arXiv subjects

Li Wan

Publications and source records attributed to Li Wan.

At least 19 recordsLinked to original sources

arXiv2026

Performance Analysis and Optimal Design of ORB-Type GRAND Algorithms

Guessing Random Additive Noise Decoding (GRAND) performs decoding by sequentially guessing channel error patterns (EPs). Ordered Reliability Bits GRAND (ORBGRAND) is a notable instance suitable for efficient implementation, as it schedules EPs solely according to the ranking of soft channel outputs. In this paper, we generalize this principle to a broader class of GRAND algorithms whose testing order depends only on reliability ranking, referred to as ORB-type GRAND. We develop a unified analytical framework based on a key quantity termed the average guessing posterior (AGP), which captures the effectiveness of each EP and reduces decoding into an ordering problem over the EP space. For random code ensembles, we derive exact expressions for the block error rate (BLER), stopping-time distribution, and average number of tests under a fixed test budget. The analysis separates target-miss and target-preemption errors and shows that ordering EPs by non-increasing AGP is optimal over the EP set under consideration. For fixed linear block codes, we derive the BLER expression that isolates the code-dependent target-preemption term and characterize this term through higher-order weight relationships of codeword tuples, with a computable first-order upper bound as a useful special case. Guided by these insights, we formulate ReShuffled-ORBGRAND (RS-ORBGRAND) as an offline AGP-based reshuffling scheme. Numerical results for the Bose--Chaudhuri--Hocquenghem (BCH)$(127,113)$ code show that RS-ORBGRAND consistently improves existing ORB-type GRAND algorithms and lies within $0.1$~dB of a maximum-likelihood decoding lower-bound benchmark at a BLER of $10^{-6}$.

cs.IT↗

arXiv2026

Enabling Extensible Embodied Capabilities with Tools

Most existing embodied intelligence methods formulate perception, reasoning, planning, and control within a unified parameterized policy. Yet these capabilities are inherently hierarchical and heterogeneous, making them difficult to reliably learn and modularize within a single model. We propose a capability externalization approach that decouples heterogeneous capabilities into independently optimized tools, dynamically invoked at inference time. To this end, we introduce Embodied Tool Protocol (ETP), a standardized protocol for embodied tool registration, discovery, invocation, and execution, and curate 100+ validated tools spanning perception, cognition, reasoning, and execution as the tool base. Building on this, we construct EmbodiedToolBench to evaluate both whether tool augmentation improves embodied performance and how well current models use tools across tool-necessity recognition, tool selection, tool execution, and tool-chain composition. Experiments across simulation and real-world platforms confirm that capability externalization consistently improves embodied performance (avg. gain 31% on EB-ALFRED and 36% on EB-Navigation), yet reveal a clear boundary: gains are substantial for cognition and perception but are limited for execution-type capabilities. Moreover, our analysis reveals that knowing when, which, and how to invoke tools remains a persistent challenge across all models, thereby highlighting embodied tool competence as a critical direction for future research.

cs.RO↗

arXiv2026

Reducing ORBGRAND Latency via Partial Gaussian Elimination

Guessing Random Additive Noise Decoding (GRAND) is a universal framework for decoding all block codes by testing candidate error patterns (EPs). Ordered Reliability Bits GRAND (ORBGRAND) facilitates parallel implementation of GRAND by exploiting log-likelihood ratio (LLR) rankings but still suffers from high tail latency under unfavorable channel conditions, limiting its use in real-time systems. We propose an elimination-aided ORBGRAND scheme that reduces decoding latency by integrating the Rank of the Most Reliable Erroneous (RMRE) bit with a partial Gaussian-elimination (GE) filtering mechanism. The scheme groups and jointly verifies EPs that share the same RMRE, and once a valid EP is identified, the ORBGRAND search is resumed. By leveraging prior GE steps to filter out unnecessary guesses, this approach significantly reduces the number of EPs to be tested, thereby lowering both average and worst-case latency while maintaining error-correction performance. Simulation results show that compared to the original ORBGRAND, the elimination-aided ORBGRAND filters out more than 50\% of EPs and correspondingly reduce overall computational complexity, all with no loss in block error rate. This demonstrates that this approach is suitable for ultra-reliable low-latency communication scenarios.

