Selected work

Representative Papers

Two Views, One Voice: Evidence-Grounded Conversational Music Recommendation

Jul 24, 2026

This work addresses the limitation in traditional conversational recommender systems where tightly coupling retrieval and response generation weakens entity-level signals during dialogue intent evolution, thereby undermining explanation credibility. To overcome this, the study introduces the first explicit decoupling of retrieval and generation in conversational music recommendation. The retrieval module integrates lexical and dense representations, employs a fine-tuned Qwen-8B adapter for task-adaptive pooling, and refines candidates via LightGBM calibration. The generation module adopts an evidence-anchored Propose-Allocate-Select (PAS) framework to structurally leverage retrieved evidence for producing interpretable responses. This approach substantially enhances explanation reliability, achieving third place overall and second in explanation quality in the ACM RecSys Challenge 2026 Blind-B track.

0 citationsRead paper

RACORN-1: Adaptive Recall-Preserving Speedup for Low-Selectivity Filtered Vector Search

Jul 01, 2026

This work addresses the instability in connectivity and sharp recall degradation experienced by existing filtered vector search methods, such as ACORN-1, under extremely low selectivity regimes (<5%). To overcome these limitations, the authors propose RACORN-1, which introduces Adaptive Search Fallback (ASF) and Adaptive Exact Fallback (AEF) mechanisms. These strategies leverage filtered-out nodes as temporary bridges to bypass disconnected paths and incorporate step-size sampling to enhance spatial diversity. Compatible with the ACORN-1 architecture, RACORN-1 achieves 9–26× latency reduction while boosting recall from 0.03 to 0.98 across datasets of 1M–40M scale at selectivity levels between 0.3% and 1%. Its enhanced variant, RACORN-1+, attains perfect recall (1.00) and up to 75× speedup under selectivity ≤0.1%.

0 citationsRead paper

Constrained Auto-Bidding via Generative Response Modeling

May 26, 2026

This work addresses the challenge of simultaneously satisfying long-term budget constraints and optimizing advertiser value under non-stationary traffic and auction environments, where existing autobidding methods often fall short. The authors propose a Generative Response Model (GRM) that shifts the learning objective from bidding actions to system responses, modeling historical conditional sequences to forecast future traffic and cost/value curves. GRM integrates a lightweight analytical controller that leverages one-dimensional root-finding to precisely enforce budget constraints. This approach uniquely combines response modeling with analytical control, offering theoretical constraint guarantees for single-multiplier problems and demonstrating robustness under distributional shifts. Experiments on the AuctionNet dataset show that GRM significantly improves constraint stability and overall performance, outperforming current state-of-the-art baselines.

0 citationsRead paper

Plans for Evaluating Structured Generative Search Summaries

May 25, 2026

This study addresses the absence of an effective evaluation framework for structured generative search summaries—comprising overviews, titled sections, and cited source documents—that appear at the top of natural search results. It presents the first systematic effort to construct a comprehensive evaluation framework tailored to these summaries, explicitly defining their core components and multidimensional assessment criteria. By integrating large language model generation techniques with established information retrieval evaluation methodologies, the work proposes a practical and scalable evaluation framework and outlines a clear empirical validation pathway. This contribution establishes a foundational methodological basis for future research on generative search summaries and their impact on user experience and information access.

0 citationsRead paper

WorldKV: Efficient World Memory with World Retrieval and Compression

May 21, 2026

Autoregressive video diffusion models face significant memory and computational bottlenecks during long-horizon generation due to the linear growth of key-value (KV) cache, making it challenging to simultaneously achieve real-time performance and scene consistency. This work proposes WorldKV, a framework that enables efficient long-term memory without requiring fine-tuning. WorldKV retrieves historical KV blocks using camera or action semantics and compresses redundant tokens by leveraging key-key similarity across critical frames. The approach transcends the conventional trade-off between sliding-window and full-cache strategies, achieving memory fidelity on par with or superior to full KV caching on benchmarks such as Matrix-Game-2.0 and LingBot-World-Fast, while delivering approximately 2× higher inference throughput.

