arXiv subjects

Yan Song

Publications and source records attributed to Yan Song.

At least 37 records · Page 2Linked to original sources

arXiv2026

Towards Fine-grained Temporal Perception: Post-Training Large Audio-Language Models with Audio-Side Time Prompt

Large Audio-Language Models (LALMs) enable general audio understanding and demonstrate remarkable performance across various audio tasks. However, these models still face challenges in temporal perception (e.g., inferring event onset and offset), leading to limited utility in fine-grained scenarios. To address this issue, we propose Audio-Side Time Prompt and leverage Reinforcement Learning (RL) to develop the TimePro-RL framework for fine-grained temporal perception. Specifically, we encode timestamps as embeddings and interleave them within the audio feature sequence as temporal coordinates to prompt the model. Furthermore, we introduce RL following Supervised Fine-Tuning (SFT) to directly optimize temporal alignment performance. Experiments demonstrate that TimePro-RL achieves significant performance gains across a range of audio temporal tasks, such as audio grounding, sound event detection, and dense audio captioning, validating its robust effectiveness.

cs.SD↗

arXiv2026

Learning Shared Sentiment Prototypes for Adaptive Multimodal Sentiment Analysis

Multimodal sentiment analysis (MSA) aims to predict human sentiment from textual, acoustic, and visual information in videos. Recent studies improve multimodal fusion by modeling modality interaction and assigning different modality weights. However, they usually compress diverse sentiment cues into a single compact representation before sentiment reasoning. This early aggregation makes it difficult to preserve the internal structure of sentiment evidence, where different cues may complement, conflict with, or differ in reliability from each other. In addition, modality importance is often determined only once during fusion, so later reasoning cannot further adjust modality contributions. To address these issues, we propose PRISM, a framework that unifies structured affective extraction and adaptive modality evaluation. PRISM organizes multimodal evidence in a shared prototype space, which supports structured cross-modal comparison and adaptive fusion. It further applies dynamic modality reweighting during reasoning, allowing modality contributions to be continuously refined as semantic interactions become deeper. Experiments on three benchmark datasets show that PRISM outperforms representative baselines.

cs.MM↗

arXiv2026

Spectrogram features for audio and speech analysis

Spectrogram-based representations have grown to dominate the feature space for deep learning audio analysis systems, and are often adopted for speech analysis also. Initially, the primary motivator for spectrogram-based representations was their ability to present sound as a two dimensional signal in the time-frequency plane, which not only provides an interpretable physical basis for analysing sound, but also unlocks the use of a wide range of machine learning techniques such as convolutional neural networks, that had been developed for image processing. A spectrogram is a matrix characterised by the resolution and span of its two dimensions, as well as by the representation and scaling of each element. Many possibilities for these three characteristics have been explored by researchers across numerous application areas, with different settings showing affinity for various tasks. This paper reviews the use of spectrogram-based representations and surveys the state-of-the-art to question how front-end feature representation choice allies with back-end classifier architecture for different tasks.

eess.AS↗

arXiv2026

Hardware Co-Design Scaling Laws via Roofline Modelling for On-Device LLMs

Vision-Language-Action Models (VLAs) have emerged as a key paradigm of Physical AI and are increasingly deployed in autonomous vehicles, robots, and smart spaces. In these resource-constrained on-device settings, selecting an appropriate large language model (LLM) backbone is a critical challenge: models must balance accuracy with strict inference latency and hardware efficiency constraints. This makes hardware-software co-design a game-changing requirement for on-device LLM deployment, where each hardware platform demands a tailored architectural solution. We propose a hardware co-design law that jointly captures model accuracy and inference performance. Specifically, we model training loss as an explicit function of architectural hyperparameters and characterise inference latency via roofline modelling. We empirically evaluate 1,942 candidate architectures on NVIDIA Jetson Orin, training 170 selected models for 10B tokens each to fit a scaling law relating architecture to training loss. By coupling this scaling law with latency modelling, we establish a direct accuracy-latency correspondence and identify the Pareto frontier for hardware co-designed LLMs. We further formulate architecture search as a joint optimisation over precision and performance, deriving feasible design regions under industrial hardware and application budgets. Our approach reduces architecture selection from months to days. At the same latency as Qwen2.5-0.5B on the target hardware, our co-designed architecture achieves 19.42% lower perplexity on WikiText-2. To our knowledge, this is the first principled and operational framework for hardware co-design scaling laws in on-device LLM deployment. We will make the code and related checkpoints publicly available.

