Robust deployment of large multimodal models (LMMs) in real-world scenarios requires access to external knowledge sources, given the complexity and dynamic nature of real-world information. Existing approaches such as retrieval-augmented generation (RAG) and prompt engineered search agents rely on rigid pipelines, often leading to inefficient or excessive search behaviors. We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments. Our framework integrates both image and text search tools, allowing the model to reason about when and how to invoke them guided by an outcome-based reward with a search penalty. To support training, We collect a multimodal search VQA dataset through a semi-automated pipeline that covers diverse visual and textual knowledge needs and curate a search-balanced subset with both search-required and search-free samples, which proves essential for shaping efficient and on-demand search behavior. Extensive experiments on knowledge-intensive and info-seeking VQA tasks show that our model not only outperforms RAG-based baselines of the same model size, but also matches the performance of a larger RAG-based model while reducing search calls by over 30%. We further analyze key empirical findings to offer actionable insights for advancing research in multimodal search.

Yudong Yang, Jimin Zhuang, Guangzhi Sun, Changli Tang, Yixuan Li, Peihan Li, Yifan Jiang, Wei Li, Zejun Ma, Chao Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Event time series appear in many software scenarios and are a necessary data type in data analytics systems. Transformers are the preferred type of sequential neural network for advanced analytics on event time series, particularly due to their significant contributions to the recent surge of large language models (LLMs). Event series analytics heavily depends on the quality of input data, which may contain natural measurement errors or adversarial noises. Since the input data deviates from the true state, the opaque nature of neural networks presents a challenge in ensuring the reliability of output, which might be deemed untrustworthy. In this paper, we introduce an innovative formal verification framework for Transformer-based event series systems, leveraging sampling, linear programming, and the extreme value theorem. This framework can support the verification of the dependability of Transformers in managing inputs characterized by unpredictability and uncertainty. To exemplify its utility, we apply our verification approach to verify natural requirements from a real-world event series environments: network traffic classification. It outperforms the current state-of-the-art verifier in terms of effectiveness, providing more stringent verified bounds. Our experimental findings provide valuable benchmarks for guaranteeing reliable deployment of systems in scenarios where the credibility of event data is compromised, and for exposing specific cases in which the expected requirements are not satisfied.

Visual instruction tuning has made considerable strides in enhancing the capabilities of Large Multimodal Models (LMMs). However, existing open LMMs largely focus on single-image tasks, their applications to multi-image scenarios remains less explored. Additionally, prior LMM research separately tackles different scenarios, leaving it impossible to generalize cross scenarios with new emerging capabilities. To this end, we introduce LLaVA-Interleave, which simultaneously tackles Multi-image, Multi-frame (video), Multi-view (3D), and Multi-patch (single-image) scenarios in LMMs. To enable these capabilities, we regard the interleaved data format as a general template and compile the M4-Instruct dataset with 1,177.6k samples, spanning 4 primary domains with 14 tasks and 41 datasets. We also curate the LLaVA-Interleave Bench to comprehensively evaluate the multi-image performance of LMMs. Through extensive experiments, LLaVA-Interleave achieves leading results in multi-image, video, and 3D benchmarks, while maintaining the performance of single-image tasks. Besides, our model also exhibits several emerging capabilities, e.g., transferring tasks across different settings and modalities.

Human vision is dynamic and continuous. However, in video understanding with multimodal large language models (LLMs), existing methods primarily rely on static features extracted from images sampled at a fixed low frame rate of frame-per-second (FPS) $\leqslant$2, leading to critical visual information loss. In this paper, we introduce F-16, the first multimodal LLM designed for high-frame-rate video understanding. By increasing the frame rate to 16 FPS and compressing visual tokens within each 1-second clip, F-16 efficiently captures dynamic visual features while preserving key semantic information. Experimental results demonstrate that higher frame rates considerably enhance video understanding across multiple benchmarks, providing a new approach to improving video LLMs beyond scaling model size or training data. F-16 achieves state-of-the-art performance among 7-billion-parameter video LLMs on both general and fine-grained video understanding benchmarks, such as Video-MME and TemporalBench. Furthermore, F-16 excels in complex spatiotemporal tasks, including high-speed sports analysis (*e.g.*, basketball, football, gymnastics, and diving), outperforming SOTA proprietary visual models like GPT-4o and Gemini-1.5-pro. Additionally, we introduce a novel decoding method for F-16 that enables highly efficient low-frame-rate inference without requiring model retraining. We will release the source code, model checkpoints, and data at [https://github.com/bytedance/F-16](https://github.com/bytedance/F-16).

