Selected work

Representative Papers

PairCoder++: Pair Programming as a Universal Paradigm for Verified Code-Driven Multimodal and Structured-Artifact Generation

Jul 02, 2026

Existing code-driven generation of structured artifacts—such as charts and 3D scenes—is prone to errors due to the absence of validation feedback from toolchains like compilers and renderers. This work proposes a dual-agent pair programming framework: a Driver agent generates code, while a Navigator agent critiques it based on execution outcomes, diagnostic messages, and rendered visual comparisons. When persistent errors occur, the agents dynamically swap roles, establishing a closed-loop iterative refinement process. This approach represents the first systematic integration of toolchain feedback into multimodal structured generation, substantially improving output executability. Evaluated across 17 benchmarks, the method consistently outperforms baselines—for instance, increasing Blender scene executability from 0.20 to 0.78 and boosting TikZ compilation rates by 10–30 percentage points—at an inference cost approximately seven times that of a single model.

0 citationsRead paper

Agent-Assisted Side-Channel Attacks on Non-Prefix KV Cache in RAG

Jun 19, 2026

Existing KV cache side-channel attacks rely on strict prefix alignment, rendering them ineffective in realistic retrieval-augmented generation (RAG) scenarios involving private prefixes. This work introduces SpliceLeak, the first side-channel attack targeting non-prefix-aligned KV cache fusion. By analyzing the “staircase” timing patterns induced by memory block scheduling, SpliceLeak enables end-to-end extraction of both the length and semantic content of private prompts through a novel two-stage privacy theft mechanism. To counter this threat, the paper also proposes SpliceDefense, an efficient defense framework combining Quantized Chunk Padding (QCP) and Constant-Time Boundary Fusion (CTBF). Evaluated on vLLM+LMCache, SpliceLeak achieves 100% content extraction success with only 63 queries per token, while SpliceDefense incurs negligible throughput overhead and effectively eliminates the side-channel signal.

0 citationsRead paper

CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels

May 22, 2026

Real-world data often exhibit both long-tailed distributions and non-uniform label noise, with tail classes particularly susceptible to noisy annotations. To address this challenge, this work proposes the CARE framework, which introduces, for the first time, a class-adaptive expert consensus mechanism. CARE integrates three types of supervision signals—noisy labels, textual embeddings, and visual features—derived from vision-language models, imposing stronger consistency constraints on tail classes while moderately relaxing them for head classes to achieve precise and balanced label correction. Extensive experiments demonstrate that CARE significantly outperforms existing methods across multiple synthetic and real-world benchmarks, achieving performance gains of up to 3.0%.

0 citationsRead paper

One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems

May 21, 2026

This work addresses significant limitations in existing short-drama generation methods, particularly in narrative pacing, spatial consistency, and production quality. The authors propose a hierarchical multi-agent framework that transforms a user’s one-sentence prompt into a high-quality, personalized short drama in an end-to-end manner. Key innovations include a multi-agent debate–based story generation mechanism, spatial consistency modeling via 3D-anchored initial frames, and an iterative refinement pipeline integrating multi-stage review loops, scene-level background music matching, and transition planning. Experimental results demonstrate that the proposed approach substantially outperforms current methods in narrative coherence, cross-segment consistency, and overall viewer experience. To facilitate systematic evaluation, the authors also introduce Short-Drama-Bench, a dedicated benchmark for short-drama generation.

0 citationsRead paper

Resolving Long-Tail Ambiguity in Unsupervised 3D Point Cloud Segmentation with Language Priors

May 20, 2026

In unsupervised 3D point cloud segmentation, long-tailed categories are often absorbed by dominant classes during purely visual clustering, leading to severely imbalanced predictions. To address this issue, this work proposes LangTail, a novel framework that, for the first time, incorporates balanced semantic priors from language models into unsupervised 3D segmentation. By aligning language and visual features at multiple levels, LangTail guides hierarchical clustering and explicitly associates entity-level semantics with rare categories, thereby enhancing the representational capacity for minority classes. The method achieves significant improvements over existing approaches, yielding gains of 13.5, 12.9, and 8.9 mIoU on ScanNet-v2, S3DIS, and nuScenes benchmarks, respectively.