cs.LG↗

arXiv2026

From $O(mn)$ to $O(r^2)$: Two-Sided Low-Rank Communication for Adam in Distributed Training with Memory Efficiency

As foundation models continue to scale, pretraining increasingly relies on data-parallel distributed optimization, making bandwidth-limited gradient synchronization a key bottleneck. Orthogonally, projection-based low-rank optimizers were mainly designed for memory efficiency, but remain suboptimal for communication-limited training: one-sided synchronization still transmits an $O(rn)$ object for an $m\times n$ matrix gradient and refresh steps can dominate peak communicated bytes. We propose TSR, which brings two-sided low-rank communication to Adam-family updates (TSR-Adam) by synchronizing a compact core $U^\top G V\in\mathbb{R}^{r\times r}$, reducing the dominant per-step payload from $O(mn)$ to $O(r^2)$ while keeping moment states in low-dimensional cores. To further reduce the peak communication from subspace refresh, TSR-Adam adopts a randomized SVD-based refresh that avoids full-gradient synchronization. We additionally extend low-rank communication to embedding gradients with embedding-specific ranks and refresh schedules, yielding additional communication and memory savings over keeping embeddings dense. Across pretraining from 60M to 1B model scales, TSR-Adam reduces average communicated bytes per step by $13\times$, and on GLUE fine-tuning it reduces communication by $25\times$, while achieving comparable performance; we further provide a theoretical stationarity analysis for the proposed update. Code is available at https://github.com/DKmiyan/TSR-Adam.

cs.LG↗

arXiv2026

Training Data Selection with Gradient Orthogonality for Efficient Domain Adaptation

Fine-tuning large language models (LLMs) for specialized domains often necessitates a trade-off between acquiring domain expertise and retaining general reasoning capabilities, a phenomenon known as catastrophic forgetting. Existing remedies face a dichotomy: gradient surgery methods offer geometric safety but incur prohibitive computational costs via online projections, while efficient data selection approaches reduce overhead but remain blind to conflict-inducing gradient directions. In this paper, we propose Orthogonal Gradient Selection (OGS), a data-centric method that harmonizes domain performance, general capability retention, and training efficiency. OGS shifts the geometric insights of gradient projection from the optimizer to the data selection stage by treating data selection as a constrained decision-making process. By leveraging a lightweight Navigator model and reinforcement learning techniques, OGS dynamically identifies training samples whose gradients are orthogonal to a general-knowledge anchor. This approach ensures naturally safe updates for target models without modifying the optimizer or incurring runtime projection costs. Experiments across medical, legal, and financial domains demonstrate that OGS achieves excellent results, significantly improving domain performance and training efficiency while maintaining or even enhancing performance on general tasks such as GSM8K.

cs.LG↗

arXiv2026

LEAD: Layer-wise Expert-aligned Decoding for Faithful Radiology Report Generation

Radiology Report Generation (RRG) aims to produce accurate and coherent diagnostics from medical images. Although large vision language models (LVLM) improve report fluency and accuracy, they exhibit hallucinations, generating plausible yet image-ungrounded pathological details. Existing methods primarily rely on external knowledge guidance to facilitate the alignment between generated text and visual information. However, these approaches often ignore the inherent decoding priors and vision-language alignment biases in pretrained models and lack robustness due to reliance on constructed guidance. In this paper, we propose Layer-wise Expert-aligned Decoding (LEAD), a novel method to inherently modify the LVLM decoding trajectory. A multiple experts module is designed for extracting distinct pathological features which are integrated into each decoder layer via a gating mechanism. This layer-wise architecture enables the LLM to consult expert features at every inference step via a learned gating function, thereby dynamically rectifying decoding biases and steering the generation toward factual consistency. Experiments conducted on multiple public datasets demonstrate that the LEAD method yields effective improvements in clinical accuracy metrics and mitigates hallucinations while preserving high generation quality.

cs.CL↗

arXiv2026

CAG-Avatar: Cross-Attention Guided Gaussian Avatars for High-Fidelity Head Reconstruction

