arXiv subjects

Cheng Luo

Publications and source records attributed to Cheng Luo.

At least 19 recordsLinked to original sources

arXiv2026

Shaping Delay-Doppler Ambiguity in Practical OFDM-ISAC

Orthogonal frequency-division multiplexing (OFDM) is a key waveform for integrated sensing and communication (ISAC). Existing OFDM ambiguity analyses, however, typically assume fully occupied data-only waveforms, whereas practical frames contain direct-current and edge-guard nulls, fixed pilots, and random payload symbols. This mixed resource structure reshapes the self-ambiguity function and induces prominent sidelobes in the sensing region of interest (ROI). We therefore propose ROI-oriented deep block-unitary precoded OFDM (DBU-OFDM), which combines resource-specific trainable unitary transformations with dedicated sensing subcarriers to suppress ROI sidelobes while preserving the prescribed resource support. We develop a constraint-preserving parameterization capable of representing arbitrary unitary matrices and an ROI-aware sensing-support initialization. We prove periodic autocorrelation function (P-ACF) invariance under phase-only optimization and zero-delay Doppler-cut invariance under arbitrary unitary transformations. Numerical results corroborate the P-ACF optimality of cyclic-prefix OFDM (CP-OFDM) and provide numerical support for the conjecture that conventional OFDM is globally optimal in the considered aperiodic autocorrelation function (A-ACF) setting. Unitary pilot optimization improves the ROI peak-to-sidelobe ratio by over 2 dB, while combining it with dedicated sensing subcarriers yields gains ranging from several to tens of dB. Under the considered configuration, the A-ACF mode also provides substantially greater ROI ambiguity-shaping capability than the P-ACF mode.

eess.SP↗

arXiv2026

Soft Posterior Speaker Injection for Multi-Talker Speech Recognition

Multi-talker automatic speech recognition (MT-ASR) remains challenging in the presence of overlapping speech. Hard segmentation introduces irreversible errors, whereas serialized output training (SOT) avoids explicit segmentation but does not condition a pretrained encoder on speaker activity. We propose Soft Posterior Speaker Injection (SPSI). A Soft Posterior Head predicts per-frame speaker posteriors $\hat{\mathbf{P}}$ and injects them into Whisper through Multi-layer Feature-wise Linear Modulation (MFLM) and Speaker Memory Prompts (SMP). The benefit of SPSI is largest where overlap is heaviest and under domain transfer. On controlled two-speaker LibriSpeech overlap, SPSI reduces concatenated minimum-permutation word error rate (cpWER) from $61.5\%$ to $60.0\%$ in the high-overlap bin, and from $51.9\%$ to $51.0\%$ on the full set, relative to SOT. Same-backbone speaker-auxiliary objectives, voice activity detection (VAD) pipelines, and a diarization-conditioned Whisper replica do not outperform SOT. Freeze-posterior overlap-heavy adaptation reduces held-out LibriCSS cpWER from $42.3\%$ to $36.8\%$ on sessions $8$--$9$, a $5.5$-point gain over SOT. The source code is available at https://github.com/HackerHyper/SPSI.git.

cs.SD↗

arXiv2026

Preserving General Capabilities during Domain Specialization with Uncertainty-Calibrated MOPD

Specializing large language models to vertical domains improves domain-specific behavior but often degrades general capabilities such as reasoning, coding, instruction following, and creative writing. We study this domain--general trade-off in Multi-Teacher On-Policy Distillation (MOPD), where a specialized student is supervised on its own sampled trajectories by domain and general teachers. Standard MOPD faces two limitations: ordinary on-policy sampling rarely exposes tokens with large positive teacher--student advantages, while the advantage sign alone does not establish whether the resulting update direction is reliable. We propose uncertainty-calibrated MOPD to address these limitations. Dual-temperature sampling broadens the candidate trajectory pool, and positive-advantage-density filtering selects trajectories with stronger positive learning signals. Centered log-likelihood (CLL) filtering then computes an entropy-calibrated teacher-endorsement score and probabilistically retains token updates according to direction--endorsement consistency. Experiments on role-playing and medical-domain specialization show that our method improves the general-capability average over standard MOPD by $4.73\%$ and $10.84\%$, respectively, while maintaining vertical-domain performance. Ablations and diagnostic analyses further confirm that the gains do not merely result from a larger rollout budget and that the proposed trajectory- and token-level mechanisms address their intended failure modes.