cs.IT↗

arXiv2025

SLM-TTA: A Framework for Test-Time Adaptation of Generative Spoken Language Models

Spoken Language Models (SLMs) are increasingly central to modern speech-driven applications, but performance degrades under acoustic shift - real-world noise, reverberation, and microphone variation. Prior solutions rely on offline domain adaptation, which is post-hoc, data-intensive, and slow. We introduce the first test-time adaptation (TTA) framework for generative SLMs that process interleaved audio-text prompts. Our method updates a small, targeted subset of parameters during inference using only the incoming utterance, requiring no source data or labels. This stabilizes token distributions and improves robustness to acoustic variability without degrading core task accuracy. Evaluated on automatic speech recognition, speech translation, and 19 audio understanding tasks from AIR-Bench, our approach yields consistent gains under diverse corruptions. Because adaptation touches only a small fraction of weights, it is both compute- and memory-efficient, supporting deployment on resource-constrained platforms. This work enhances the robustness and adaptability of generative SLMs for real-world speech-driven applications.

cs.SD↗

arXiv2025

Anharmonic interaction as random field for thermal transport in FPU-$\beta$ lattice

We present an open quantum theory for the thermal transport in the Fermi-Pasta-Ulam-$\beta$(FPU-$\beta$) lattice. In the theory, local bosons(LBs) are introduced as carriers for the transport. The LBs are stimulated by individual atoms in the lattice, which are different from the phonons that are collective motions of the atoms. The LBs move in the FPU chain and are governed by a set of stochastic differential equations(SDEs). The anharmonic interaction between the atoms in the lattice is transformed to a random field by the Hubbard-Stratonovich transformation, and has been implemented in the set of SDEs. By solving the set of SDEs at the steady state, we study the influence of the anharmonic interaction on the thermal transport. Results show that the anharmonic interaction decreases the thermal current by trapping the LBs on the lattice sites, as well as increase the thermal current by enhancing the amount of the LBs for the transport. The competition between these two mechanisms makes the thermal conductivity of the lattice dependent on the anharmonic interaction non-monotonically. The finite size effect of the thermal conductivity has also been captured by the theory.

cond-mat.mes-hall↗

arXiv2025

Tree Training: Accelerating Agentic LLMs Training via Shared Prefix Reuse

Agentic large language model (LLM) training often involves multi-turn interaction trajectories that branch into multiple execution paths due to concurrent tool use, think-mode, sub-agent, context management and other runtime designs. As a result, the tokens produced by a single task naturally form a tree-structured token trajectory with shared prefixes, rather than a linear sequence. Existing training pipelines linearize such trajectories and treat each branch independently, leading to substantial redundant computation in both forward and backward passes. We derive that averaging the loss over all branches independently is algebraically identical to a per-token weighted loss, where each token's weight equals the fraction of branches passing through it. The problem therefore reduces to computing the log-probability of every token in the prefix tree exactly once, with no repeated computation across shared prefixes: we propose DFS serialization of the tree, which visits every token exactly once, and adapt full-attention and SSM layers to ensure the resulting log-probabilities match independent per-branch calculation exactly. In practice, a single trajectory tree can be too large to fit in GPU memory; we therefore propose Redundancy-Free Tree Partitioning, which handles memory-constrained settings with zero redundant computation and peak memory bounded by a single root-to-leaf path. Together, these contributions form Tree Training, an efficient framework for training LLMs on tree-structured trajectories, achieving up to 6.2x end-to-end training speedup on dense and MoE models for both supervised fine-tuning and reinforcement learning.

cs.LG↗

arXiv2025

A Parallelization Strategy for GRAND with Optimality Guarantee by Exploiting Error Pattern Tree Representation