Creating high-fidelity, real-time drivable 3D head avatars is a core challenge in digital animation. While 3D Gaussian Splashing (3D-GS) offers unprecedented rendering speed and quality, current animation techniques often rely on a "one-size-fits-all" global tuning approach, where all Gaussian primitives are uniformly driven by a single expression code. This simplistic approach fails to unravel the distinct dynamics of different facial regions, such as deformable skin versus rigid teeth, leading to significant blurring and distortion artifacts. We introduce Conditionally-Adaptive Gaussian Avatars (CAG-Avatar), a framework that resolves this key limitation. At its core is a Conditionally Adaptive Fusion Module built on cross-attention. This mechanism empowers each 3D Gaussian to act as a query, adaptively extracting relevant driving signals from the global expression code based on its canonical position. This "tailor-made" conditioning strategy drastically enhances the modeling of fine-grained, localized dynamics. Our experiments confirm a significant improvement in reconstruction fidelity, particularly for challenging regions such as teeth, while preserving real-time rendering performance.

cs.GR↗

arXiv2026

GAT-NeRF: Geometry-Aware-Transformer Enhanced Neural Radiance Fields for High-Fidelity 4D Facial Avatars

High-fidelity 4D dynamic facial avatar reconstruction from monocular video is a critical yet challenging task, driven by increasing demands for immersive virtual human applications. While Neural Radiance Fields (NeRF) have advanced scene representation, their capacity to capture high-frequency facial details, such as dynamic wrinkles and subtle textures from information-constrained monocular streams, requires significant enhancement. To tackle this challenge, we propose a novel hybrid neural radiance field framework, called Geometry-Aware-Transformer Enhanced NeRF (GAT-NeRF) for high-fidelity and controllable 4D facial avatar reconstruction, which integrates the Transformer mechanism into the NeRF pipeline. GAT-NeRF synergistically combines a coordinate-aligned Multilayer Perceptron (MLP) with a lightweight Transformer module, termed as Geometry-Aware-Transformer (GAT) due to its processing of multi-modal inputs containing explicit geometric priors. The GAT module is enabled by fusing multi-modal input features, including 3D spatial coordinates, 3D Morphable Model (3DMM) expression parameters, and learnable latent codes to effectively learn and enhance feature representations pertinent to fine-grained geometry. The Transformer's effective feature learning capabilities are leveraged to significantly augment the modeling of complex local facial patterns like dynamic wrinkles and acne scars. Comprehensive experiments unequivocally demonstrate GAT-NeRF's state-of-the-art performance in visual fidelity and high-frequency detail recovery, forging new pathways for creating realistic dynamic digital humans for multimedia applications.

cs.CV↗

arXiv2026

Fine-grained Verbal Attack Detection via a Hierarchical Divide-and-Conquer Framework

In the digital era, effective identification and analysis of verbal attacks are essential for maintaining online civility and ensuring social security. However, existing research is limited by insufficient modeling of conversational structure and contextual dependency, particularly in Chinese social media where implicit attacks are prevalent. Current attack detection studies often emphasize general semantic understanding while overlooking user response relationships, hindering the identification of implicit and context-dependent attacks. To address these challenges, we present the novel "Hierarchical Attack Comment Detection" dataset and propose a divide-and-conquer, fine-grained framework for verbal attack recognition based on spatiotemporal information. The proposed dataset explicitly encodes hierarchical reply structures and chronological order, capturing complex interaction patterns in multi-turn discussions. Building on this dataset, the framework decomposes attack detection into hierarchical subtasks, where specialized lightweight models handle explicit detection, implicit intent inference, and target identification under constrained context. Extensive experiments on the proposed dataset and benchmark intention detection datasets show that smaller models using our framework significantly outperform larger monolithic models relying on parameter scaling, demonstrating the effectiveness of structured task decomposition.

cs.CL↗

arXiv2025

Chain-of-thought Reviewing and Correction for Time Series Question Answering

With the advancement of large language models (LLMs), diverse time series analysis tasks are reformulated as time series question answering (TSQA) through a unified natural language interface. However, existing LLM-based approaches largely adopt general natural language processing techniques and are prone to reasoning errors when handling complex numerical sequences. Different from purely textual tasks, time series data are inherently verifiable, enabling consistency checking between reasoning steps and the original input. Motivated by this property, we propose T3LLM, which performs multi-step reasoning with an explicit correction mechanism for time series question answering. The T3LLM framework consists of three LLMs, namely, a worker, a reviewer, and a student, that are responsible for generation, review, and reasoning learning, respectively. Within this framework, the worker generates step-wise chains of thought (CoT) under structured prompts, while the reviewer inspects the reasoning, identifies erroneous steps, and provides corrective comments. The collaboratively generated corrected CoT are used to fine-tune the student model, internalizing multi-step reasoning and self-correction into its parameters. Experiments on multiple real-world TSQA benchmarks demonstrate that T3LLM achieves state-of-the-art performance over strong LLM-based baselines.

