Selected work

Representative Papers

Spin-Weighted Spherical Harmonics Enable Complete and Scalable $\mathrm{E}(3)$-Equivariant Networks

Jul 01, 2026

Existing $\mathrm{E}(3)$-equivariant networks are constrained by the $O(L^6)$ computational complexity of Clebsch–Gordan tensor products, while efficient alternatives such as Gaunt tensor products sacrifice expressivity due to the absence of antisymmetric pathways. This work introduces spin-weighted spherical harmonics (SWSH) into equivariant learning for the first time and proposes the SpinGTP framework: a novel tensor product operator derived from the algebraic structure of SWSH that retains Gaunt-level computational efficiency while recovering full symmetry expressivity, including parity-odd components. Experiments demonstrate that SpinGTP achieves accuracy comparable to full Clebsch–Gordan tensor products on benchmarks such as Tetris, 3BPA, SPICE-MACE-OFF, and OC20, and further exhibits superior performance on tasks involving chiral materials and non-centrosymmetric structures.

0 citationsRead paper

Forget Without Compromise: Nexus Sampling for Streaming KV-Cache Eviction Under Fixed Budgets

Jun 22, 2026

This work addresses the challenge of KV cache overflow in long-context and agent-style large language model inference, where existing token eviction methods often mistakenly discard low-attention yet critical bridging tokens. To overcome this limitation, the authors propose Nexus Sampling—a training-free KV cache eviction strategy that uniquely integrates iterative attention path analysis with weighted reservoir sampling. By assigning a Nexus score to each token and retaining them probabilistically based on this metric, the method transcends the rigidity of conventional deterministic top-K eviction. Experimental results demonstrate that even at an 80% cache eviction rate, Nexus Sampling incurs only a 1% performance drop relative to dense attention on LongBench, significantly outperforms baseline approaches on retrieval-intensive tasks, and reduces per-sequence memory consumption by up to 10×.

0 citationsRead paper

SATURN: Symbolic Spatial Reasoning for Multi-Perspective Grounding

Jun 21, 2026

This work addresses the unreliability of vision-language models in reasoning about compositional spatial relationships that depend on reference frames. To tackle this challenge, the authors propose SATURN, a neuro-symbolic framework that decouples perception from reasoning by reconstructing an approximate 3D scene, incorporating viewpoint-aware soft spatial predicates, and leveraging a training-free Pythonic symbolic executor to propagate multi-hop uncertainties. By integrating 3D geometric understanding, soft predicate modeling, and symbolic reasoning, SATURN achieves robust performance on the 3D FORCE benchmark and attains 78.57% accuracy on the real-world MindCube benchmark—outperforming the strongest baseline by 14 percentage points.

0 citationsRead paper

MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing with Flow Matching

Jun 01, 2026

Existing single-turn image editing models often fail in multi-turn interactions due to error accumulation, limiting their ability to support iterative user edits. This work proposes MT-EditFlow, a novel framework that integrates flow matching with reinforcement learning (GRPO/NFT) to enhance the global success rate of sequential editing through multi-turn perspective modeling and a multi-reward optimization mechanism. The key innovation lies in broadcasting aggregated advantage signals across the entire editing trajectory, effectively aligning local editing actions with the global objective. Furthermore, the method systematically designs low-bias, low-variance reward functions to stabilize training. Evaluated on FLUX.1-Kontext-dev, MT-EditFlow achieves a 6.85-point improvement in third-turn editing performance over strong baselines such as Qwen-Image-Edit, while significantly reducing exposure bias and maintaining high marginal success rates.

0 citationsRead paper

OpenJarvis: Personal AI, On Personal Devices

May 16, 2026

This work addresses the trade-off between privacy and performance in personal AI systems, where cloud-based large models compromise data sensitivity while purely local deployment suffers from a 25–39 percentage point drop in accuracy. The authors propose OpenJarvis, a decoupled personal AI architecture that decomposes the system into five independently optimizable primitives: intelligence, engine, agent, tools & memory, and learning—enabling, for the first time, both modular decoupling and end-to-end optimizability. By introducing typed specifications and an LLM-guided collaborative search mechanism between local and cloud resources, the framework jointly optimizes prompts, tool descriptions, and memory configurations. Evaluated across eight benchmarks, OpenJarvis matches or exceeds cloud-only models on four tasks and trails the best cloud baseline by only 3.2 percentage points on average, while reducing API costs by approximately 800× and cutting end-to-end latency by 4×.

