Keyword spotting (KWS) has become a widely used front-end task in voice-interface systems, where compact speech models detect predefined commands before downstream speech processing is activated. However, the increasing use of outsourced data, pretrained checkpoints, and third-party training pipelines exposes KWS models to backdoor threats. Existing repair methods are largely inherited from visual-domain defenses or rely on broad fine-tuning signals, making it difficult to suppress diverse audio backdoors while preserving benign keyword recognition. We propose a competition-guided local repair framework that constructs waveform-domain probes to expose abnormal true-versus-non-target competition and suppresses this competition by updating only a late-local repair region. The repair uses a small clean set and does not require trigger or target-label knowledge. Extensive experiments across seven audio backdoor attacks and diverse KWS model families on two KWS corpora, supplemented by a larger-vocabulary Speech Commands setting, show that the proposed method substantially reduces attack success rate (ASR) while preserving clean accuracy (CA) over representative repair baselines. Across the main SCD-10 and FKD-14 evaluation settings, the proposed method reduces average ASR from 16.6% to 8.4% relative to the strongest audio-domain repair baseline, while improving CA by 5.0 percentage points, with particularly clear gains on learned and concealed trigger attacks. The repaired model preserves the original inference architecture without adding any inference-time branch.
Reinforcement Learning from Human Feedback (RLHF) increasingly relies on listwise preference data, yet most methods formulate preference alignment as a fine-grained ranking problem, which mismatches the coarse-grained distinctions (e.g., acceptable vs. unacceptable) of reliable human judgment. We show that this mismatch can induce contradictory gradient signals in Plackett-Luce (PL)-based objectives, where intermediate responses receive opposing signals across decomposed sub-tasks. To address this, we propose Preference Classification Optimization (PCO), which reframes alignment as a groupwise classification task. By leveraging Preference Group Data—partitioning responses into positive and negative sets—PCO combines the contrastive stability of Bradley-Terry models with listwise parallel efficiency, without enforcing a total order over candidates. Gradient analysis characterizes how PCO can mitigate ranking-induced conflicts. Experiments on Anthropic HH, Summarize From Feedback, and UltraInteract Pair show favorable alignment performance relative to the evaluated baselines, together with additional evidence on training stability, safety compliance, and out-of-distribution generalization.
Neural vocoders have achieved remarkable progress in high-fidelity waveform generation in recent years, but most existing methods still rely on full-precision operators, leading to non-negligible storage and computational costs during inference. Binary neural networks (BNNs) provide an alternative direction by quantizing weights and activations to 1 bit. However, directly applying binarization to neural vocoders is rather challenging as it inherently incurs severe information loss. To this end, we propose BiVoC, a BNN-oriented time-frequency neural vocoder for high-fidelity waveform generation. Building on our previous full-precision RNDVoC framework, BiVoC adapts the inverse-problem formulation, range-null decomposition, and fine-to-coarse non-uniform subband modeling to the 1-bit setting. Specifically, we develop a simple binarized convolutional unit to provide stable time-frequency feature modeling under binary representations. Furthermore, we analyze the quality-complexity sensitivity of different modules and propose a mixed-precision design to retain quality-sensitive components in full precision while binarizing computation-dominant parts. Extensive experiments are conducted on LibriTTS and multiple out-of-distribution benchmarks. Both objective and subjective results demonstrate that BiVoC substantially outperforms BNN variants obtained by directly binarizing existing vocoders, while achieving competitive performance against several mainstream full-precision vocoder baselines under much lower theoretical computational cost. These results demonstrate the promising potential of BNNs in the audio generation field. Codes and pre-trained weights are available at https://github.com/Andong-Li-speech/BiVoC.
Graph Neural Networks (GNNs) have demonstrated significant advancements in social media event detection (SMED); however, critical limitations persist in multilingual environments, particularly concerning knowledge transfer and semantic alignment among different languages. This study introduces Multilingual Social Media Event Detection (MSMED), a novel framework that integrates cross-lingual knowledge transfer and strict sampling strategies for SMED. The proposed framework first constructs a multilingual message graph with message attributes, semantic similarity, and Wikipedia QID. Then, we use random walk algorithms to systematically traverse multilingual message graphs, facilitating the generation of unified message representations. Meanwhile, a novel strict angular Constraint sampling mechanism is introduced to distinguish semantically identical events in different languages, thereby improving the training of the SMED model. Extensive experiments are conducted on the Event2012 benchmark. MSMED obtains an average of 9.2% improvement over baseline models in SMED. Meanwhile, our method obtains a Normalized Mutual Information (NMI) score of 0.72 under trilingual mixed-input conditions. We give a comprehensive analysis of the experimental results. Our code and data will be released on GitHub.
Current role-playing dialogue agents (RPAs) are typically constructed by imitating surface-level behaviors, but this approach lacks internal cognitive consistency, often causing out-of-character errors in complex situations. To address this, we propose Character-R1, a framework designed to provide comprehensive verifiable reward signals for effective role-aware reasoning, which are missing in recent studies. Specifically, our framework comprises three core designs: (1) Cognitive Focus Reward, which enforces explicit label-based analysis of 10 character elements (e.g., worldview) to structure internal cognition; (2) Reference-Guided Reward, which utilizes overlap-based metrics with reference responses as optimization anchors to enhance exploration and performance; and (3) Character-Conditioned Reward Normalization, which adjusts reward distributions based on character categories to ensure robust optimization across heterogeneous roles. Extensive experiments indicate that Character-R1 significantly outperforms role-playing baselines of comparable size in knowledge, memory and others.
