Selected work

Representative Papers

Alignment Drift in Single-Model Speculative Decoding for ASR: Mechanism, Correction, and Cost

Aug 12, 2026

This work addresses a critical limitation in single-model speculative decoding for automatic speech recognition, where the draft module struggles to accurately track audio positions, leading to alignment drift and degraded prediction quality. The study reveals that precise audio position tracking is pivotal for effective speculation and proposes AnchorDraft, a training method that corrects such drift without altering the inference graph. AnchorDraft either leverages attention readout positions during verification or guides the draft module to implicitly learn positional information. Experiments demonstrate that AnchorDraft significantly accelerates end-to-end inference across two target model scales. Properly aligned audio windows substantially increase token acceptance rates, with the median error in verification-stage attention confined to merely two frames.

0 citationsRead paper

IHBench: Evaluating Post-Interruption Recovery in Voice Agents with Structured Workflows

Jun 17, 2026

This work addresses the absence of evaluation frameworks for assessing voice agents’ ability to recover from interruptions during structured multi-step tasks. The authors introduce IHBench, the first benchmark designed to systematically evaluate how well voice agents handle user interjections, avoid redundant actions, and correctly resume task execution. Built upon a state-machine-driven workflow, IHBench injects six types of interruptions across ten enterprise domains and automatically scores performance along two dimensions—task completion and recovery quality—with human validation. Evaluation of 27 audio-language model configurations reveals that closed-source models significantly outperform open-source counterparts in task completion, robustness over long conversations, and cross-modal consistency. Furthermore, the proposed LLM-based scorer demonstrates strong alignment with human judgments.

0 citationsRead paper

TARQ: Tail-Aware Reconstruction Quantization for Rare-Word Robust Automatic Speech Recognition

May 26, 2026

This work addresses the inadequate calibration of low-frequency words—such as proper nouns and numerals—in traditional post-training quantization for automatic speech recognition (ASR), which leads to high rare-word error rates. The authors propose a label-free post-training quantization framework that rebalances the calibration data distribution via a closed-form, layer-wise reweighting scheme termed rareBAL, thereby equalizing calibration quality between frequent and rare words. Additionally, they introduce a metric-consistent residual correction mechanism that operates without entity labels or additional training. By integrating data-aware quantization, closed-form layer weight adjustment, and residual correction, the method achieves efficient deployment under W4G128 settings, significantly reducing rare-word error rates across eight ASR backbones and six datasets, exhibiting minimal cross-corpus variance, and generalizing effectively to entity-dense benchmarks in fully unsupervised conditions.

0 citationsRead paper

Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training

May 12, 2026

Policy updates in reinforcement learning are highly sensitive to distributional shifts, a problem exacerbated in large-scale settings where discrepancies in numerical precision and sampling between training and inference introduce further instability. Existing approaches often rely on fixed hyperparameters, limiting their adaptability to variations in tasks, model scales, or data distributions. This work proposes a batch-adaptive policy optimization objective that dynamically modulates update intensity based on the effective sample size of policy ratios within each batch. By replacing fixed clipping with an adaptive mechanism grounded in the empirical distribution of ratios, the method jointly addresses trust-region constraints and off-policy data reliability without introducing additional hyperparameters. Empirical results demonstrate that the proposed approach matches or surpasses carefully tuned baselines across diverse settings, significantly enhancing algorithmic robustness and generalization.

0 citationsRead paper

The Pokémon Theorem and other Fairness Impossibility Results

May 09, 2026

This work uncovers a unified geometric origin underlying the incompatibility of multiple fairness criteria under unequal base rates. By modeling fairness constraints as linear conditions on conditional mean embeddings in a reproducing kernel Hilbert space (RKHS) and leveraging an over-determined analysis of the law of total expectation, it elucidates the fundamental nature of fairness conflicts. The paper introduces the “Pokémon Theorem,” proving that any finite set of linear mean-based fairness criteria inevitably entails residual violations and revealing the unavoidable class collapse in fair representation learning. Building on RKHS theory, maximum mean discrepancy (MMD), Kolmogorov m-width, and spectral regularization, the authors derive a signal–error frontier under approximate fairness relaxations, with experiments validating the theoretical bounds on standard fairness benchmarks.