Parallelism has become a central concern in modern decoding frameworks aiming to meet stringent throughput and latency requirements. Guessing Random Additive Noise Decoding (GRAND) is a recently proposed decoding paradigm that tests candidate Error Patterns (EPs) until a valid codeword is found. Among its variants, Soft GRAND (SGRAND) achieves maximum-likelihood (ML) decoding but relies on real-time generation and likelihood ordering of EPs, making parallel execution nontrivial under the ML optimality constraint. In this work, we introduce a unified binary tree representation of EPs, termed the EP tree, which formalizes the hierarchical structure underlying SGRAND and Ordered Reliability Bits (ORB) GRAND algorithms, enabling structured organization of EPs and algorithmic-level parallel exploration. Building upon this unified framework, we propose a parallel design of SGRAND that preserves ML optimality while significantly reducing decoding complexity through pruning strategies and tree-based computation. Furthermore, we develop an enhanced ORBGRAND algorithm based on the same EP tree representation, improving decoding performance toward ML while retaining parallel efficiency. Numerical experiments show that the proposed parallel SGRAND achieves a $3.96\times$ reduction in decoding latency compared with its serial counterpart, while the enhanced ORBGRAND achieves a $4.21\times$ speedup, demonstrating the effectiveness of the unified tree-based framework and its strong potential for future algorithmic and hardware optimizations.

cs.IT↗

arXiv2025

Multi-Channel Differential ASR for Robust Wearer Speech Recognition on Smart Glasses

With the growing adoption of wearable devices such as smart glasses for AI assistants, wearer speech recognition (WSR) is becoming increasingly critical to next-generation human-computer interfaces. However, in real environments, interference from side-talk speech remains a significant challenge to WSR and may cause accumulated errors for downstream tasks such as natural language processing. In this work, we introduce a novel multi-channel differential automatic speech recognition (ASR) method for robust WSR on smart glasses. The proposed system takes differential inputs from different frontends that complement each other to improve the robustness of WSR, including a beamformer, microphone selection, and a lightweight side-talk detection model. Evaluations on both simulated and real datasets demonstrate that the proposed system outperforms the traditional approach, achieving up to an 18.0% relative reduction in word error rate.

eess.AS↗

arXiv2025

Fine-tuning ORBGRAND with Very Few Channel Soft Values

Guessing random additive noise decoding (GRAND) is a universal decoding paradigm that decodes by repeatedly testing error patterns until identifying a codeword, where the ordering of tests is generated by the received channel values. On one hand, while testing error patterns in a descending order of posterior probabilities leads to maximum likelihood decoding, its implementation complexity is prohibitive. On the other hand, testing error patterns with a prescribed set of error patterns permuted by the ranking among magnitudes of log-likelihood ratios (i.e., ordered reliability bits, ORB) enables efficient implementation, but results in performance loss for finite-length codes. Aiming at harnessing the strengths of these two approaches, this work proposes a fine-tuning method to improve ORBGRAND, adjusting the ordering of tests with the aid of very few exact channel soft values. This method is based on a metric for assessing the ``well-orderedness'' of error patterns. The metric is studied via the lens of the asymptotic theory of integer partitioning, which provides highly accurate estimation in numerical experiments. The metric then leads to an effective identification of fine-tuning to conduct, at the cost of a negligible increment of complexity. Numerical experiments demonstrate that the proposed fine-tuning method achieves a substantial performance enhancement compared with ORBGRAND.

cs.IT↗

arXiv2025

MMW: Side Talk Rejection Multi-Microphone Whisper on Smart Glasses

Smart glasses are increasingly positioned as the next-generation interface for ubiquitous access to large language models (LLMs). Nevertheless, achieving reliable interaction in real-world noisy environments remains a major challenge, particularly due to interference from side speech. In this work, we introduce a novel side-talk rejection multi-microphone Whisper (MMW) framework for smart glasses, incorporating three key innovations. First, we propose a Mix Block based on a Tri-Mamba architecture to effectively fuse multi-channel audio at the raw waveform level, while maintaining compatibility with streaming processing. Second, we design a Frame Diarization Mamba Layer to enhance frame-level side-talk suppression, facilitating more efficient fine-tuning of Whisper models. Third, we employ a Multi-Scale Group Relative Policy Optimization (GRPO) strategy to jointly optimize frame-level and utterance-level side speech suppression. Experimental evaluations demonstrate that the proposed MMW system can reduce the word error rate (WER) by 4.95\% in noisy conditions.