cs.CL↗

arXiv2026

Multi-Head Attention Residuals

Transformers propagate information across depth through a single additive residual stream: every sublayer reads only the most recent state. Attention residuals relax this by letting each sublayer attend, through a learned softmax. However, that read uses a single query shared across the entire width, so every feature subspace must read the depth history through one distribution. The cost of this forced compromise grows with how much the subspaces disagree about which layers to read, and disagreement grows with model width. We introduce Multi-Head Attention Residuals (MHAR): the routing query is reshaped into H per-subspace heads, each with its own softmax over the depth history. The read becomes block-diagonal, the reshape adds zero parameters and negligible compute, and H = 1 recovers attention residuals exactly. Trained from scratch on a deduplicated Nemotron-based anneal corpus that is quality-filtered and STEM- and code-heavy, MHAR improves validation loss over a standard Transformer at 100M, 350M, and 1B (-0.061, -0.149, and -0.140). It achieves the best result among four methods in every setting, with the gain increasing from 100M to the larger scales. The head count is a real design axis rather than a free knob: validation loss is U-shaped with respect to H, with a flat optimum at H = 4 or H = 8 across scales. We adopt H = 8 for large-scale models; over-splitting beyond this point (H = 16) consistently gives back part of the gain. A direct probe of the trained queries confirms that learned subspace disagreement is the underlying driver. Fused Triton routing kernels increase attention-residual training throughput from 0.2-0.5x to 0.55-0.88x of the baseline while maintaining near-baseline peak memory. An identity-preserving conversion using delta attention residuals supports 8B mid-training, yielding improvements of +3.2 on GSM8K and +3.1 on GPQA.

cs.AI↗

arXiv2026

KAT-Coder-V2.5 Technical Report

We present KAT-Coder-V2.5, a coding-focused agentic model trained to act autonomously inside real, executable repositories rather than as a single-turn code generator. Its capability is bottlenecked less by model scale than by the scarcity of reproducible environments, verifiable rewards, and high-value trajectories, which we address with an end-to-end agentic post-training framework. AutoBuilder reconstructs multilingual repositories into sandboxed environments with fail-to-pass and pass-to-pass verification at scale, from which we regenerate self-contained task specifications, recover near-miss trajectories, and distill supervision through process-aware filtering, while KwaiClawEnv synthesizes large-scale tool-use trajectories from executable services and real task seeds. We further scale reinforcement learning with harness randomization, a reliability-hardened sandbox, an asymmetric actor--critic PPO with hindsight-augmented value estimation, and a harness-oriented reward framework, and unify SWE, Agent-Claw, and WebCoding experts via Multi-Teacher On-Policy Distillation. Across six software-engineering and agentic benchmarks, KAT-Coder-V2.5 delivers the best agentic tool-use result on PinchBench and ranks second only to the frontier Opus 4.8 on repository-level software engineering. Our service is available at https://streamlake.com/product/kat-coder.

cs.SE↗

arXiv2026

REACT 2026: The Fourth Multiple Appropriate Facial Reaction Generation Challenge: Personalised MAFRG and Appropriate EEG Reaction Prediction

In dyadic interactions, various human facial reactions could be appropriate for responding to each human speaker behaviour. Following the successful organisation of the REACT 2023, 2024 and 2025 challenge series, a body of generative deep learning (DL) models have been developed for the problem of multiple appropriate facial reaction generation (MAFRG). This year, we propose the REACT 2026 challenge encouraging the development and benchmarking of Machine Learning (ML) models that can generate multiple personalised, appropriate, diverse, realistic and synchronised human-style facial reactions expressed by a specific human listener for responding to each given speaker behaviour. As a key of the challenge, we continuously provide challenge participants with MARS dataset introduced by REACT 2025 but additionally provide individual-level Big-Five personality labels and EEG recordings. This introduces a new one-to-many personalised facial reaction generation setting combining human expressive behavioural, affective and neurophysiological signals, which remains largely unexplored in current dyadic interaction modelling. This paper also presents the challenge guidelines and new baselines on the four proposed sub-challenges: Offline generic and personalised MAFRG as well as Online generic and personalised MAFRG, respectively, which are publicly available at https://github.com/reactmultimodalchallenge/baseline_react2026.