0 citationsRead paper

Recent publications

Latest Papers

Spin-Weighted Spherical Harmonics Enable Complete and Scalable $\mathrm{E}(3)$-Equivariant Networks

Jul 01, 2026

Existing $\mathrm{E}(3)$-equivariant networks are constrained by the $O(L^6)$ computational complexity of Clebsch–Gordan tensor products, while efficient alternatives such as Gaunt tensor products sacrifice expressivity due to the absence of antisymmetric pathways. This work introduces spin-weighted spherical harmonics (SWSH) into equivariant learning for the first time and proposes the SpinGTP framework: a novel tensor product operator derived from the algebraic structure of SWSH that retains Gaunt-level computational efficiency while recovering full symmetry expressivity, including parity-odd components. Experiments demonstrate that SpinGTP achieves accuracy comparable to full Clebsch–Gordan tensor products on benchmarks such as Tetris, 3BPA, SPICE-MACE-OFF, and OC20, and further exhibits superior performance on tasks involving chiral materials and non-centrosymmetric structures.

0 citationsRead paper

Forget Without Compromise: Nexus Sampling for Streaming KV-Cache Eviction Under Fixed Budgets

Jun 22, 2026

This work addresses the challenge of KV cache overflow in long-context and agent-style large language model inference, where existing token eviction methods often mistakenly discard low-attention yet critical bridging tokens. To overcome this limitation, the authors propose Nexus Sampling—a training-free KV cache eviction strategy that uniquely integrates iterative attention path analysis with weighted reservoir sampling. By assigning a Nexus score to each token and retaining them probabilistically based on this metric, the method transcends the rigidity of conventional deterministic top-K eviction. Experimental results demonstrate that even at an 80% cache eviction rate, Nexus Sampling incurs only a 1% performance drop relative to dense attention on LongBench, significantly outperforms baseline approaches on retrieval-intensive tasks, and reduces per-sequence memory consumption by up to 10×.

0 citationsRead paper

SATURN: Symbolic Spatial Reasoning for Multi-Perspective Grounding

Jun 21, 2026

This work addresses the unreliability of vision-language models in reasoning about compositional spatial relationships that depend on reference frames. To tackle this challenge, the authors propose SATURN, a neuro-symbolic framework that decouples perception from reasoning by reconstructing an approximate 3D scene, incorporating viewpoint-aware soft spatial predicates, and leveraging a training-free Pythonic symbolic executor to propagate multi-hop uncertainties. By integrating 3D geometric understanding, soft predicate modeling, and symbolic reasoning, SATURN achieves robust performance on the 3D FORCE benchmark and attains 78.57% accuracy on the real-world MindCube benchmark—outperforming the strongest baseline by 14 percentage points.

0 citationsRead paper

MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing with Flow Matching

Jun 01, 2026

Existing single-turn image editing models often fail in multi-turn interactions due to error accumulation, limiting their ability to support iterative user edits. This work proposes MT-EditFlow, a novel framework that integrates flow matching with reinforcement learning (GRPO/NFT) to enhance the global success rate of sequential editing through multi-turn perspective modeling and a multi-reward optimization mechanism. The key innovation lies in broadcasting aggregated advantage signals across the entire editing trajectory, effectively aligning local editing actions with the global objective. Furthermore, the method systematically designs low-bias, low-variance reward functions to stabilize training. Evaluated on FLUX.1-Kontext-dev, MT-EditFlow achieves a 6.85-point improvement in third-turn editing performance over strong baselines such as Qwen-Image-Edit, while significantly reducing exposure bias and maintaining high marginal success rates.

0 citationsRead paper

OpenJarvis: Personal AI, On Personal Devices

May 16, 2026

This work addresses the trade-off between privacy and performance in personal AI systems, where cloud-based large models compromise data sensitivity while purely local deployment suffers from a 25–39 percentage point drop in accuracy. The authors propose OpenJarvis, a decoupled personal AI architecture that decomposes the system into five independently optimizable primitives: intelligence, engine, agent, tools & memory, and learning—enabling, for the first time, both modular decoupling and end-to-end optimizability. By introducing typed specifications and an LLM-guided collaborative search mechanism between local and cloud resources, the framework jointly optimizes prompts, tool descriptions, and memory configurations. Evaluated across eight benchmarks, OpenJarvis matches or exceeds cloud-only models on four tasks and trails the best cloud baseline by only 3.2 percentage points on average, while reducing API costs by approximately 800× and cutting end-to-end latency by 4×.

0 citationsRead paper