eess.AS↗

arXiv2024

MASV: Speaker Verification with Global and Local Context Mamba

Deep learning models like Convolutional Neural Networks and transformers have shown impressive capabilities in speech verification, gaining considerable attention in the research community. However, CNN-based approaches struggle with modeling long-sequence audio effectively, resulting in suboptimal verification performance. On the other hand, transformer-based methods are often hindered by high computational demands, limiting their practicality. This paper presents the MASV model, a novel architecture that integrates the Mamba module into the ECAPA-TDNN framework. By introducing the Local Context Bidirectional Mamba and Tri-Mamba block, the model effectively captures both global and local context within audio sequences. Experimental results demonstrate that the MASV model substantially enhances verification performance, surpassing existing models in both accuracy and efficiency.

eess.AS↗

arXiv2024

Constructing the body source for high-order lattice Boltzmann method

This paper presents a novel strategy for constructing body source terms in the high-order lattice Boltzmann method (LBM), designed to efficiently introduce various physical phenomena by modifying the non-equilibrium distribution function. The source term, expressed as a Hermite polynomial, provides a flexible framework for simulating complex fluid flows. Three typical source terms are given: a body force source for gravity-driven flows, a thermal dissipation source for controlling the Prandtl number, and a pressure tensor source for modeling multiphase flows. Chapman-Enskog analysis confirms that the source terms recover the expected macroscopic equations. Notably, the proposed strategy eliminates the need for explicit construction of the collision operator, a challenge in conventional approaches for handling diverse physical scenarios. Furthermore, the method is compatible with the traditional BGK model, ensuring its applicability to various high-order lattices. The model's accuracy and versatility are validated through a series of benchmark tests, showing excellent agreement with existing literature results.

physics.flu-dyn↗

arXiv2024

Query-by-Example Keyword Spotting Using Spectral-Temporal Graph Attentive Pooling and Multi-Task Learning

Existing keyword spotting (KWS) systems primarily rely on predefined keyword phrases. However, the ability to recognize customized keywords is crucial for tailoring interactions with intelligent devices. In this paper, we present a novel Query-by-Example (QbyE) KWS system that employs spectral-temporal graph attentive pooling and multi-task learning. This framework aims to effectively learn speaker-invariant and linguistic-informative embeddings for QbyE KWS tasks. Within this framework, we investigate three distinct network architectures for encoder modeling: LiCoNet, Conformer and ECAPA_TDNN. The experimental results on a substantial internal dataset of $629$ speakers have demonstrated the effectiveness of the proposed QbyE framework in maximizing the potential of simpler models such as LiCoNet. Particularly, LiCoNet, which is 13x more efficient, achieves comparable performance to the computationally intensive Conformer model (1.98% vs. 1.63\% FRR at 0.3 FAs/Hr).

cs.CL↗

arXiv2024

Disentangled Training with Adversarial Examples For Robust Small-footprint Keyword Spotting

A keyword spotting (KWS) engine that is continuously running on device is exposed to various speech signals that are usually unseen before. It is a challenging problem to build a small-footprint and high-performing KWS model with robustness under different acoustic environments. In this paper, we explore how to effectively apply adversarial examples to improve KWS robustness. We propose datasource-aware disentangled learning with adversarial examples to reduce the mismatch between the original and adversarial data as well as the mismatch across original training datasources. The KWS model architecture is based on depth-wise separable convolution and a simple attention module. Experimental results demonstrate that the proposed learning strategy improves false reject rate by $40.31%$ at $1%$ false accept rate on the internal dataset, compared to the strongest baseline without using adversarial examples. Our best-performing system achieves $98.06%$ accuracy on the Google Speech Commands V1 dataset.

