Current state-of-the-art systems formulate SimulST as a multi-turn dialogue task, requiring specialized interleaved training data and relying on computationally expensive large language model (LLM) inference for decision-making. In this paper, we propose SimulSense, a novel framework for SimulST that mimics human interpreters by continuously reading input speech and triggering write decisions to produce translation when a new sense unit is perceived. Experiments against two state-of-the-art baseline systems demonstrate that our proposed method achieves a superior quality-latency tradeoff and substantially improved real-time efficiency, where its decision-making is up to 9.6x faster than the baselines.