Multimodal large language models have made substantial progress in medical visual understanding, but temporal reasoning and knowledge-guided reasoning over real surgical videos remain underexplored. Existing surgical benchmarks mainly focus on static frames, single procedures, or simple questions, which limits their ability to capture the complexity of real clinical workflows. To address this gap, we introduce UniSurgBench, a unified benchmark for open-ended surgical video reasoning, comprising 2,302 question-answer samples collected from real surgical videos spanning 45 procedures and 29 anatomical structures. Each sample is annotated with temporally grounded evidence and structured reasoning steps. We also propose SurgExpEval, a two-stage evaluation protocol that extracts structured reasoning components from free-form model outputs and jointly assesses answer accuracy, reasoning quality, and temporal grounding. Evaluations with 23 proprietary, open-source, and medical-related MLLMs show that MLLMs can produce clinically reasonable and causally coherent responses. However, they often fail to associate these responses with fine-grained visual evidence, highlighting evidence-level visual grounding as a key bottleneck for reliable surgical intelligence.

Lip-to-speech (LTS) synthesis aims to generate intelligible speech from silent lip movements, but the ambiguous visual-to-phoneme mapping makes content-accurate synthesis challenging. Existing methods mainly follow two routes: (1) direct synthesis from lip movements and (2) synthesis dependent on an external lip-reading model. The former often suffers from content errors, while the latter improves content accuracy at the cost of computational overhead and error propagation. To address these limitations, we introduce FlexLTS, a unified paradigm for content-accurate LTS with optional lip reading. It generates accurate speech without requiring lip reading, while still benefiting from lip-reading text when available. FlexLTS learns phonetic prior embeddings from training transcripts and employs a visual-centric prior fusion mechanism to selectively incorporate relevant phonetic priors into the visual stream, yielding content-enriched semantic representations. Moreover, our proposed Progressive Textual Dropout strategy further promotes prior internalization by reducing reliance on explicit text during training. Experiments show that FlexLTS achieves state-of-the-art content accuracy and speech quality on the popular LRS2-BBC and LRS3-TED benchmarks and also performs well in cross-dataset zero-shot evaluation. Notably, without using any lip-reading model, FlexLTS surpasses existing lip-reading-aided methods in content accuracy.

Despite rapid progress, Video Large Language Models (Video-LLMs) remain unreliable due to hallucinations, which are outputs that contradict either video evidence (faithfulness) or verifiable world knowledge (factuality).Existing benchmarks provide limited coverage of factuality hallucinations and predominantly evaluate models only in clean settings.We introduce INFACT, a diagnostic benchmark comprising 9,800 QA instances with fine-grained taxonomies for faithfulness and factuality, spanning real and synthetic videos.INFACT evaluates models in four modes: Base (clean), Visual Degradation, Evidence Corruption, and Temporal Intervention for order-sensitive items.Reliability under induced modes is quantified using Resist Rate (RR) and Temporal Sensitivity Score (TSS).Experiments on 14 representative Video-LLMs reveal that higher Base-mode accuracy does not reliably translate to higher reliability in the induced modes, with evidence corruption reducing stability and temporal intervention yielding the largest degradation.Notably, many open-source baselines exhibit near-zero TSS on factuality, indicating pronounced temporal inertia on order-sensitive questions.

Dense video event captioning aims to generate a sequence of descriptive captions for each event in a long untrimmed video. Video-level context provides important information and facilities the model to generate consistent and less redundant captions between events. In this paper, we introduce a novel Hierarchical Context-aware Network for dense video event captioning (HCN) to capture context from various aspects. In detail, the model leverages local and global context with different mechanisms to jointly learn to generate coherent captions. The local context module performs full interaction between neighbor frames and the global context module selectively attends to previous or future events. According to our extensive experiment on both Youcook2 and Activitynet Captioning datasets, the video-level HCN model outperforms the event-level context-agnostic model by a large margin. The code is available at https://github.com/KirkGuo/HCN.

Non-Autoregressive Transformer (NAT) aims to accelerate the Transformer model through discarding the autoregressive mechanism and generating target words independently, which fails to exploit the target sequential information. Over-translation and under-translation errors often occur for the above reason, especially in the long sentence translation scenario. In this paper, we propose two approaches to retrieve the target sequential information for NAT to enhance its translation ability while preserving the fast-decoding property. Firstly, we propose a sequence-level training method based on a novel reinforcement algorithm for NAT (Reinforce-NAT) to reduce the variance and stabilize the training procedure. Secondly, we propose an innovative Transformer decoder named FS-decoder to fuse the target sequential information into the top layer of the decoder. Experimental results on three translation tasks show that the Reinforce-NAT surpasses the baseline NAT system by a significant margin on BLEU without decelerating the decoding speed and the FS-decoder achieves comparable translation performance to the autoregressive Transformer with considerable speedup.

Neural machine translation (NMT) models are usually trained with the word-level loss using the teacher forcing algorithm, which not only evaluates the translation improperly but also suffers from exposure bias. Sequence-level training under the reinforcement framework can mitigate the problems of the word-level loss, but its performance is unstable due to the high variance of the gradient estimation. On these grounds, we present a method with a differentiable sequence-level training objective based on probabilistic n-gram matching which can avoid the reinforcement framework. In addition, this method performs greedy search in the training which uses the predicted words as context just as at inference to alleviate the problem of exposure bias. Experiment results on the NIST Chinese-to-English translation tasks show that our method significantly outperforms the reinforcement-based algorithms and achieves an improvement of 1.5 BLEU points on average over a strong baseline system.