arXiv subjects

Kun Su

Publications and source records attributed to Kun Su.

At least 19 recordsLinked to original sources

arXiv2026

REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent

Post-training quantization (PTQ) is essential for deploying large language models (LLMs) under strict resource constraints. State-of-the-art PTQ methods quantize each layer with a single closed-form second-order solver: to remain analytically tractable, they heavily approximate the global loss (dropping cross-channel coupling, pooling output rows into groups), and they then freeze the resulting Hessian across the entire layer, with no way to refresh it as the loss landscape shifts column by column--a phenomenon we call information misalignment. We propose REAL-Q (Real-time E2E-loss Aligned LLM Quantization), a novel PTQ paradigm that breaks this compromise: instead of diluting the objective for the sake of analytic tractability, REAL-Q targets an end-to-end-aligned surrogate of the global loss and refines it via fine-grained, dynamic Block-wise Gradient Descent applied after every column block (128 columns). By coupling this fine-grained correction with a sliding window mechanism for smooth cross-layer transitions, REAL-Q effectively mitigates error propagation across the network. On LLaMA-3.1 (8B and 70B) and Qwen3 (0.6B-32B) at W4A16, REAL-Q reduces end-to-end KL divergence by up to ~49% relative to state-of-the-art globally-guided methods.

cs.LG↗

arXiv2026

Code-on-Graph: Iterative Programmatic Reasoning via Large Language Models on Knowledge Graphs

Knowledge Graphs (KGs) are widely used to mitigate the limitations of Large Language Models (LLMs), such as outdated knowledge and hallucinations. Existing LLM-KG integration frameworks typically rely on predefined operators to retrieve factual knowledge from KGs and inject it into prompts for answer generation. This paradigm faces two critical bottlenecks: 1) Inflexibility: The predefined operators are limited in scope and thus lack sufficient compositional expressiveness to fully capture the complex semantics required by KG questions. 2) Unscalability: Direct injection of factual knowledge into prompts limits scalability in handling large-scale factual knowledge. To address these two bottlenecks, we propose Code-on-Graph (CoG), a programmatic reasoning framework for LLM-KG integration. Specifically, given the factual knowledge retrieved at each reasoning step, CoG first identifies the corresponding KG schemas and represents these schemas as Python classes, which serve as abstract interfaces to the retrieved facts. It then generates executable code grounded in these classes, with the retrieved facts instantiated as objects of the corresponding classes during execution. This design enables flexible code-based reasoning while avoiding the direct injection of large-scale factual knowledge into prompts. Experiments on WebQSP, CWQ, and GrailQA demonstrate that CoG outperforms prior state-of-the-art models by up to 10.5%.

cs.AI↗

arXiv2026

Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion

Many modern retrieval problems are set-valued: given a broad intent, the system must return a collection of results that optimizes higher-order properties (e.g., diversity, coverage, complementarity, coherence) while remaining grounded with respect to a fixed database. Set-valued objectives are typically non-decomposable and are not captured by existing supervised (query, content) datasets which only prioritize top-1 retrieval. Consequently, fan-out retrieval is often employed to generate diverse subqueries to retrieve item sets. While reinforcement learning (RL) can optimize set-level objectives via interaction, deploying an RL-tuned LLM for fan-out retrieval is prohibitively expensive at inference time. Conversely, diffusion-based generative retrieval enables efficient single-pass fan-out in embedding space, but requires objective-aligned training targets. To address these issues, we propose R4T (Retrieve-for-Train), which uses RL once as an objective transducer in a three-step process: (i) train a fan-out LLM with composite set-level rewards, (ii) synthesize objective-consistent training pairs, and (iii) train a lightweight diffusion retriever to model the conditional distribution of set-valued outputs. Across large-scale fashion and music benchmarks consisting of curated item sets, we show that R4T improves retrieval quality relative to strong baselines while reducing query-time fan-out latency by an order of magnitude.