While recent advancements in reasoning optimization have significantly enhanced the capabilities of large language models (LLMs), existing efforts to improve reasoning have been limited to solving mathematical problems and focusing on visual graphical inputs, neglecting broader applications in general video understanding. This paper proposes video-SALMONN-o1, the first open-source reasoning-enhanced audio-visual LLM designed for general video understanding tasks. To enhance its reasoning abilities, we develop a reasoning-intensive dataset featuring challenging audio-visual questions with step-by-step solutions. We also propose process direct preference optimization (pDPO), which leverages contrastive step selection to achieve efficient step-level reward modelling tailored for multimodal inputs. Additionally, we introduce RivaBench, the first reasoning-intensive video understanding benchmark, featuring over 4,000 high-quality, expert-curated question-answer pairs across scenarios such as standup comedy, academic presentations, and synthetic video detection. video-SALMONN-o1 achieves 3-8% accuracy improvements over the LLaVA-OneVision baseline across different video reasoning benchmarks. Besides, pDPO achieves 6-8% improvements compared to the supervised fine-tuning model on RivaBench. Enhanced reasoning enables video-SALMONN-o1 zero-shot synthetic video detection capabilities.

Extending multilingual automatic speech recognition (mASR) systems to new languages poses challenges, particularly when training data for existing languages is limited or unavailable. To tackle this issue, we suggest utilizing supplementary encoder and decoder components. Specifically, we propose appending and fine-tuning a distinct decoder designed for new languages, while preserving the parameters of existing languages to minimize disruption to their performance. Furthermore, we advocate attaching an additional encoder component to enhance acoustic representation learning for new languages, resulting in substantial improvements in word error rate performance. Our experimental findings demonstrate the effectiveness of the proposed methods for the task of extending language support within mASR systems.

Multimodal large language models (LLMs) have shown promising visual perception abilities by connecting with image encoders, but their performance on auditory tasks has not yet been widely investigated. Meanwhile, automatic speech recognition (ASR) and automatic audio captioning (AAC) are often achieved with separate systems, resulting in incomplete auditory perception abilities. To fill in these gaps, in this paper, we present the first study that achieves both ASR and AAC by connecting an LLM with auditory encoders. A dual auditory encoder structure is proposed, integrating the Whisper encoder for speech and the BEATs encoder for audio events with a high temporal resolution by using a Q-Former at the window level. Experiments for ASR and AAC are performed correspondingly on the widely used LibriSpeech, GigaSpeech, WavCaps, AudioCaps, and Clotho datasets and yield promising results. In particular, state-of-the-art results are achieved on GigaSpeech, AudioCaps and Clotho. Our model is also able to caption speech and audio events simultaneously from clips with mixed speech and background audio events, which is a step towards more complete machine auditory perception.

The impressive capability and versatility of large language models (LLMs) have aroused increasing attention in automatic speech recognition (ASR), with several pioneering studies attempting to build integrated ASR models by connecting a speech encoder with an LLM. This paper presents a comparative study of three commonly used structures as connectors, including fully connected layers, multi-head cross-attention, and Q-Former. Speech encoders from the Whisper model series as well as LLMs from the Vicuna model series with different model sizes were studied. Experiments were performed on the commonly used LibriSpeech, Common Voice, and GigaSpeech datasets, where the LLMs with Q-Formers demonstrated consistent and considerable word error rate (WER) reductions over LLMs with other connector structures. Q-Former-based LLMs can generalise well to out-of-domain datasets, where 12% relative WER reductions over the Whisper baseline ASR model were achieved on the Eval2000 test set without using any in-domain training data from Switchboard. Moreover, a novel segment-level Q-Former is proposed to enable LLMs to recognise speech segments with a duration exceeding the limitation of the encoders, which results in 17% relative WER reductions over other connector structures on 90-second-long speech data.