cs.CL↗

arXiv2025

MoE Pathfinder: Trajectory-driven Expert Pruning

Mixture-of-experts (MoE) architectures used in large language models (LLMs) achieve state-of-the-art performance across diverse tasks yet face practical challenges such as deployment complexity and low activation efficiency. Expert pruning has thus emerged as a promising solution to reduce computational overhead and simplify the deployment of MoE models. However, existing expert pruning approaches conventionally rely on local importance metrics and often apply uniform layer-wise pruning, leveraging only partial evaluation signals and overlooking the heterogeneous contributions of experts across layers. To address these limitations, we propose an expert pruning approach based on the trajectory of activated experts across layers, which treats MoE as a weighted computation graph and casts expert selection as a global optimal path planning problem. Within this framework, we integrate complementary importance signals from reconstruction error, routing probabilities, and activation strength at the trajectory level, which naturally yields non-uniform expert retention across layers. Experiments show that our approach achieves superior pruning performance on nearly all tasks compared with most existing approaches.

cs.LG↗

arXiv2025

From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory

Large Language Models (LLMs) based agents have demonstrated remarkable potential in autonomous task-solving across complex, open-ended environments. A promising approach for improving the reasoning capabilities of LLM agents is to better utilize prior experiences in guiding current decisions. However, LLMs acquire experience either through implicit memory via training, which suffers from catastrophic forgetting and limited interpretability, or explicit memory via prompting, which lacks adaptability. In this paper, we introduce a novel agent-centric, trainable, multi-layered graph memory framework and evaluate how context memory enhances the ability of LLMs to utilize parametric information. The graph abstracts raw agent trajectories into structured decision paths in a state machine and further distills them into high-level, human-interpretable strategic meta-cognition. In order to make memory adaptable, we propose a reinforcement-based weight optimization procedure that estimates the empirical utility of each meta-cognition based on reward feedback from downstream tasks. These optimized strategies are then dynamically integrated into the LLM agent's training loop through meta-cognitive prompting. Empirically, the learnable graph memory delivers robust generalization, improves LLM agents' strategic reasoning performance, and provides consistent benefits during Reinforcement Learning (RL) training.

cs.CL↗

arXiv2025

Detect Any Sound: Open-Vocabulary Sound Event Detection with Multi-Modal Queries

Most existing sound event detection~(SED) algorithms operate under a closed-set assumption, restricting their detection capabilities to predefined classes. While recent efforts have explored language-driven zero-shot SED by exploiting audio-language models, their performance is still far from satisfactory due to the lack of fine-grained alignment and cross-modal feature fusion. In this work, we propose the Detect Any Sound Model (DASM), a query-based framework for open-vocabulary SED guided by multi-modal queries. DASM formulates SED as a frame-level retrieval task, where audio features are matched against query vectors derived from text or audio prompts. To support this formulation, DASM introduces a dual-stream decoder that explicitly decouples event recognition and temporal localization: a cross-modality event decoder performs query-feature fusion and determines the presence of sound events at the clip-level, while a context network models temporal dependencies for frame-level localization. Additionally, an inference-time attention masking strategy is proposed to leverage semantic relations between base and novel classes, substantially enhancing generalization to novel classes. Experiments on the AudioSet Strong dataset demonstrate that DASM effectively balances localization accuracy with generalization to novel classes, outperforming CLAP-based methods in open-vocabulary setting (+ 7.8 PSDS) and the baseline in the closed-set setting (+ 6.9 PSDS). Furthermore, in cross-dataset zero-shot evaluation on DESED, DASM achieves a PSDS1 score of 42.2, even exceeding the supervised CRNN baseline. The project page is available at https://cai525.github.io/Transformer4SED/demo_page/DASM/.