cs.SD↗

arXiv2024

Non-electrostatic interactions as random fields in charged liquids

We derive an equation capable of treating both the electrostatic and non-electrostatic interactions in the charged liquids. The equation derived is a complex version of the Poisson-Boltzmann equation, in which the non-electrostatic interactions are transformed to complex random fields by the field theory. Thus, the non-electrostatic interactions of the ions in the charged liquids can be simulated easily by generating random numbers according to the random fields. To solve the equation, the finite element method has been applied. The non-electrostatic interactions treated in the equation are general. In this study, we take the steric effect of ions as an example. Results show that the steric effect can be clearly caught by the equation.

cond-mat.soft↗

arXiv2024

Lightweight Conceptual Dictionary Learning for Text Classification Using Information Compression

We propose a novel, lightweight supervised dictionary learning framework for text classification based on data compression and representation. This two-phase algorithm initially employs the Lempel-Ziv-Welch (LZW) algorithm to construct a dictionary from text datasets, focusing on the conceptual significance of dictionary elements. Subsequently, dictionaries are refined considering label data, optimizing dictionary atoms to enhance discriminative power based on mutual information and class distribution. This process generates discriminative numerical representations, facilitating the training of simple classifiers such as SVMs and neural networks. We evaluate our algorithm's information-theoretic performance using information bottleneck principles and introduce the information plane area rank (IPAR) as a novel metric to quantify the information-theoretic performance. Tested on six benchmark text datasets, our algorithm competes closely with top models, especially in limited-vocabulary contexts, using significantly fewer parameters. \review{Our algorithm closely matches top-performing models, deviating by only ~2\% on limited-vocabulary datasets, using just 10\% of their parameters. However, it falls short on diverse-vocabulary datasets, likely due to the LZW algorithm's constraints with low-repetition data. This contrast highlights its efficiency and limitations across different dataset types.

cs.CL↗

arXiv2024

Approaching Maximum Likelihood Decoding Performance via Reshuffling ORBGRAND

Guessing random additive noise decoding (GRAND) is a recently proposed decoding paradigm particularly suitable for codes with short length and high rate. Among its variants, ordered reliability bits GRAND (ORBGRAND) exploits soft information in a simple and effective fashion to schedule its queries, thereby allowing efficient hardware implementation. Compared with maximum likelihood (ML) decoding, however, ORBGRAND still exhibits noticeable performance loss in terms of block error rate (BLER). In order to improve the performance of ORBGRAND while still retaining its amenability to hardware implementation, a new variant of ORBGRAND termed RS-ORBGRAND is proposed, whose basic idea is to reshuffle the queries of ORBGRAND so that the expected number of queries is minimized. Numerical simulations show that RS-ORBGRAND leads to noticeable gains compared with ORBGRAND and its existing variants, and is only 0.1dB away from ML decoding, for BLER as low as $10^{-6}$.

cs.IT↗

arXiv2024

FADI-AEC: Fast Score Based Diffusion Model Guided by Far-end Signal for Acoustic Echo Cancellation

Despite the potential of diffusion models in speech enhancement, their deployment in Acoustic Echo Cancellation (AEC) has been restricted. In this paper, we propose DI-AEC, pioneering a diffusion-based stochastic regeneration approach dedicated to AEC. Further, we propose FADI-AEC, fast score-based diffusion AEC framework to save computational demands, making it favorable for edge devices. It stands out by running the score model once per frame, achieving a significant surge in processing efficiency. Apart from that, we introduce a novel noise generation technique where far-end signals are utilized, incorporating both far-end and near-end signals to refine the score model's accuracy. We test our proposed method on the ICASSP2023 Microsoft deep echo cancellation challenge evaluation dataset, where our method outperforms some of the end-to-end methods and other diffusion based echo cancellation methods.

eess.AS↗