cs.CV↗

arXiv2026

When BBR Meets Live Streaming

Recently, industrial pioneers like Amazon, Tencent, ByteDance, and Huawei have been adopting BBR as their congestion control algorithm for live-streaming applications, including TikTok Live. However, BBR, originally crafted for bulk data transmission, faces multiple challenges in live-streaming scenarios. In this paper, we first explore two key issues associated with BBR due to inaccurate bandwidth estimation in live-streaming scenarios: (i) BBR cannot easily exit its startup phase, resulting in a fierce self-inflicted loss. (ii) BBR sends data at a lower rate than the available bandwidth during its stable phase. We then propose BBR-Copilot, an auxiliary congestion control component that cooperates with BBR, making BBR better adapt to live-streaming scenarios. BBR-Copilot allows for proactively generating accurate bandwidth measurement samples by smartly creating and sending extra data. We implement the BBR-Copilot prototype upon QUIC and evaluate it via testbed. Experimental evaluation results show that BBR-Copilot effectively enhances BBR's performance in live-streaming scenarios.

eess.IV↗

arXiv2026

Delta Attention Residuals

Attention Residuals replace standard additive residual connections with learned softmax attention over previous layer outputs, enabling selective cross-layer routing. However, standard Attention Residuals still attend over cumulative hidden states in previous layers, which are highly redundant. We show that this redundancy leads to routing collapse in deeper layers: attention weights become low-contrast and closer to uniform (max weight ${\approx}$0.2), limiting the model's ability to select informative states in previous layers. This raises a key but underexplored design question: what layer-wise representations should be routed in Attention Residuals? To answer this question, we propose Delta Attention Residuals, which attend over deltas -- the change introduced by each sublayer ($\mathbf{v}_i = \mathbf{h}_{i+1} - \mathbf{h}_i$) -- instead of cumulative states. Delta representations are structurally diverse and yield higher-contrast attention distributions (max weight ${\approx}$0.6), enabling more selective and effective routing across layers. This principle applies at both per-sublayer and block granularity. Across all tested scales (220M--7.6B), Delta Attention Residuals consistently outperform both standard residuals and Attention Residuals, with 1.7--8.2\% validation perplexity gains. Delta Attention Residuals also enables converting pretrained checkpoints into Delta Attention Residuals via standard fine-tuning. Code is available at https://github.com/wdlctc/delta-attention-residuals-code.

cs.LG↗

arXiv2026

DBU-OFDM: A Trainable Deep Block-Unitary OFDM Waveform for Integrated Sensing and Communication

Orthogonal frequency-division multiplexing (OFDM) is a dominant waveform in modern wireless systems, yet its high peak-to-average power ratio (PAPR) and limited adaptability hinder efficient support for integrated communication and sensing. This paper proposes deep block-unitary precoded OFDM (DBU-OFDM), a structure-preserving learning framework that enables trainable waveform adaptation while preserving the DFT-based signal structure, pilot/null resource protection, and compatibility with low-complexity frequency-domain equalization. The proposed design restricts learning to a block-unitary transformation over data subcarriers and preserves pilot and null resources for structural compatibility. The transform is parameterized by recursive Householder reflections, ensuring strict unitarity as well as differentiable, numerically stable, and complexity-controllable implementation. Results show that DBU-OFDM achieves PAPR tails close to block-pilot DFT-s-OFDM while retaining comb-type pilots, improves communication reliability in frequency-selective fading via frequency-domain diversity, and enhances range and velocity estimation in direct sensing, especially in dimension-limited settings. Over-the-air USRP experiments and FPGA prototyping further verify its practical feasibility, demonstrating low error vector magnitude (EVM), clear PAPR reduction in real transmission, and hardware throughput up to 200~MS/s with microsecond-level latency. DBU-OFDM therefore offers a practical intermediate solution between conventional model-based OFDM waveforms and unconstrained neural transceivers for next-generation integrated communication and sensing systems.