cs.IR↗

arXiv2025

Efficient Item ID Generation for Large-Scale LLM-based Recommendation

Integrating product catalogs and user behavior into LLMs can enhance recommendations with broad world knowledge, but the scale of real-world item catalogs, often containing millions of discrete item identifiers (Item IDs), poses a significant challenge. This contrasts with the smaller, tokenized text vocabularies typically used in LLMs. The predominant view within the LLM-based recommendation literature is that it is infeasible to treat item ids as a first class citizen in the LLM and instead some sort of tokenization of an item into multiple tokens is required. However, this creates a key practical bottleneck in serving these models for real-time low-latency applications. Our paper challenges this predominant practice and integrates item ids as first class citizens into the LLM. We provide simple, yet highly effective, novel training and inference modifications that enable single-token representations of items and single-step decoding. Our method shows improvements in recommendation quality (Recall and NDCG) over existing techniques on the Amazon shopping datasets while significantly improving inference efficiency by 5x-14x. Our work offers an efficiency perspective distinct from that of other popular approaches within LLM-based recommendation, potentially inspiring further research and opening up a new direction for integrating IDs into LLMs. Our code is available here https://drive.google.com/file/d/1cUMj37rV0Z1bCWMdhQ6i4q4eTRQLURtC

cs.IR↗

arXiv2025

Generative Modeling for Adversarial Lane-Change Scenarios

Decision-making in long-tail scenarios is pivotal to autonomous-driving development, and realistic and challenging simulations play a crucial role in testing safety-critical situations. However, existing open-source datasets lack systematic coverage of long-tail scenes, and lane-change maneuvers being emblematic, rendering such data exceedingly scarce. To bridge this gap, we introduce a data mining framework that exhaustively analyzes two widely used datasets, NGSIM and INTERACTION, to identify sequences marked by hazardous behavior, thereby replenishing these neglected scenarios. Using Generative Adversarial Imitation Learning (GAIL) enhanced with Proximal Policy Optimization (PPO), and enriched by vehicular-environment interaction analytics, our method iteratively refines and parameterizes newly generated trajectories. Distinguished by a rationally adversarial and sensitivity-aware perspective, the approach optimizes the creation of challenging scenes. Experiments show that, compared to unfiltered data and baseline models, our method produces behaviors that are simultaneously both adversarial and natural, judged by collision frequency, acceleration profiles, and lane-change dynamics, offering constructive insights to amplifying long-tailed lane-change instances in datasets and advancing decision-making training.

cs.RO↗

arXiv2025

REGEN: A Dataset and Benchmarks with Natural Language Critiques and Narratives

This paper introduces a novel dataset REGEN (Reviews Enhanced with GEnerative Narratives), designed to benchmark the conversational capabilities of recommender Large Language Models (LLMs), addressing the limitations of existing datasets that primarily focus on sequential item prediction. REGEN extends the Amazon Product Reviews dataset by inpainting two key natural language features: (1) user critiques, representing user "steering" queries that lead to the selection of a subsequent item, and (2) narratives, rich textual outputs associated with each recommended item taking into account prior context. The narratives include product endorsements, purchase explanations, and summaries of user preferences. Further, we establish an end-to-end modeling benchmark for the task of conversational recommendation, where models are trained to generate both recommendations and corresponding narratives conditioned on user history (items and critiques). For this joint task, we introduce a modeling framework LUMEN (LLM-based Unified Multi-task Model with Critiques, Recommendations, and Narratives) which uses an LLM as a backbone for critiquing, retrieval and generation. We also evaluate the dataset's quality using standard auto-rating techniques and benchmark it by training both traditional and LLM-based recommender models. Our results demonstrate that incorporating critiques enhances recommendation quality by enabling the recommender to learn language understanding and integrate it with recommendation signals. Furthermore, LLMs trained on our dataset effectively generate both recommendations and contextual narratives, achieving performance comparable to state-of-the-art recommenders and language models.

cs.CL↗

arXiv2024

Diff4Steer: Steerable Diffusion Prior for Generative Music Retrieval with Semantic Guidance

Modern music retrieval systems often rely on fixed representations of user preferences, limiting their ability to capture users' diverse and uncertain retrieval needs. To address this limitation, we introduce Diff4Steer, a novel generative retrieval framework that employs lightweight diffusion models to synthesize diverse seed embeddings from user queries that represent potential directions for music exploration. Unlike deterministic methods that map user query to a single point in embedding space, Diff4Steer provides a statistical prior on the target modality (audio) for retrieval, effectively capturing the uncertainty and multi-faceted nature of user preferences. Furthermore, Diff4Steer can be steered by image or text inputs, enabling more flexible and controllable music discovery combined with nearest neighbor search. Our framework outperforms deterministic regression methods and LLM-based generative retrieval baseline in terms of retrieval and ranking metrics, demonstrating its effectiveness in capturing user preferences, leading to more diverse and relevant recommendations. Listening examples are available at tinyurl.com/diff4steer.