Hearing is arguably an essential ability of artificial intelligence (AI) agents in the physical world, which refers to the perception and understanding of general auditory information consisting of at least three types of sounds: speech, audio events, and music. In this paper, we propose SALMONN, a speech audio language music open neural network, built by integrating a pre-trained text-based large language model (LLM) with speech and audio encoders into a single multimodal model. SALMONN enables the LLM to directly process and understand general audio inputs and achieve competitive performances on a number of speech and audio tasks used in training, such as automatic speech recognition and translation, auditory-information-based question answering, emotion recognition, speaker verification, and music and audio captioning etc. SALMONN also has a diverse set of emergent abilities unseen in the training, which includes but is not limited to speech translation to untrained languages, speech-based slot filling, spoken-query-based question answering, audio-based storytelling, and speech audio co-reasoning etc. The presence of cross-modal emergent abilities is studied, and a novel few-shot activation tuning approach is proposed to activate such abilities. To our knowledge, SALMONN is the first model of its type and can be regarded as a step towards AI with generic hearing abilities. The source code, model checkpoints and data are available at https://github.com/bytedance/SALMONN.

Speech understanding as an element of the more generic video understanding using audio-visual large language models (av-LLMs) is a crucial yet understudied aspect. This paper proposes video-SALMONN, a single end-to-end av-LLM for video processing, which can understand not only visual frame sequences, audio events and music, but speech as well. To obtain fine-grained temporal information required by speech understanding, while keeping efficient for other video elements, this paper proposes a novel multi-resolution causal Q-Former (MRC Q-Former) structure to connect pre-trained audio-visual encoders and the backbone large language model. Moreover, dedicated training approaches including the diversity loss and the unpaired audio-visual mixed training scheme are proposed to avoid frames or modality dominance. On the introduced audio-visual evaluation benchmark, video-SALMONN achieves more than 25% absolute accuracy improvements on the video-QA task and over 30% absolute accuracy improvements on audio-visual QA tasks with human speech. In addition, video-SALMONN demonstrates remarkable video comprehension and reasoning abilities on tasks that are unprecedented by other av-LLMs. Our training code and model checkpoints are available at https://github.com/bytedance/SALMONN/

Recent studies on pronunciation scoring have explored the effect of introducing phone embeddings as reference pronunciation, but mostly in an implicit manner, i.e., addition or concatenation of reference phone embedding and actual pronunciation of the target phone as the phone-level pronunciation quality representation. In this paper, we propose to use linguistic-acoustic similarity to explicitly measure the deviation of non-native production from its native reference for pronunciation assessment. Specifically, the deviation is first estimated by the cosine similarity between reference phone embedding and corresponding acoustic embedding. Next, a phone-level Goodness of pronunciation (GOP) pre-training stage is introduced to guide this similarity-based learning for better initialization of the aforementioned two embeddings. Finally, a transformer-based hierarchical pronunciation scorer is used to map a sequence of phone embeddings, acoustic embeddings along with their similarity measures to predict the final utterance-level score. Experimental results on the non-native databases suggest that the proposed system significantly outperforms the baselines, where the acoustic and phone embeddings are simply added or concatenated. A further examination shows that the phone embeddings learned in the proposed approach are able to capture linguistic-acoustic attributes of native pronunciation as references.

