Selected work

Representative Papers

Arbor: Explicit Geometric Conditioning for Controllable 3D Asset Generation

Jun 22, 2026

Existing text- or image-conditioned 3D generation models struggle to explicitly express user intent regarding spatial regions an object should occupy or avoid. To address this limitation, this work proposes Arbor—a trainable plug-in module that, for the first time, introduces locally typed geometric constraint grids (specifying presence, avoidance, and contact zones) as non-target evidence into latent 3D diffusion models. Arbor learns to inject these geometric constraints positionally within a frozen denoiser via a constraint-to-token transformation and a spatial routing attention mechanism, enabling fine-grained control over the generated object’s spatial layout. Experiments demonstrate that Arbor significantly improves adherence to spatial constraints in both automatic and human evaluations, without requiring specialized compliance losses, while preserving generation quality and diversity under fixed constraints.

0 citationsRead paper

LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models

Jun 03, 2026

This work addresses the limitations of conventional tabular foundation models, whose scalar tokenization restricts feature expressiveness, leading to low-rank collapse and attention bottlenecks that diminish shallow-layer sensitivity to feature values and induce redundant representations. To overcome these issues, the authors propose a unified “token-routing” framework: RaBEL expands scalars into local radial basis function (RBF) features augmented with exponential gating, thereby improving the condition number and effective rank; additionally, an S→N→F reordered bidirectional block aggregates cross-sample context prior to feature mixing and integrates attention-based pooling for efficient modeling. The resulting model, LimiX-2M, with only 2 million parameters, surpasses larger counterparts such as TabPFN-v2 and TabICL on mainstream tabular benchmarks, achieving a significantly improved trade-off between accuracy and efficiency.

0 citationsRead paper

A Cookbook of 3D Vision: Data, Learning Paradigms, and Application

Jun 02, 2026

The field of 3D vision suffers from fragmented data representations, learning paradigms, and benchmarking protocols, leading to a lack of unified understanding regarding efficiency, fidelity, and scalability. This work proposes the first cohesive conceptual framework that integrates geometric representations—such as point clouds, meshes, voxels, and 3D Gaussians—with diverse learning paradigms—including 2D-supervised learning, implicit neural representations, and 4D modeling—and connects them to real-world application scenarios. By constructing a structured knowledge graph of 3D vision, the study systematically relates dataset design, supervision mechanisms, and task requirements, clarifying the trade-offs between efficiency and fidelity and charting pathways for multimodal geometric grounding. This framework offers systematic guidance for reconstruction, generation, and dynamic scene modeling, advancing the field toward a unified and efficient paradigm.

0 citationsRead paper

Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning

May 28, 2026

This work addresses the challenges in image layer decomposition arising from the absence of paired supervision and the reliance on unreliable visual-language model (VLM) scores for policy optimization. To overcome these limitations, the authors propose a reinforcement learning approach built upon the pretrained Qwen-Image-Layered model, employing Flow-GRPO with LoRA-based efficient fine-tuning. A two-stage VLM evaluation mechanism is introduced: first, structured scoring according to five edit-oriented criteria, followed by grid-based juxtaposition for recalibration, which substantially enhances score discriminability and training stability. Evaluated on the Crello dataset, the method produces cleaner layer separation with fewer artifacts and achieves significantly lower single-layer reconstruction error compared to existing baselines.

0 citationsRead paper

OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under optimal Squared error quantization

May 20, 2026

This work addresses the substantial memory bandwidth and storage demands of key-value (KV) caching in long-context autoregressive inference. The authors propose a data-independent, online, and deterministic quantization method that jointly maps coordinate triplets into an octahedral parameter space via structured random rotations, subsequently transforming them into a square domain. By integrating Lloyd-Max quantization with mean-squared-error-optimized non-uniform bit allocation for both direction and magnitude, the approach achieves dimension-adaptive optimal compression. Notably, it introduces triplet-wise joint quantization and octahedral parameterization for the first time. Evaluated across text, video, and audio tasks, the method consistently outperforms existing rotation-based codecs at all bitrates—particularly under extreme compression—without incurring additional inference bandwidth or latency overhead.