cs.SD↗

arXiv2024

Tell What You Hear From What You See -- Video to Audio Generation Through Text

The content of visual and audio scenes is multi-faceted such that a video can be paired with various audio and vice-versa. Thereby, in video-to-audio generation task, it is imperative to introduce steering approaches for controlling the generated audio. While Video-to-Audio generation is a well-established generative task, existing methods lack such controllability. In this work, we propose VATT, a multi-modal generative framework that takes a video and an optional text prompt as input, and generates audio and optional textual description of the audio. Such a framework has two advantages: i) Video-to-Audio generation process can be refined and controlled via text which complements the context of visual information, and ii) The model can suggest what audio to generate for the video by generating audio captions. VATT consists of two key modules: VATT Converter, a LLM that is fine-tuned for instructions and includes a projection layer that maps video features to the LLM vector space; and VATT Audio, a transformer that generates audio tokens from visual frames and from optional text prompt using iterative parallel decoding. The audio tokens are converted to a waveform by pretrained neural codec. Experiments show that when VATT is compared to existing video-to-audio generation methods in objective metrics, it achieves competitive performance when the audio caption is not provided. When the audio caption is provided as a prompt, VATT achieves even more refined performance (lowest KLD score of 1.41). Furthermore, subjective studies show that VATT Audio has been chosen as preferred generated audio than audio generated by existing methods. VATT enables controllable video-to-audio generation through text as well as suggesting text prompts for videos through audio captions, unlocking novel applications such as text-guided video-to-audio generation and video-to-audio captioning.

cs.CV↗

arXiv2024

Beyond Retrieval: Generating Narratives in Conversational Recommender Systems

The recent advances in Large Language Model's generation and reasoning capabilities present an opportunity to develop truly conversational recommendation systems. However, effectively integrating recommender system knowledge into LLMs for natural language generation which is tailored towards recommendation tasks remains a challenge. This paper addresses this challenge by making two key contributions. First, we introduce a new dataset (REGEN) for natural language generation tasks in conversational recommendations. REGEN (Reviews Enhanced with GEnerative Narratives) extends the Amazon Product Reviews dataset with rich user narratives, including personalized explanations of product preferences, product endorsements for recommended items, and summaries of user purchase history. REGEN is made publicly available to facilitate further research. Furthermore, we establish benchmarks using well-known generative metrics, and perform an automated evaluation of the new dataset using a rater LLM. Second, the paper introduces a fusion architecture (CF model with an LLM) which serves as a baseline for REGEN. And to the best of our knowledge, represents the first attempt to analyze the capabilities of LLMs in understanding recommender signals and generating rich narratives. We demonstrate that LLMs can effectively learn from simple fusion architectures utilizing interaction-based CF embeddings, and this can be further enhanced using the metadata and personalization data associated with items. Our experiments show that combining CF and content embeddings leads to improvements of 4-12% in key language metrics compared to using either type of embedding individually. We also provide an analysis to interpret how CF and content embeddings contribute to this new generative task.

cs.CL↗

arXiv2024

UniMuMo: Unified Text, Music and Motion Generation

We introduce UniMuMo, a unified multimodal model capable of taking arbitrary text, music, and motion data as input conditions to generate outputs across all three modalities. To address the lack of time-synchronized data, we align unpaired music and motion data based on rhythmic patterns to leverage existing large-scale music-only and motion-only datasets. By converting music, motion, and text into token-based representation, our model bridges these modalities through a unified encoder-decoder transformer architecture. To support multiple generation tasks within a single framework, we introduce several architectural improvements. We propose encoding motion with a music codebook, mapping motion into the same feature space as music. We introduce a music-motion parallel generation scheme that unifies all music and motion generation tasks into a single transformer decoder architecture with a single training task of music-motion joint generation. Moreover, the model is designed by fine-tuning existing pre-trained single-modality models, significantly reducing computational demands. Extensive experiments demonstrate that UniMuMo achieves competitive results on all unidirectional generation benchmarks across music, motion, and text modalities. Quantitative results are available in the \href{https://hanyangclarence.github.io/unimumo_demo/}{project page}.

cs.SD↗

arXiv2024

From Vision to Audio and Beyond: A Unified Model for Audio-Visual Representation and Generation

