Institution profile

Telepix

Industry research
Research library15linked papers
Opportunities0open roles
Selected work

Representative Papers

Platonic Task Arithmetic

Sep 30, 2026

This study addresses the limitation that task arithmetic among heterogeneous models cannot be transferred across architectures. To overcome this, it proposes a “Platonic” universal task descriptor that abstracts model-specific updates into architecture-agnostic shared representations. By integrating matrix operations, least-squares optimization, and Low-Rank Adaptation (LoRA), the method enables label-free cross-model task knowledge editing and linear composition. Extensive evaluations across six model families and eight tasks demonstrate that the proposed cross-architecture transfer preserves 74%–80% of the performance gains achieved by target models independently. These results indicate that the approach effectively overcomes the architectural barriers inherent in conventional task arithmetic, facilitating robust and flexible knowledge sharing among structurally diverse foundation models without requiring labeled data.

0 citationsRead paper

Beyond Simulation: Retain-and-Repair Neural Operators for Real-World Adaptation

Sep 30, 2026

This study addresses the structural degradation of pretrained neural operators during simulation-to-real (Sim2Real) transfer and proposes the R²NO framework. The method freezes the fine-tuned source predictor and introduces a shared repair module, innovatively formulating adaptation depth as the selective activation of Fourier-domain units learned independently from real data. It further integrates orthogonal Fourier projection, spectral ensembling, and ridge regression to fuse multiple candidate refinements. Experiments on RealPDEBench across six backbone architectures demonstrate that R²NO consistently outperforms full fine-tuning and iterative refinement baselines, achieving robust Sim2Real transfer for neural operators.

0 citationsRead paper

ELMZip: Onboard Satellite Image Compression via Extreme Learning Machines for Efficient Downlink

Aug 07, 2026

This study addresses the downlink bottleneck in small satellite multispectral imaging, where large data volumes and limited communication windows challenge conventional compression methods that struggle with the nonlinear statistical characteristics of multi-band, multi-resolution imagery. To overcome this, the paper proposes ELMZip, a novel on-board image compression framework that introduces extreme learning machines (ELMs) into spaceborne processing. By integrating domain decomposition and random feature mapping, ELMZip formulates image representation as a convex least-squares problem, enabling efficient neural implicit modeling without backpropagation. An asymmetric protocol transmits only compact output weights, drastically reducing downlink payload. The approach achieves high-fidelity reconstruction while substantially minimizing data return volume, thereby enabling real-time remote sensing analytics on resource-constrained platforms.

0 citationsRead paper

Summarize First, Download Later: Onboard VLMs for Bandwidth-Efficient Earth Observation

Aug 07, 2026

This study addresses the challenge of limited downlink bandwidth in Earth observation satellites, which often causes delays or loss of high-resolution remote sensing data, thereby compromising time-sensitive applications. To overcome this, the authors propose a semantic-driven downlink paradigm—“summarize first, download later”—wherein a lightweight vision-language model is deployed onboard to generate natural language summaries of acquired imagery. Ground users then interactively verify critical information via visual question answering (VQA) and selectively request full-resolution images only when necessary. This approach pioneers the integration of vision-language models and interactive VQA into space-to-ground communications, shifting from passive bulk transmission to semantic-aware, active dialogue. Experiments on an NVIDIA Jetson platform demonstrate that the proposed method substantially reduces bandwidth consumption while accelerating information retrieval for time-critical tasks.

0 citationsRead paper
Recent publications

Latest Papers

Platonic Task Arithmetic

Sep 30, 2026

This study addresses the limitation that task arithmetic among heterogeneous models cannot be transferred across architectures. To overcome this, it proposes a “Platonic” universal task descriptor that abstracts model-specific updates into architecture-agnostic shared representations. By integrating matrix operations, least-squares optimization, and Low-Rank Adaptation (LoRA), the method enables label-free cross-model task knowledge editing and linear composition. Extensive evaluations across six model families and eight tasks demonstrate that the proposed cross-architecture transfer preserves 74%–80% of the performance gains achieved by target models independently. These results indicate that the approach effectively overcomes the architectural barriers inherent in conventional task arithmetic, facilitating robust and flexible knowledge sharing among structurally diverse foundation models without requiring labeled data.

0 citationsRead paper

Beyond Simulation: Retain-and-Repair Neural Operators for Real-World Adaptation

Sep 30, 2026

This study addresses the structural degradation of pretrained neural operators during simulation-to-real (Sim2Real) transfer and proposes the R²NO framework. The method freezes the fine-tuned source predictor and introduces a shared repair module, innovatively formulating adaptation depth as the selective activation of Fourier-domain units learned independently from real data. It further integrates orthogonal Fourier projection, spectral ensembling, and ridge regression to fuse multiple candidate refinements. Experiments on RealPDEBench across six backbone architectures demonstrate that R²NO consistently outperforms full fine-tuning and iterative refinement baselines, achieving robust Sim2Real transfer for neural operators.

0 citationsRead paper

ELMZip: Onboard Satellite Image Compression via Extreme Learning Machines for Efficient Downlink

Aug 07, 2026

This study addresses the downlink bottleneck in small satellite multispectral imaging, where large data volumes and limited communication windows challenge conventional compression methods that struggle with the nonlinear statistical characteristics of multi-band, multi-resolution imagery. To overcome this, the paper proposes ELMZip, a novel on-board image compression framework that introduces extreme learning machines (ELMs) into spaceborne processing. By integrating domain decomposition and random feature mapping, ELMZip formulates image representation as a convex least-squares problem, enabling efficient neural implicit modeling without backpropagation. An asymmetric protocol transmits only compact output weights, drastically reducing downlink payload. The approach achieves high-fidelity reconstruction while substantially minimizing data return volume, thereby enabling real-time remote sensing analytics on resource-constrained platforms.

0 citationsRead paper

Summarize First, Download Later: Onboard VLMs for Bandwidth-Efficient Earth Observation

Aug 07, 2026

This study addresses the challenge of limited downlink bandwidth in Earth observation satellites, which often causes delays or loss of high-resolution remote sensing data, thereby compromising time-sensitive applications. To overcome this, the authors propose a semantic-driven downlink paradigm—“summarize first, download later”—wherein a lightweight vision-language model is deployed onboard to generate natural language summaries of acquired imagery. Ground users then interactively verify critical information via visual question answering (VQA) and selectively request full-resolution images only when necessary. This approach pioneers the integration of vision-language models and interactive VQA into space-to-ground communications, shifting from passive bulk transmission to semantic-aware, active dialogue. Experiments on an NVIDIA Jetson platform demonstrate that the proposed method substantially reduces bandwidth consumption while accelerating information retrieval for time-critical tasks.

0 citationsRead paper