A typical fluency scoring system generally relies on an automatic speech recognition (ASR) system to obtain time stamps in input speech for the subsequent calculation of fluency-related features or directly modeling speech fluency with an end-to-end approach. This paper describes a novel ASR-free approach for automatic fluency assessment using self-supervised learning (SSL). Specifically, wav2vec2.0 is used to extract frame-level speech features, followed by K-means clustering to assign a pseudo label (cluster index) to each frame. A BLSTM-based model is trained to predict an utterance-level fluency score from frame-level SSL features and the corresponding cluster indexes. Neither speech transcription nor time stamp information is required in the proposed system. It is ASR-free and can potentially avoid the ASR errors effect in practice. Experimental results carried out on non-native English databases show that the proposed approach significantly improves the performance in the "open response" scenario as compared to previous methods and matches the recently reported performance in the "read aloud" scenario.

We propose an environment adaptation approach that improves deep speech enhancement models via minimizing the Kullback-Leibler divergence between posterior probabilities produced by a multi-condition senone classifier (teacher) fed with noisy speech features and a clean-condition senone classifier (student) fed with enhanced speech features to transfer an existing deep neural network (DNN) speech enhancer to specific noisy environments without using noisy/clean paired target waveforms needed in conventional DNN-based spectral regression. Our solution not only improves listening quality in the enhanced speech but also boosts noise robustness of existing automatic speech recognition (ASR) systems trained on clean data if employed as a pre-processing step before speech feature extraction. Experimental results show steady gains in objective quality measurements as a result of a teacher network producing adaptation targets for a student enhancement model to adjust its parameters in unseen noise conditions. The proposed technique is particularly advantageous in environments that are not handled effectively by the unadapted DNN-based enhancer, as we find that only very little data from a specific operating condition is required to yield good improvements. Finally, higher gains in speech quality directly translate to larger improvements in ASR.

In this paper, we propose a cross-modal student-teacher learning framework to make a full use of externally abundant acoustic data in addition to a given task-specific audio-visual training database for improving speech recognition performance under the low signal-to-noise-ratio (SNR) and acoustic mismatch conditions. First, a teacher model is trained with large-sized audio-only databases. Next, a student, namely a deep neural network (DNN) model, is trained on a small-sized audio-visual database to minimize the Kullback-Leibler (KL) divergence between its output and the posterior distribution of the teacher. We evaluate the proposed approach in both matched and mismatch acoustic conditions for phone recognition with the NTCD-TIMIT database. Compared to the DNN recognition system trained with the original audio-visual data only, the proposed solution reduces the phone error rate (PER) from 26.7% to 21.3% on a matched acoustic scenario. In the mismatch conditions, the PER is reduced from 47.9% to 42.9%. Moreover, we show that posteriors generated by the teacher contain environmental information, which enables our proposed student-teacher learning to work as an environmental-aware training and good PER reductions are observed in all SNR conditions.

Face detection has achieved great success with the development of convolution neural network. However, it remains a challenging problem to detect small and blurred faces in unconstrained environment. This paper presents a novel cascade single-shot face detector, named Cascade Single Shot Face Detector (CSSD), which introduces novel cascade classification and regression network in an anchor-based face detector to reject false positives and improve location accuracy. We have contributed in the following three aspects: 1) proposing a feature enchanted and scale-invariable face detection architecture to process faces with different scales; 2) regressing bounding boxes of faces in two steps with a cascade method; 3) filtering negative anchors online after anchor refinement and rebalancing puzzle negative anchors and positive anchors with rate of 3:1. As a consequence, our method achieves state-of-the-art detection performance on FDDB and WIDER FACE dataset.

We investigate the effectiveness of soft-target tone labels and sequential context information for mispronunciation detection of Mandarin lexical tones pronounced by second language (L2) learners whose first language (L1) is of European origin. In conventional approaches, prosodic information (e.g., F0 and tone posteriors extracted from trained tone models) is used to calculate goodness of pronunciation (GOP) scores or train binary classifiers to verify pronunciation correctness. We propose three techniques to improve detection of mispronunciation of Mandarin tones for non-native learners. First, we extend our tone model from a deep neural network (DNN) to a bidirectional long short-term memory (BLSTM) network in order to more accurately model the high variability of non-native tone productions and the contextual information expressed in tone-level co-articulation. Second, we characterize ambiguous pronunciations where L2 learners' tone realizations are between two canonical tone categories by relaxing hard target labels to soft targets with probabilistic transcriptions. Third, segmental tone features fed into verifiers are extracted by a BLSTM to exploit sequential context information to improve mispronunciation detection. Compared to DNN-GOP trained with hard targets, the proposed BLSTM-GOP framework trained with soft targets reduces the tones' averaged equal error rate (ERR) from 7.58% to 5.83% and the averaged area under ROC curve (AUC) is increased from 97.85% to 98.31%. By utilizing BLSTM-based verifiers the EER further decreases to 5.16%, and the AUC is increased to 98.47%.