0 citationsRead paper

Recent publications

Latest Papers

Alignment Drift in Single-Model Speculative Decoding for ASR: Mechanism, Correction, and Cost

Aug 12, 2026

This work addresses a critical limitation in single-model speculative decoding for automatic speech recognition, where the draft module struggles to accurately track audio positions, leading to alignment drift and degraded prediction quality. The study reveals that precise audio position tracking is pivotal for effective speculation and proposes AnchorDraft, a training method that corrects such drift without altering the inference graph. AnchorDraft either leverages attention readout positions during verification or guides the draft module to implicitly learn positional information. Experiments demonstrate that AnchorDraft significantly accelerates end-to-end inference across two target model scales. Properly aligned audio windows substantially increase token acceptance rates, with the median error in verification-stage attention confined to merely two frames.

0 citationsRead paper

IHBench: Evaluating Post-Interruption Recovery in Voice Agents with Structured Workflows

Jun 17, 2026

This work addresses the absence of evaluation frameworks for assessing voice agents’ ability to recover from interruptions during structured multi-step tasks. The authors introduce IHBench, the first benchmark designed to systematically evaluate how well voice agents handle user interjections, avoid redundant actions, and correctly resume task execution. Built upon a state-machine-driven workflow, IHBench injects six types of interruptions across ten enterprise domains and automatically scores performance along two dimensions—task completion and recovery quality—with human validation. Evaluation of 27 audio-language model configurations reveals that closed-source models significantly outperform open-source counterparts in task completion, robustness over long conversations, and cross-modal consistency. Furthermore, the proposed LLM-based scorer demonstrates strong alignment with human judgments.

0 citationsRead paper

TARQ: Tail-Aware Reconstruction Quantization for Rare-Word Robust Automatic Speech Recognition

May 26, 2026

This work addresses the inadequate calibration of low-frequency words—such as proper nouns and numerals—in traditional post-training quantization for automatic speech recognition (ASR), which leads to high rare-word error rates. The authors propose a label-free post-training quantization framework that rebalances the calibration data distribution via a closed-form, layer-wise reweighting scheme termed rareBAL, thereby equalizing calibration quality between frequent and rare words. Additionally, they introduce a metric-consistent residual correction mechanism that operates without entity labels or additional training. By integrating data-aware quantization, closed-form layer weight adjustment, and residual correction, the method achieves efficient deployment under W4G128 settings, significantly reducing rare-word error rates across eight ASR backbones and six datasets, exhibiting minimal cross-corpus variance, and generalizing effectively to entity-dense benchmarks in fully unsupervised conditions.

0 citationsRead paper

Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training

May 12, 2026

Policy updates in reinforcement learning are highly sensitive to distributional shifts, a problem exacerbated in large-scale settings where discrepancies in numerical precision and sampling between training and inference introduce further instability. Existing approaches often rely on fixed hyperparameters, limiting their adaptability to variations in tasks, model scales, or data distributions. This work proposes a batch-adaptive policy optimization objective that dynamically modulates update intensity based on the effective sample size of policy ratios within each batch. By replacing fixed clipping with an adaptive mechanism grounded in the empirical distribution of ratios, the method jointly addresses trust-region constraints and off-policy data reliability without introducing additional hyperparameters. Empirical results demonstrate that the proposed approach matches or surpasses carefully tuned baselines across diverse settings, significantly enhancing algorithmic robustness and generalization.

0 citationsRead paper

The Pokémon Theorem and other Fairness Impossibility Results

May 09, 2026

This work uncovers a unified geometric origin underlying the incompatibility of multiple fairness criteria under unequal base rates. By modeling fairness constraints as linear conditions on conditional mean embeddings in a reproducing kernel Hilbert space (RKHS) and leveraging an over-determined analysis of the law of total expectation, it elucidates the fundamental nature of fairness conflicts. The paper introduces the “Pokémon Theorem,” proving that any finite set of linear mean-based fairness criteria inevitably entails residual violations and revealing the unavoidable class collapse in fair representation learning. Building on RKHS theory, maximum mean discrepancy (MMD), Kolmogorov m-width, and spectral regularization, the authors derive a signal–error frontier under approximate fairness relaxations, with experiments validating the theoretical bounds on standard fairness benchmarks.

0 citationsRead paper