cs.SD↗

arXiv2025

Frustratingly Easy Task-aware Pruning for Large Language Models

Pruning provides a practical solution to reduce the resources required to run large language models (LLMs) to benefit from their effective capabilities as well as control their cost for training and inference. Research on LLM pruning often ranks the importance of LLM parameters using their magnitudes and calibration-data activations and removes (or masks) the less important ones, accordingly reducing LLMs' size. However, these approaches primarily focus on preserving the LLM's ability to generate fluent sentences, while neglecting performance on specific domains and tasks. In this paper, we propose a simple yet effective pruning approach for LLMs that preserves task-specific capabilities while shrinking their parameter space. We first analyze how conventional pruning minimizes loss perturbation under general-domain calibration and extend this formulation by incorporating task-specific feature distributions into the importance computation of existing pruning algorithms. Thus, our framework computes separate importance scores using both general and task-specific calibration data, partitions parameters into shared and exclusive groups based on activation-norm differences, and then fuses their scores to guide the pruning process. This design enables our method to integrate seamlessly with various foundation pruning techniques and preserve the LLM's specialized abilities under compression. Experiments on widely used benchmarks demonstrate that our approach is effective and consistently outperforms the baselines with identical pruning ratios and different settings.

cs.CL↗

arXiv2025

ICA-RAG: Information Completeness Guided Adaptive Retrieval-Augmented Generation for Disease Diagnosis

Retrieval-Augmented Large Language Models (LLMs), which integrate external knowledge, have shown remarkable performance in medical domains, including clinical diagnosis. However, existing RAG methods often struggle to tailor retrieval strategies to diagnostic difficulty and input sample informativeness. This limitation leads to excessive and often unnecessary retrieval, impairing computational efficiency and increasing the risk of introducing noise that can degrade diagnostic accuracy. To address this, we propose ICA-RAG (\textbf{I}nformation \textbf{C}ompleteness Guided \textbf{A}daptive \textbf{R}etrieval-\textbf{A}ugmented \textbf{G}eneration), a novel framework for enhancing RAG reliability in disease diagnosis. ICA-RAG utilizes an adaptive control module to assess the necessity of retrieval based on the input's information completeness. By optimizing retrieval and incorporating knowledge filtering, ICA-RAG better aligns retrieval operations with clinical requirements. Experiments on three Chinese electronic medical record datasets demonstrate that ICA-RAG significantly outperforms baseline methods, highlighting its effectiveness in clinical diagnosis.

cs.CL↗

arXiv2025

Memory-Driven Self-Improvement for Decision Making with Large Language Models

Large language models (LLMs) have emerged as effective action policies for sequential decision-making (SDM) tasks due to their extensive prior knowledge. However, this broad yet general knowledge is often insufficient for specific decision-making tasks with limited task-related data, making it challenging to efficiently adapt LLMs to specific SDM tasks. To address this challenge, we propose a memory-driven self-improvement framework that combines LLM general prior knowledge with a compact memory of domain-specific experiences. Memory retains past interactions and associated Q-values, thereby capturing decision-relevant knowledge that facilitates accurate value estimation and informs the LLM prior refinement. The refined LLM prior, in turn, generates higher-reward trajectories that further enrich memory, forming a natural self-improvement framework where memory and LLM prior mutually reinforce each other. Experiments show that our memory-driven approach significantly outperforms both traditional RL and LLM-based baselines, e.g., improving performance by over 40\% on in-distribution tasks and over 75\% when generalized to unseen tasks in ALFWorld.

cs.LG↗

arXiv2025

An Efficient Transfer Learning Method Based on Adapter with Local Attributes for Speech Emotion Recognition

Existing speech emotion recognition (SER) methods commonly suffer from the lack of high-quality large-scale corpus, partly due to the complex, psychological nature of emotion which makes accurate labeling difficult and time consuming. Recently, transfer learning based methods that exploit the encoders pretrained on large-scale speech corpus (e.g., Wav2Vec2.0 and HuBERT) have shown strong potential for downstream SER tasks. However, task-specific fine-tuning remains necessary for various conversational scenarios of different topics, speakers and languages to achieve satisfactory performance. It generally requires costly encoder retraining for individual SER tasks. To address this issue, we propose to train an adapter with local attributes for efficient transfer learning. Specifically, a weighted average pooling-Transformer (WAP-Transformer) is proposed as a lightweight backbone to enrich the frame-level representation. An adapter with teacher-student branches is exploited for task-agnostic transfer learning, where the student branch is jointly optimized via mask prediction and self-distillation objectives, and the teacher branch is obtained online from the student via exponential moving average (EMA). Meanwhile, local attributes are learned from the teacher branch via unsupervised clustering, which aims to act as a universal model that provides additional semantic-rich supervisions. A statistical attentive pooling (SAP) module is proposed to obtain utterance representation for fine-tuning. To evaluate the effectiveness of the proposed adapter with local attributes, extensive experiments have been conducted on IEMOCAP. Superior performance has been reported, compared to the previous state-of-the-art methods in similar settings.

cs.SD↗