eess.SP↗

arXiv2026

HAFM: Hierarchical Autoregressive Foundation Model for Music Accompaniment Generation

Music accompaniment generation aims to automatically produce instrumental accompaniments that are rhythmically, harmonically, and timbrally coherent with a given vocal input, with broad applications in personalized music creation, arrangement assistance, and music education. Existing approaches, primarily operating in the symbolic domain or relying on single-stage audio generation frameworks, commonly suffer from insufficient high-level semantic structure modeling, limited acoustic detail reconstruction, and weak conditional controllability. To address these limitations, this paper proposes HAFM, a Hierarchical Autoregressive Foundation Model for vocal-conditioned music accompaniment generation. The model employs a dual-rate tokenization strategy in which $50$ Hz HuBERT semantic tokens capture high-level musical structure and $75$ Hz EnCodec acoustic tokens encode fine-grained acoustic content, enabling explicit disentanglement of semantic and acoustic representations. Building on this foundation, a three-stage cascaded generation framework is designed to progressively generate semantic tokens, coarse acoustic tokens, and fine acoustic tokens, refining the accompaniment from global structure to local detail. . Objective evaluation on the MUSDB18 dataset demonstrates that the full three-stage model achieves a Fr{\'e}chet Audio Distance (FAD) score of 1.71, representing an 18.6% relative improvement over the two-stage baseline (FAD = 2.10). Subjective listening tests show that the generated accompaniments achieve a 51.5% preference rate against ground-truth accompaniments in head-to-head comparisons, and substantially outperform the random baseline in terms of rhythmic alignment, harmonic compatibility, and overall musical coherence. The source code and demo are available at https://github.com/HackerHyper/HAFM.git.

cs.SD↗

arXiv2026

When Do Hallucinations Arise? A Graph Perspective on the Evolution of Path Reuse and Path Compression

Reasoning hallucinations in large language models (LLMs) often appear as fluent yet unsupported conclusions that violate either the given context or underlying factual knowledge. Although such failures are widely observed, the mechanisms by which decoder-only Transformers produce them remain poorly understood. We model next-token prediction as a graph search process over an underlying graph, where entities correspond to nodes and learned transitions form edges. From this perspective, contextual reasoning is a constrained search over a sampled subgraph (intrinsic reasoning), while context-free queries rely on memorized structures in the underlying graph (extrinsic reasoning). We show that reasoning hallucinations arise from two fundamental mechanisms: \textbf{Path Reuse}, where memorized knowledge overrides contextual constraints during early training, and \textbf{Path Compression}, where frequently traversed multi-step paths collapse into shortcut edges in later training. Together, these mechanisms provide a unified explanation for reasoning hallucinations in LLMs and connected to well-known behaviors observed in downstream applications.

cs.AI↗

arXiv2026

ReactMotion: Generating Reactive Listener Motions from Speaker Utterance

In this paper, we introduce a new task, Reactive Listener Motion Generation from Speaker Utterance, which aims to generate naturalistic listener body motions that appropriately respond to a speaker's utterance. However, modeling such nonverbal listener behaviors remains underexplored and challenging due to the inherently non-deterministic nature of human reactions. To facilitate this task, we present ReactMotionNet, a large-scale dataset that pairs speaker utterances with multiple candidate listener motions annotated with varying degrees of appropriateness. This dataset design explicitly captures the one-to-many nature of listener behavior and provides supervision beyond a single ground-truth motion. Building on this dataset design, we develop preference-oriented evaluation protocols tailored to evaluate reactive appropriateness, where conventional motion metrics focusing on input-motion alignment ignore. We further propose ReactMotion, a unified generative framework that jointly models text, audio, emotion, and motion, and is trained with preference-based objectives to encourage both appropriate and diverse listener responses. Extensive experiments show that ReactMotion outperforms retrieval baselines and cascaded LLM-based pipelines, generating more natural, diverse, and appropriate listener motions.