0 citationsRead paper

Recent publications

Latest Papers

Two Views, One Voice: Evidence-Grounded Conversational Music Recommendation

Jul 24, 2026

This work addresses the limitation in traditional conversational recommender systems where tightly coupling retrieval and response generation weakens entity-level signals during dialogue intent evolution, thereby undermining explanation credibility. To overcome this, the study introduces the first explicit decoupling of retrieval and generation in conversational music recommendation. The retrieval module integrates lexical and dense representations, employs a fine-tuned Qwen-8B adapter for task-adaptive pooling, and refines candidates via LightGBM calibration. The generation module adopts an evidence-anchored Propose-Allocate-Select (PAS) framework to structurally leverage retrieved evidence for producing interpretable responses. This approach substantially enhances explanation reliability, achieving third place overall and second in explanation quality in the ACM RecSys Challenge 2026 Blind-B track.

0 citationsRead paper

RACORN-1: Adaptive Recall-Preserving Speedup for Low-Selectivity Filtered Vector Search

Jul 01, 2026

This work addresses the instability in connectivity and sharp recall degradation experienced by existing filtered vector search methods, such as ACORN-1, under extremely low selectivity regimes (<5%). To overcome these limitations, the authors propose RACORN-1, which introduces Adaptive Search Fallback (ASF) and Adaptive Exact Fallback (AEF) mechanisms. These strategies leverage filtered-out nodes as temporary bridges to bypass disconnected paths and incorporate step-size sampling to enhance spatial diversity. Compatible with the ACORN-1 architecture, RACORN-1 achieves 9–26× latency reduction while boosting recall from 0.03 to 0.98 across datasets of 1M–40M scale at selectivity levels between 0.3% and 1%. Its enhanced variant, RACORN-1+, attains perfect recall (1.00) and up to 75× speedup under selectivity ≤0.1%.

0 citationsRead paper

Constrained Auto-Bidding via Generative Response Modeling

May 26, 2026

This work addresses the challenge of simultaneously satisfying long-term budget constraints and optimizing advertiser value under non-stationary traffic and auction environments, where existing autobidding methods often fall short. The authors propose a Generative Response Model (GRM) that shifts the learning objective from bidding actions to system responses, modeling historical conditional sequences to forecast future traffic and cost/value curves. GRM integrates a lightweight analytical controller that leverages one-dimensional root-finding to precisely enforce budget constraints. This approach uniquely combines response modeling with analytical control, offering theoretical constraint guarantees for single-multiplier problems and demonstrating robustness under distributional shifts. Experiments on the AuctionNet dataset show that GRM significantly improves constraint stability and overall performance, outperforming current state-of-the-art baselines.

0 citationsRead paper

Plans for Evaluating Structured Generative Search Summaries

May 25, 2026

This study addresses the absence of an effective evaluation framework for structured generative search summaries—comprising overviews, titled sections, and cited source documents—that appear at the top of natural search results. It presents the first systematic effort to construct a comprehensive evaluation framework tailored to these summaries, explicitly defining their core components and multidimensional assessment criteria. By integrating large language model generation techniques with established information retrieval evaluation methodologies, the work proposes a practical and scalable evaluation framework and outlines a clear empirical validation pathway. This contribution establishes a foundational methodological basis for future research on generative search summaries and their impact on user experience and information access.

0 citationsRead paper

WorldKV: Efficient World Memory with World Retrieval and Compression

May 21, 2026

Autoregressive video diffusion models face significant memory and computational bottlenecks during long-horizon generation due to the linear growth of key-value (KV) cache, making it challenging to simultaneously achieve real-time performance and scene consistency. This work proposes WorldKV, a framework that enables efficient long-term memory without requiring fine-tuning. WorldKV retrieves historical KV blocks using camera or action semantics and compresses redundant tokens by leveraging key-key similarity across critical frames. The approach transcends the conventional trade-off between sliding-window and full-cache strategies, achieving memory fidelity on par with or superior to full KV caching on benchmarks such as Matrix-Game-2.0 and LingBot-World-Fast, while delivering approximately 2× higher inference throughput.

0 citationsRead paper