0 citationsRead paper

Recent publications

Latest Papers

PairCoder++: Pair Programming as a Universal Paradigm for Verified Code-Driven Multimodal and Structured-Artifact Generation

Jul 02, 2026

Existing code-driven generation of structured artifacts—such as charts and 3D scenes—is prone to errors due to the absence of validation feedback from toolchains like compilers and renderers. This work proposes a dual-agent pair programming framework: a Driver agent generates code, while a Navigator agent critiques it based on execution outcomes, diagnostic messages, and rendered visual comparisons. When persistent errors occur, the agents dynamically swap roles, establishing a closed-loop iterative refinement process. This approach represents the first systematic integration of toolchain feedback into multimodal structured generation, substantially improving output executability. Evaluated across 17 benchmarks, the method consistently outperforms baselines—for instance, increasing Blender scene executability from 0.20 to 0.78 and boosting TikZ compilation rates by 10–30 percentage points—at an inference cost approximately seven times that of a single model.

0 citationsRead paper

Agent-Assisted Side-Channel Attacks on Non-Prefix KV Cache in RAG

Jun 19, 2026

Existing KV cache side-channel attacks rely on strict prefix alignment, rendering them ineffective in realistic retrieval-augmented generation (RAG) scenarios involving private prefixes. This work introduces SpliceLeak, the first side-channel attack targeting non-prefix-aligned KV cache fusion. By analyzing the “staircase” timing patterns induced by memory block scheduling, SpliceLeak enables end-to-end extraction of both the length and semantic content of private prompts through a novel two-stage privacy theft mechanism. To counter this threat, the paper also proposes SpliceDefense, an efficient defense framework combining Quantized Chunk Padding (QCP) and Constant-Time Boundary Fusion (CTBF). Evaluated on vLLM+LMCache, SpliceLeak achieves 100% content extraction success with only 63 queries per token, while SpliceDefense incurs negligible throughput overhead and effectively eliminates the side-channel signal.

0 citationsRead paper

CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels

May 22, 2026

Real-world data often exhibit both long-tailed distributions and non-uniform label noise, with tail classes particularly susceptible to noisy annotations. To address this challenge, this work proposes the CARE framework, which introduces, for the first time, a class-adaptive expert consensus mechanism. CARE integrates three types of supervision signals—noisy labels, textual embeddings, and visual features—derived from vision-language models, imposing stronger consistency constraints on tail classes while moderately relaxing them for head classes to achieve precise and balanced label correction. Extensive experiments demonstrate that CARE significantly outperforms existing methods across multiple synthetic and real-world benchmarks, achieving performance gains of up to 3.0%.

0 citationsRead paper

One Sentence, One Drama: Personalized Short-Form Drama Generation via Multi-Agent Systems

May 21, 2026

This work addresses significant limitations in existing short-drama generation methods, particularly in narrative pacing, spatial consistency, and production quality. The authors propose a hierarchical multi-agent framework that transforms a user’s one-sentence prompt into a high-quality, personalized short drama in an end-to-end manner. Key innovations include a multi-agent debate–based story generation mechanism, spatial consistency modeling via 3D-anchored initial frames, and an iterative refinement pipeline integrating multi-stage review loops, scene-level background music matching, and transition planning. Experimental results demonstrate that the proposed approach substantially outperforms current methods in narrative coherence, cross-segment consistency, and overall viewer experience. To facilitate systematic evaluation, the authors also introduce Short-Drama-Bench, a dedicated benchmark for short-drama generation.

0 citationsRead paper

Resolving Long-Tail Ambiguity in Unsupervised 3D Point Cloud Segmentation with Language Priors

May 20, 2026

In unsupervised 3D point cloud segmentation, long-tailed categories are often absorbed by dominant classes during purely visual clustering, leading to severely imbalanced predictions. To address this issue, this work proposes LangTail, a novel framework that, for the first time, incorporates balanced semantic priors from language models into unsupervised 3D segmentation. By aligning language and visual features at multiple levels, LangTail guides hierarchical clustering and explicitly associates entity-level semantics with rare categories, thereby enhancing the representational capacity for minority classes. The method achieves significant improvements over existing approaches, yielding gains of 13.5, 12.9, and 8.9 mIoU on ScanNet-v2, S3DIS, and nuScenes benchmarks, respectively.

0 citationsRead paper