Think Shallow, Solve Deep: Controlling Recurrent Dynamics for Reliable Test-Time Depth
研究通过控制递归动态来解决测试时深度问题，提出了一种基于有限时间动态机制的方法以确保增加迭代次数不会降低答案质量。
0 citationsRead paper
Selected work
Representative Papers
The Null Token Knows: Reducing Message-Free Hallucination in ASR and NMT
This study addresses the issue of fluent hallucinations generated by ASR and NMT models in the absence of valid input, proposing the native null token as a diagnostic lens. Through auditing null token scores, scalar logit offsets, and decoder state probing, we demonstrate that the null token encodes valid abstention signals. We advocate evaluating abstention strategies by jointly considering suppression efficacy and deletion costs. Experiments indicate that while boosting null token scores significantly mitigates hallucinations, it necessitates balancing the risk of erroneously deleting legitimate content. Consequently, this work establishes a novel paradigm for alleviating no-signal hallucinations that effectively reconciles safety with practical utility.
0 citationsRead paper
LLM-Guided Prompt Evolution for Password Guessing
This work addresses the limitations of traditional password guessing approaches, which poorly emulate real-world attacker behavior, and existing large language model (LLM)-based methods that rely heavily on handcrafted prompts. The authors propose OpenEvolve, a novel system that introduces LLM-guided prompt evolution into password guessing for the first time, integrating MAP-Elites quality-diversity search with island-based population evolution to enable fully automated, human-intervention-free prompt optimization. Evaluated on RockYou-derived test sets, OpenEvolve significantly improves cracking rates from 2.02% to 8.48% and generates passwords whose character distributions more closely mirror those of real users. Experiments employing Qwen3-8B (local), Gemini-2.5 Flash (cloud), and an ensemble configuration demonstrate consistent attack performance gains across models, substantially enhancing the effectiveness of LLM-driven password auditing.
0 citationsRead paper
Signal Constellations with Enhanced Energy Efficiency for High-Speed Communication Systems
This work proposes a novel multidimensional signal constellation, termed SCOPT, to enhance the energy efficiency of high-speed communication systems without increasing transmit power or employing additional coding. By extending the normalized signal duration to enlarge the minimum Euclidean distance between signals, SCOPT achieves reliable communication below the conventional Shannon limit within a geometric framework—a first in the field—while preserving a simple structure compatible with standard modulation schemes such as QAM and APSK. Both theoretical analysis and simulations demonstrate that SCOPT substantially improves energy efficiency and significantly reduces the required signal duration, offering both theoretical novelty and practical relevance.
0 citationsRead paper
Evaluating Generalization and Robustness in Russian Anti-Spoofing: The RuASD Initiative
This work addresses the lack of reproducible and robust evaluation benchmarks for Russian speech anti-spoofing. To this end, we introduce RuASD, a novel dataset comprising spoofed utterances generated by 37 Russian text-to-speech and voice cloning systems alongside multi-source genuine speech. The dataset incorporates controlled channel distortions—including reverberation, additive noise/music, and codec transcoding—to simulate realistic distribution shifts encountered in practical deployments. Using this benchmark, we systematically evaluate the performance of lightweight supervised models, graph attention networks, self-supervised learning (SSL) detectors, and large-scale pre-trained systems under both clean and perturbed conditions. RuASD constitutes the first large-scale, multi-source, perturbation-controlled, and reproducible benchmark for Russian anti-spoofing, offering comprehensive insights into the generalization and robustness of current approaches. The dataset is publicly released to foster further research.
0 citationsRead paper
Recent publications
Latest Papers
Think Shallow, Solve Deep: Controlling Recurrent Dynamics for Reliable Test-Time Depth
研究通过控制递归动态来解决测试时深度问题，提出了一种基于有限时间动态机制的方法以确保增加迭代次数不会降低答案质量。
0 citationsRead paper
The Null Token Knows: Reducing Message-Free Hallucination in ASR and NMT
This study addresses the issue of fluent hallucinations generated by ASR and NMT models in the absence of valid input, proposing the native null token as a diagnostic lens. Through auditing null token scores, scalar logit offsets, and decoder state probing, we demonstrate that the null token encodes valid abstention signals. We advocate evaluating abstention strategies by jointly considering suppression efficacy and deletion costs. Experiments indicate that while boosting null token scores significantly mitigates hallucinations, it necessitates balancing the risk of erroneously deleting legitimate content. Consequently, this work establishes a novel paradigm for alleviating no-signal hallucinations that effectively reconciles safety with practical utility.
0 citationsRead paper
LLM-Guided Prompt Evolution for Password Guessing
This work addresses the limitations of traditional password guessing approaches, which poorly emulate real-world attacker behavior, and existing large language model (LLM)-based methods that rely heavily on handcrafted prompts. The authors propose OpenEvolve, a novel system that introduces LLM-guided prompt evolution into password guessing for the first time, integrating MAP-Elites quality-diversity search with island-based population evolution to enable fully automated, human-intervention-free prompt optimization. Evaluated on RockYou-derived test sets, OpenEvolve significantly improves cracking rates from 2.02% to 8.48% and generates passwords whose character distributions more closely mirror those of real users. Experiments employing Qwen3-8B (local), Gemini-2.5 Flash (cloud), and an ensemble configuration demonstrate consistent attack performance gains across models, substantially enhancing the effectiveness of LLM-driven password auditing.
0 citationsRead paper
Signal Constellations with Enhanced Energy Efficiency for High-Speed Communication Systems
This work proposes a novel multidimensional signal constellation, termed SCOPT, to enhance the energy efficiency of high-speed communication systems without increasing transmit power or employing additional coding. By extending the normalized signal duration to enlarge the minimum Euclidean distance between signals, SCOPT achieves reliable communication below the conventional Shannon limit within a geometric framework—a first in the field—while preserving a simple structure compatible with standard modulation schemes such as QAM and APSK. Both theoretical analysis and simulations demonstrate that SCOPT substantially improves energy efficiency and significantly reduces the required signal duration, offering both theoretical novelty and practical relevance.
0 citationsRead paper
Evaluating Generalization and Robustness in Russian Anti-Spoofing: The RuASD Initiative
This work addresses the lack of reproducible and robust evaluation benchmarks for Russian speech anti-spoofing. To this end, we introduce RuASD, a novel dataset comprising spoofed utterances generated by 37 Russian text-to-speech and voice cloning systems alongside multi-source genuine speech. The dataset incorporates controlled channel distortions—including reverberation, additive noise/music, and codec transcoding—to simulate realistic distribution shifts encountered in practical deployments. Using this benchmark, we systematically evaluate the performance of lightweight supervised models, graph attention networks, self-supervised learning (SSL) detectors, and large-scale pre-trained systems under both clean and perturbed conditions. RuASD constitutes the first large-scale, multi-source, perturbation-controlled, and reproducible benchmark for Russian anti-spoofing, offering comprehensive insights into the generalization and robustness of current approaches. The dataset is publicly released to foster further research.
0 citationsRead paper