cs.CV↗

arXiv2026

MonarchRT: Efficient Attention for Real-Time Video Generation

Real-time video generation with Diffusion Transformers is bottlenecked by the quadratic cost of 3D self-attention, especially in real-time regimes that are both few-step and autoregressive, where errors compound across time and each denoising step must carry substantially more information. In this setting, we find that prior sparse-attention approximations break down, despite showing strong results for bidirectional, many-step diffusion. Specifically, we observe that video attention is not reliably sparse, but instead combines pronounced periodic structure driven by spatiotemporal position with dynamic, sparse semantic correspondences and dense mixing, exceeding the representational capacity of even oracle top-k attention. Building on this insight, we propose Monarch-RT, a structured attention parameterization for video diffusion models that factorizes attention using Monarch matrices. Through appropriately aligned block structure and our extended tiled Monarch parameterization, we achieve high expressivity while preserving computational efficiency. We further overcome the overhead of parameterization through finetuning, with custom Triton kernels. We first validate the high efficacy of Monarch-RT over existing sparse baselines designed only for bidirectional models. We further observe that Monarch-RT attains up to 95% attention sparsity with no loss in quality when applied to the state-of-the-art model Self-Forcing, making Monarch-RT a pioneering work on highly-capable sparse attention parameterization for real-time video generation. Our optimized implementation outperforms FlashAttention-2, FlashAttention-3, and FlashAttention-4 kernels on Nvidia RTX 5090, H100, and B200 GPUs respectively, providing kernel speedups in the range of 1.4-11.8X. This enables us, for the first time, to achieve true real-time video generation with Self-Forcing at 16 FPS on a single RTX 5090.

cs.CV↗

arXiv2026

Multi-Source Retrieval and Reasoning for Legal Sentencing Prediction

Legal judgment prediction (LJP) aims to predict judicial outcomes from case facts and typically includes law article, charge, and sentencing prediction. While recent methods perform well on the first two subtasks, legal sentencing prediction (LSP) remains difficult due to its need for fine-grained objective knowledge and flexible subjective reasoning. To address these limitations, we propose $MSR^2$, a framework that integrates multi-source retrieval and reasoning in LLMs with reinforcement learning. $MSR^2$ enables LLMs to perform multi-source retrieval based on reasoning needs and applies a process-level reward to guide intermediate subjective reasoning steps. Experiments on two real-world datasets show that $MSR^2$ improves both accuracy and interpretability in LSP, providing a promising step toward practical legal AI. Our code is available at https://github.com/cjj826/MSR2.

cs.IR↗

arXiv2026

LegalOne: A Family of Foundation Models for Reliable Legal Reasoning

While Large Language Models (LLMs) have demonstrated impressive general capabilities, their direct application in the legal domain is often hindered by a lack of precise domain knowledge and complexity of performing rigorous multi-step judicial reasoning. To address this gap, we present LegalOne, a family of foundational models specifically tailored for the Chinese legal domain. LegalOne is developed through a comprehensive three-phase pipeline designed to master legal reasoning. First, during mid-training phase, we propose Plasticity-Adjusted Sampling (PAS) to address the challenge of domain adaptation. This perplexity-based scheduler strikes a balance between the acquisition of new knowledge and the retention of original capabilities, effectively establishing a robust legal foundation. Second, during supervised fine-tuning, we employ Legal Agentic CoT Distillation (LEAD) to distill explicit reasoning from raw legal texts. Unlike naive distillation, LEAD utilizes an agentic workflow to convert complex judicial processes into structured reasoning trajectories, thereby enforcing factual grounding and logical rigor. Finally, we implement a Curriculum Reinforcement Learning (RL) strategy. Through a progressive reinforcement process spanning memorization, understanding, and reasoning, LegalOne evolves from simple pattern matching to autonomous and reliable legal reasoning. Experimental results demonstrate that LegalOne achieves state-of-the-art performance across a wide range of legal tasks, surpassing general-purpose LLMs with vastly larger parameter counts through enhanced knowledge density and efficiency. We publicly release the LegalOne weights and the LegalKit evaluation framework to advance the field of Legal AI, paving the way for deploying trustworthy and interpretable foundation models in high-stakes judicial applications.