Video encompasses both visual and auditory data, creating a perceptually rich experience where these two modalities complement each other. As such, videos are a valuable type of media for the investigation of the interplay between audio and visual elements. Previous studies of audio-visual modalities primarily focused on either audio-visual representation learning or generative modeling of a modality conditioned on the other, creating a disconnect between these two branches. A unified framework that learns representation and generates modalities has not been developed yet. In this work, we introduce a novel framework called Vision to Audio and Beyond (VAB) to bridge the gap between audio-visual representation learning and vision-to-audio generation. The key approach of VAB is that rather than working with raw video frames and audio data, VAB performs representation learning and generative modeling within latent spaces. In particular, VAB uses a pre-trained audio tokenizer and an image encoder to obtain audio tokens and visual features, respectively. It then performs the pre-training task of visual-conditioned masked audio token prediction. This training strategy enables the model to engage in contextual learning and simultaneous video-to-audio generation. After the pre-training phase, VAB employs the iterative-decoding approach to rapidly generate audio tokens conditioned on visual features. Since VAB is a unified model, its backbone can be fine-tuned for various audio-visual downstream tasks. Our experiments showcase the efficiency of VAB in producing high-quality audio from video, and its capability to acquire semantic audio-visual features, leading to competitive results in audio-visual retrieval and classification.

cs.MM↗

arXiv2024

Medical Image Segmentation via Single-Source Domain Generalization with Random Amplitude Spectrum Synthesis

The field of medical image segmentation is challenged by domain generalization (DG) due to domain shifts in clinical datasets. The DG challenge is exacerbated by the scarcity of medical data and privacy concerns. Traditional single-source domain generalization (SSDG) methods primarily rely on stacking data augmentation techniques to minimize domain discrepancies. In this paper, we propose Random Amplitude Spectrum Synthesis (RASS) as a training augmentation for medical images. RASS enhances model generalization by simulating distribution changes from a frequency perspective. This strategy introduces variability by applying amplitude-dependent perturbations to ensure broad coverage of potential domain variations. Furthermore, we propose random mask shuffle and reconstruction components, which can enhance the ability of the backbone to process structural information and increase resilience intra- and cross-domain changes. The proposed Random Amplitude Spectrum Synthesis for Single-Source Domain Generalization (RAS^4DG) is validated on 3D fetal brain images and 2D fundus photography, and achieves an improved DG segmentation performance compared to other SSDG models.

cs.CV↗

arXiv2023

V2Meow: Meowing to the Visual Beat via Video-to-Music Generation

Video-to-music generation demands both a temporally localized high-quality listening experience and globally aligned video-acoustic signatures. While recent music generation models excel at the former through advanced audio codecs, the exploration of video-acoustic signatures has been confined to specific visual scenarios. In contrast, our research confronts the challenge of learning globally aligned signatures between video and music directly from paired music and videos, without explicitly modeling domain-specific rhythmic or semantic relationships. We propose V2Meow, a video-to-music generation system capable of producing high-quality music audio for a diverse range of video input types using a multi-stage autoregressive model. Trained on 5k hours of music audio clips paired with video frames mined from in-the-wild music videos, V2Meow is competitive with previous domain-specific models when evaluated in a zero-shot manner. It synthesizes high-fidelity music audio waveforms solely by conditioning on pre-trained general-purpose visual features extracted from video frames, with optional style control via text prompts. Through both qualitative and quantitative evaluations, we demonstrate that our model outperforms various existing music generation systems in terms of visual-audio correspondence and audio quality. Music samples are available at tinyurl.com/v2meow.

cs.SD↗

arXiv2023

Physics-Driven Diffusion Models for Impact Sound Synthesis from Videos

Modeling sounds emitted from physical object interactions is critical for immersive perceptual experiences in real and virtual worlds. Traditional methods of impact sound synthesis use physics simulation to obtain a set of physics parameters that could represent and synthesize the sound. However, they require fine details of both the object geometries and impact locations, which are rarely available in the real world and can not be applied to synthesize impact sounds from common videos. On the other hand, existing video-driven deep learning-based approaches could only capture the weak correspondence between visual content and impact sounds since they lack of physics knowledge. In this work, we propose a physics-driven diffusion model that can synthesize high-fidelity impact sound for a silent video clip. In addition to the video content, we propose to use additional physics priors to guide the impact sound synthesis procedure. The physics priors include both physics parameters that are directly estimated from noisy real-world impact sound examples without sophisticated setup and learned residual parameters that interpret the sound environment via neural networks. We further implement a novel diffusion model with specific training and inference strategies to combine physics priors and visual information for impact sound synthesis. Experimental results show that our model outperforms several existing systems in generating realistic impact sounds. More importantly, the physics-based representations are fully interpretable and transparent, thus enabling us to perform sound editing flexibly.

cs.CV↗

arXiv2022

Incremental Learning Meets Transfer Learning: Application to Multi-site Prostate MRI Segmentation