0 citationsRead paper

Recent publications

Latest Papers

Arbor: Explicit Geometric Conditioning for Controllable 3D Asset Generation

Jun 22, 2026

Existing text- or image-conditioned 3D generation models struggle to explicitly express user intent regarding spatial regions an object should occupy or avoid. To address this limitation, this work proposes Arbor—a trainable plug-in module that, for the first time, introduces locally typed geometric constraint grids (specifying presence, avoidance, and contact zones) as non-target evidence into latent 3D diffusion models. Arbor learns to inject these geometric constraints positionally within a frozen denoiser via a constraint-to-token transformation and a spatial routing attention mechanism, enabling fine-grained control over the generated object’s spatial layout. Experiments demonstrate that Arbor significantly improves adherence to spatial constraints in both automatic and human evaluations, without requiring specialized compliance losses, while preserving generation quality and diversity under fixed constraints.

0 citationsRead paper

LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models

Jun 03, 2026

This work addresses the limitations of conventional tabular foundation models, whose scalar tokenization restricts feature expressiveness, leading to low-rank collapse and attention bottlenecks that diminish shallow-layer sensitivity to feature values and induce redundant representations. To overcome these issues, the authors propose a unified “token-routing” framework: RaBEL expands scalars into local radial basis function (RBF) features augmented with exponential gating, thereby improving the condition number and effective rank; additionally, an S→N→F reordered bidirectional block aggregates cross-sample context prior to feature mixing and integrates attention-based pooling for efficient modeling. The resulting model, LimiX-2M, with only 2 million parameters, surpasses larger counterparts such as TabPFN-v2 and TabICL on mainstream tabular benchmarks, achieving a significantly improved trade-off between accuracy and efficiency.

0 citationsRead paper

A Cookbook of 3D Vision: Data, Learning Paradigms, and Application

Jun 02, 2026

The field of 3D vision suffers from fragmented data representations, learning paradigms, and benchmarking protocols, leading to a lack of unified understanding regarding efficiency, fidelity, and scalability. This work proposes the first cohesive conceptual framework that integrates geometric representations—such as point clouds, meshes, voxels, and 3D Gaussians—with diverse learning paradigms—including 2D-supervised learning, implicit neural representations, and 4D modeling—and connects them to real-world application scenarios. By constructing a structured knowledge graph of 3D vision, the study systematically relates dataset design, supervision mechanisms, and task requirements, clarifying the trade-offs between efficiency and fidelity and charting pathways for multimodal geometric grounding. This framework offers systematic guidance for reconstruction, generation, and dynamic scene modeling, advancing the field toward a unified and efficient paradigm.

0 citationsRead paper

Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning

May 28, 2026

This work addresses the challenges in image layer decomposition arising from the absence of paired supervision and the reliance on unreliable visual-language model (VLM) scores for policy optimization. To overcome these limitations, the authors propose a reinforcement learning approach built upon the pretrained Qwen-Image-Layered model, employing Flow-GRPO with LoRA-based efficient fine-tuning. A two-stage VLM evaluation mechanism is introduced: first, structured scoring according to five edit-oriented criteria, followed by grid-based juxtaposition for recalibration, which substantially enhances score discriminability and training stability. Evaluated on the Crello dataset, the method produces cleaner layer separation with fewer artifacts and achieves significantly lower single-layer reconstruction error compared to existing baselines.

0 citationsRead paper

OCTOPUS: Optimized KV Cache for Transformers via Octahedral Parametrization Under optimal Squared error quantization

May 20, 2026

This work addresses the substantial memory bandwidth and storage demands of key-value (KV) caching in long-context autoregressive inference. The authors propose a data-independent, online, and deterministic quantization method that jointly maps coordinate triplets into an octahedral parameter space via structured random rotations, subsequently transforming them into a square domain. By integrating Lloyd-Max quantization with mean-squared-error-optimized non-uniform bit allocation for both direction and magnitude, the approach achieves dimension-adaptive optimal compression. Notably, it introduces triplet-wise joint quantization and octahedral parameterization for the first time. Evaluated across text, video, and audio tasks, the method consistently outperforms existing rotation-based codecs at all bitrates—particularly under extreme compression—without incurring additional inference bandwidth or latency overhead.

0 citationsRead paper