cs.CL↗

arXiv2025

ArcGen: Generalizing Neural Backdoor Detection Across Diverse Architectures

Backdoor attacks pose a significant threat to the security and reliability of deep learning models. To mitigate such attacks, one promising approach is to learn to extract features from the target model and use these features for backdoor detection. However, we discover that existing learning-based neural backdoor detection methods do not generalize well to new architectures not seen during the learning phase. In this paper, we analyze the root cause of this issue and propose a novel black-box neural backdoor detection method called ArcGen. Our method aims to obtain architecture-invariant model features, i.e., aligned features, for effective backdoor detection. Specifically, in contrast to existing methods directly using model outputs as model features, we introduce an additional alignment layer in the feature extraction function to further process these features. This reduces the direct influence of architecture information on the features. Then, we design two alignment losses to train the feature extraction function. These losses explicitly require that features from models with similar backdoor behaviors but different architectures are aligned at both the distribution and sample levels. With these techniques, our method demonstrates up to 42.5% improvements in detection performance (e.g., AUC) on unseen model architectures. This is based on a large-scale evaluation involving 16,896 models trained on diverse datasets, subjected to various backdoor attacks, and utilizing different model architectures. Our code is available at https://github.com/SeRAlab/ArcGen.

cs.LG↗

arXiv2025

An End-to-End Neural Network Transceiver Design for OFDM System with FPGA-Accelerated Implementation

The evolution toward sixth-generation (6G) wireless networks demands high-performance transceiver architectures capable of handling complex and dynamic environments. Conventional orthogonal frequency-division multiplexing (OFDM) receivers rely on cascaded discrete Fourier transform (DFT) and demodulation blocks, which are prone to inter-stage error propagation and suboptimal global performance. In this work, we propose two neural network (NN) models DFT-Net and Demodulation-Net (Demod-Net) to jointly replace the IDFT/DFT and demodulation modules in an OFDM transceiver. The models are trained end-to-end (E2E) to minimize bit error rate (BER) while preserving operator equivalence for hybrid deployment. A customized DFT-Demodulation Net Accelerator (DDNA) is further developed to efficiently map the proposed networks onto field-programmable gate array (FPGA) platforms. Leveraging fine-grained pipelining and block matrix operations, DDNA achieves high throughput and flexibility under stringent latency constraints. Experimental results show that the DL-based transceiver consistently outperforms the conventional OFDM system across multiple modulation schemes. With only a modest increase in hardware resource usage, it achieves approximately 1.5 dB BER gain and up to 66\% lower execution time.

eess.SY↗

arXiv2025

AutoRec: Accelerating Loss Recovery for Live Streaming in a Multi-Supplier Market

Due to the limited permissions for upgrading dualside (i.e., server-side and client-side) loss tolerance schemes from the perspective of CDN vendors in a multi-supplier market, modern large-scale live streaming services are still using the automatic-repeat-request (ARQ) based paradigm for loss recovery, which only requires server-side modifications. In this paper, we first conduct a large-scale measurement study with up to 50 million live streams. We find that loss shows dynamics and live streaming contains frequent on-off mode switching in the wild. We further find that the recovery latency, enlarged by the ubiquitous retransmission loss, is a critical factor affecting live streaming's client-side QoE (e.g., video freezing). We then propose an enhanced recovery mechanism called AutoRec, which can transform the disadvantages of on-off mode switching into an advantage for reducing loss recovery latency without any modifications on the client side. AutoRec allows users to customize overhead tolerance and recovery latency tolerance and adaptively adjusts strategies as the network environment changes to ensure that recovery latency meets user demands whenever possible while keeping overhead under control. We implement AutoRec upon QUIC and evaluate it via testbed and real-world commercial services deployments. The experimental results demonstrate the practicability and profitability of AutoRec.

cs.NI↗