Many medical datasets have recently been created for medical image segmentation tasks, and it is natural to question whether we can use them to sequentially train a single model that (1) performs better on all these datasets, and (2) generalizes well and transfers better to the unknown target site domain. Prior works have achieved this goal by jointly training one model on multi-site datasets, which achieve competitive performance on average but such methods rely on the assumption about the availability of all training data, thus limiting its effectiveness in practical deployment. In this paper, we propose a novel multi-site segmentation framework called incremental-transfer learning (ITL), which learns a model from multi-site datasets in an end-to-end sequential fashion. Specifically, "incremental" refers to training sequentially constructed datasets, and "transfer" is achieved by leveraging useful information from the linear combination of embedding features on each dataset. In addition, we introduce our ITL framework, where we train the network including a site-agnostic encoder with pre-trained weights and at most two segmentation decoder heads. We also design a novel site-level incremental loss in order to generalize well on the target domain. Second, we show for the first time that leveraging our ITL training scheme is able to alleviate challenging catastrophic forgetting problems in incremental learning. We conduct experiments using five challenging benchmark datasets to validate the effectiveness of our incremental-transfer learning approach. Our approach makes minimal assumptions on computation resources and domain-specific expertise, and hence constitutes a strong starting point in multi-site medical image segmentation.

cs.CV↗

arXiv2020

Multi-Instrumentalist Net: Unsupervised Generation of Music from Body Movements

We propose a novel system that takes as an input body movements of a musician playing a musical instrument and generates music in an unsupervised setting. Learning to generate multi-instrumental music from videos without labeling the instruments is a challenging problem. To achieve the transformation, we built a pipeline named 'Multi-instrumentalistNet' (MI Net). At its base, the pipeline learns a discrete latent representation of various instruments music from log-spectrogram using a Vector Quantized Variational Autoencoder (VQ-VAE) with multi-band residual blocks. The pipeline is then trained along with an autoregressive prior conditioned on the musician's body keypoints movements encoded by a recurrent neural network. Joint training of the prior with the body movements encoder succeeds in the disentanglement of the music into latent features indicating the musical components and the instrumental features. The latent space results in distributions that are clustered into distinct instruments from which new music can be generated. Furthermore, the VQ-VAE architecture supports detailed music generation with additional conditioning. We show that a Midi can further condition the latent space such that the pipeline will generate the exact content of the music being played by the instrument in the video. We evaluate MI Net on two datasets containing videos of 13 instruments and obtain generated music of reasonable audio quality, easily associated with the corresponding instrument, and consistent with the music audio content.

cs.SD↗

arXiv2020

Audeo: Audio Generation for a Silent Performance Video

We present a novel system that gets as an input video frames of a musician playing the piano and generates the music for that video. Generation of music from visual cues is a challenging problem and it is not clear whether it is an attainable goal at all. Our main aim in this work is to explore the plausibility of such a transformation and to identify cues and components able to carry the association of sounds with visual events. To achieve the transformation we built a full pipeline named `\textit{Audeo}' containing three components. We first translate the video frames of the keyboard and the musician hand movements into raw mechanical musical symbolic representation Piano-Roll (Roll) for each video frame which represents the keys pressed at each time step. We then adapt the Roll to be amenable for audio synthesis by including temporal correlations. This step turns out to be critical for meaningful audio generation. As a last step, we implement Midi synthesizers to generate realistic music. \textit{Audeo} converts video to audio smoothly and clearly with only a few setup constraints. We evaluate \textit{Audeo} on `in the wild' piano performance videos and obtain that their generated music is of reasonable audio quality and can be successfully recognized with high precision by popular music identification software.

cs.CV↗

arXiv2020

BI-MAML: Balanced Incremental Approach for Meta Learning

We present a novel Balanced Incremental Model Agnostic Meta Learning system (BI-MAML) for learning multiple tasks. Our method implements a meta-update rule to incrementally adapt its model to new tasks without forgetting old tasks. Such a capability is not possible in current state-of-the-art MAML approaches. These methods effectively adapt to new tasks, however, suffer from 'catastrophic forgetting' phenomena, in which new tasks that are streamed into the model degrade the performance of the model on previously learned tasks. Our system performs the meta-updates with only a few-shots and can successfully accomplish them. Our key idea for achieving this is the design of balanced learning strategy for the baseline model. The strategy sets the baseline model to perform equally well on various tasks and incorporates time efficiency. The balanced learning strategy enables BI-MAML to both outperform other state-of-the-art models in terms of classification accuracy for existing tasks and also accomplish efficient adaption to similar new tasks with less required shots. We evaluate BI-MAML by conducting comparisons on two common benchmark datasets with multiple number of image classification tasks. BI-MAML performance demonstrates advantages in both accuracy and efficiency.

cs.LG↗