Selected work

Representative Papers

Masking or Mitigating? Deconstructing the Impact of Query Rewriting on Retriever Biases in RAG

Apr 07, 2026

This work investigates systematic biases—such as conciseness, position, verbatim matching, and redundancy—in dense retrievers within retrieval-augmented generation (RAG) systems, where the impact of existing query rewriting approaches remains unclear. The study introduces the first taxonomy distinguishing between query–document interaction bias and document encoding bias, and systematically evaluates the debiasing efficacy of five query augmentation methods across six retrievers. By integrating LLM-driven query rewriting, pseudo-document generation, adversarial testing, and mechanistic analysis, it reveals that different methods mitigate bias either by increasing score variance or decorrelating biased signals. Experiments show that simple LLM-based rewriting reduces bias by 54% on average but fails under adversarial conditions. No single method universally addresses all biases; performance is highly dependent on the retriever architecture, offering practical guidance for deployment choices.

0 citationsRead paper

FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching

Feb 01, 2026

Flow Matching models suffer from slow inference due to the need for numerous denoising steps, hindering their applicability in real-time scenarios; existing acceleration methods often compromise generation quality, require retraining, or exhibit poor generalization. This work proposes FlowCast, a training-free, plug-and-play speculative acceleration framework that exploits the constant velocity field property inherent in Flow Matching. By extrapolating the current trajectory to predict future states and employing a mean squared error threshold to govern prediction acceptance, FlowCast enables large step sizes in stable regions while preserving accuracy in complex areas. Theoretical analysis provides bounds on trajectory deviation. Experiments on image and video generation and editing tasks demonstrate that FlowCast achieves over 2.5× speedup without any loss in output quality, significantly outperforming existing baselines.

0 citationsRead paper

ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models

Jan 07, 2026arXiv.org

This work addresses the tendency of large language models to rely on internal memorized knowledge over conflicting but relevant external retrieval context, often resulting in unfaithful outputs. To mitigate this issue, the authors propose a lightweight activation steering method that intervenes in the activation signals of key transformer layers. Without requiring fine-tuning or incurring significant inference overhead, this approach substantially enhances contextual faithfulness while preserving generation fluency. The method is compatible with existing prompting strategies and scales effectively to large models. Evaluated on the ConFiQA benchmark, it significantly outperforms strong baselines such as ContextDPO and COIECD, demonstrating its effectiveness, robustness, and computational efficiency.

0 citationsRead paper

Mercury: Ultra-Fast Language Models Based on Diffusion

Jun 17, 2025

To address the slow autoregressive decoding of large language models (LLMs), which hinders real-time programming applications, this paper introduces Mercury—the first production-grade diffusion-based LLM accelerator. Its core innovation lies in pioneering the integration of diffusion modeling into large-scale language modeling, enabling token-level parallel generation within a Transformer architecture and overcoming the fundamental sequential decoding bottleneck. We further present Mercury Coder, a code-generation-optimized variant with Mini and Small configurations; on H100 GPUs, it achieves throughputs of 1,109 and 737 tokens/sec, respectively—up to 10× faster than state-of-the-art acceleration baselines. In the Copilot Arena benchmark, Mercury Coder ranks first in speed and second in quality. The model, along with its API and interactive web platform, has been open-sourced, empirically validating the feasibility and superiority of the diffusion paradigm for efficient LLM inference.

0 citationsRead paper

Recent publications

Latest Papers

Masking or Mitigating? Deconstructing the Impact of Query Rewriting on Retriever Biases in RAG

Apr 07, 2026

This work investigates systematic biases—such as conciseness, position, verbatim matching, and redundancy—in dense retrievers within retrieval-augmented generation (RAG) systems, where the impact of existing query rewriting approaches remains unclear. The study introduces the first taxonomy distinguishing between query–document interaction bias and document encoding bias, and systematically evaluates the debiasing efficacy of five query augmentation methods across six retrievers. By integrating LLM-driven query rewriting, pseudo-document generation, adversarial testing, and mechanistic analysis, it reveals that different methods mitigate bias either by increasing score variance or decorrelating biased signals. Experiments show that simple LLM-based rewriting reduces bias by 54% on average but fails under adversarial conditions. No single method universally addresses all biases; performance is highly dependent on the retriever architecture, offering practical guidance for deployment choices.

0 citationsRead paper

FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching

Feb 01, 2026

Flow Matching models suffer from slow inference due to the need for numerous denoising steps, hindering their applicability in real-time scenarios; existing acceleration methods often compromise generation quality, require retraining, or exhibit poor generalization. This work proposes FlowCast, a training-free, plug-and-play speculative acceleration framework that exploits the constant velocity field property inherent in Flow Matching. By extrapolating the current trajectory to predict future states and employing a mean squared error threshold to govern prediction acceptance, FlowCast enables large step sizes in stable regions while preserving accuracy in complex areas. Theoretical analysis provides bounds on trajectory deviation. Experiments on image and video generation and editing tasks demonstrate that FlowCast achieves over 2.5× speedup without any loss in output quality, significantly outperforming existing baselines.

0 citationsRead paper

ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models

Jan 07, 2026arXiv.org

This work addresses the tendency of large language models to rely on internal memorized knowledge over conflicting but relevant external retrieval context, often resulting in unfaithful outputs. To mitigate this issue, the authors propose a lightweight activation steering method that intervenes in the activation signals of key transformer layers. Without requiring fine-tuning or incurring significant inference overhead, this approach substantially enhances contextual faithfulness while preserving generation fluency. The method is compatible with existing prompting strategies and scales effectively to large models. Evaluated on the ConFiQA benchmark, it significantly outperforms strong baselines such as ContextDPO and COIECD, demonstrating its effectiveness, robustness, and computational efficiency.

0 citationsRead paper

Mercury: Ultra-Fast Language Models Based on Diffusion

Jun 17, 2025

To address the slow autoregressive decoding of large language models (LLMs), which hinders real-time programming applications, this paper introduces Mercury—the first production-grade diffusion-based LLM accelerator. Its core innovation lies in pioneering the integration of diffusion modeling into large-scale language modeling, enabling token-level parallel generation within a Transformer architecture and overcoming the fundamental sequential decoding bottleneck. We further present Mercury Coder, a code-generation-optimized variant with Mini and Small configurations; on H100 GPUs, it achieves throughputs of 1,109 and 737 tokens/sec, respectively—up to 10× faster than state-of-the-art acceleration baselines. In the Copilot Arena benchmark, Mercury Coder ranks first in speed and second in quality. The model, along with its API and interactive web platform, has been open-sourced, empirically validating the feasibility and superiority of the diffusion paradigm for efficient LLM inference.

0 citationsRead paper