We propose three techniques to improve mispronunciation detection of Mandarin tones of second language (L2) learners using tone-based extended recognition network (ERN). First, we extend our model from deep neural network (DNN) to bidirectionallon-short-term memory (BLSTM) in order to model tone-level co-articulation influenced by a broader temporal context (e.g., two or three consecutive Mandarin syllables). Second, we relax the hard labels to characterize the situations when a single tone class label is not enough because L2 learners' pronunciations are often between two canonical tone categories. Therefore, soft targets (a probabilistic transcription) are proposed for acoustic model training in place of conventional hard targets (one-hot targets). Third, we average tone scores produced by BLSTM models trained with hard and soft targets to seek the complementarity from modeling at the tone-target levels. Compared to our previous system based on the DNN-trained ERNs, the BLSTM-trained system with soft targets reduces the equal error rate (ERR) from 5.77% to 4.86%, and system combination decreases EER further to 4.34%, achieving a 24.78% relative error reduction.

In this paper, we utilize manner and place of articulation features and deep neural network models (DNNs) with long short-term memory (LSTM) to improve the detection performance of phonetic mispronunciations produced by second language learners. First, we show that speech attribute scores are complementary to conventional phone scores, so they can be concatenated as features to improve a baseline system based only on phone information. Next, pronunciation representation, usually calculated by frame-level averaging in a DNN, is now learned by LSTM, which directly uses sequential context information to embed a sequence of pronunciation scores into a pronunciation vector to improve the perfonnance of subsequent mispronunciation detectors. Finally, when both proposed techniques are incorporated into the baseline phone-based GOP (goodness of pronunciation) classifier system trained on the same data, the integrated system reduces the false acceptance rate (FAR) and false rejection rate (FRR) by 37.90% and 38.44% (relative), respectively, from the baseline system.

We propose the use of speech attributes, such as voicing and aspiration, to address two key research issues in computer assisted pronunciation training (CAPT) for L2 learners, namely detecting mispronunciation and providing diagnostic feedback. To improve the performance we focus on mispronunciations occurred at the segmental and sub-segmental levels. In this study, speech attributes scores are first used to measure the pronunciation quality at a sub-segmental level, such as manner and place of articulation. These speech attribute scores are integrated by neural network classifiers to generate segmental pronunciation scores. Compared with the conventional phone-based GOP (Goodness of Pronunciation) system we implement with our dataset, the proposed framework reduces the equal error rate by 8.78% relative. Moreover, it attains comparable results to phone-based classifier approach to mispronunciation detection while providing comprehensive feedback, including segmental and sub-segmental diagnostic information, to help L2 learners.

We propose a novel decision tree based framework to detect phonetic mispronunciations produced by L2 learners caused by using inaccurate speech attributes, such as manner and place of articulation. Compared with conventional score-based CAPT (computer assisted pronunciation training) systems, our proposed framework has three advantages: (1) each mispronunciation in a tree can be interpreted and communicated to the L2 learners by traversing the corresponding path from a leaf node to the root node; (2) corrective feedback based on speech attribute features, which are directly used to describe how consonants and vowels are produced using related articulators, can be provided to the L2 learners; and (3) by building the phone-dependent decision tree, the relative importance of the speech attribute features of a target phone can be automatically learned and used to distinguish itself from other phones. This information can provide L2 learners speech attribute feedback that is ranked in order of importance. In addition to the abovementioned advantages, experimental results confirm that the proposed approach can detect most pronunciation errors and provide accurate diagnostic feedback