Selected work

Representative Papers

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

Aug 12, 2026

This study addresses two critical questions in glacial lake outburst floods, landslides, and glacier-related flooding in the Himalayas: where hazards are prone to occur and when they are triggered. For the first time, it decouples deformation and meteorological signals into distinct spatial susceptibility and temporal triggering prediction tasks, leveraging only freely available satellite data (e.g., InSAR and weather observations) and topographic features to build a self-contained predictive model that does not rely on neighboring-region information. Rigorous spatiotemporal cross-validation mitigates geographic overfitting, revealing that simpler models—such as gradient-boosted trees—outperform complex deep learning approaches. Meteorological indicators achieve AUC scores of 0.73–0.83 for trigger timing prediction, whereas terrain-based features show limited discriminative power for susceptibility (AUC 0.54–0.76). These findings underpin a prioritized hazard risk inventory for Nepal.

0 citationsRead paper

iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

Aug 06, 2026

Existing 3D scene generation methods struggle to reliably satisfy task-critical functional constraints such as navigability and reachability, limiting the practical utility of synthetic data. This work proposes an iterative agent-based reinforcement learning framework that first enhances physical plausibility and layout quality through pretraining with generic rewards, then leverages a large language model (LLM) to generate executable, task-specific reward programs. These LLM-generated rewards are integrated into a feedback-driven reinforcement learning loop for iterative refinement. By uniquely combining LLM-synthesized reward functions with iterative reinforcement learning, the approach significantly improves adherence to functional constraints while preserving scene diversity, thereby enhancing downstream task performance.

0 citationsRead paper

Beyond Edge Maps: Wavelet-Domain Conditioning for Multi-Adapter Map-to-Satellite Diffusion

Jul 29, 2026

This study addresses the challenge of outdated satellite basemaps in low-resource regions, where limited collaboration with commercial mapping providers hinders timely updates. To tackle this issue, the authors propose a ControlNet-based diffusion model that generates high-quality satellite imagery using only readily available OpenStreetMap raster data and its stationary wavelet transform (SWT) subbands as conditioning inputs. A multi-adapter fusion strategy jointly guides both spatial structure and frequency-domain details without requiring model retraining for multiple conditions. Notably, this work is the first to incorporate SWT subbands into map-to-image generation. Evaluated on a newly curated Nepal dataset and the Pix2Pix benchmark, the proposed method outperforms existing approaches on six out of eight metrics, with SWT-only conditioning achieving the lowest Fréchet Inception Distance (FID) on both datasets.

0 citationsRead paper

An Evolutionary Approach for Designing Stable and Highly Expressible Low-Immunogenicity Therapeutic mRNA Sequences

May 27, 2026

This study addresses the challenge of designing therapeutic mRNA sequences that simultaneously maximize translation efficiency, structural stability, and low immunogenicity—objectives often in tension with one another. To this end, the authors propose a two-stage computational framework: first, a pre-trained CodonTransformer, a BERT-like large language model, generates candidate sequences encoding the target antigen; second, a multi-objective genetic algorithm incorporating human codon usage bias, synonymous mutations, and codon-aware crossover refines these candidates. This approach represents the first integration of large language models with evolutionary optimization for mRNA design and substantially outperforms baseline methods such as LinearDesign and BiLSTM-CRF. It achieves state-of-the-art performance across key metrics, including CAI (0.73–0.74), tAI (0.63–0.64), 5′-end unpaired proportion (0.87), global minimum free energy (−346 to −356 kcal/mol), and immunogenicity penalty (27.3).

0 citationsRead paper

IsoNet: Spatially-aware audio-visual target speech extraction in complex acoustic environments

May 14, 2026

This work addresses the challenge of target speech extraction on compact devices equipped with small-aperture microphone arrays, where conventional beamforming and monaural neural models struggle to perform effectively. To overcome this limitation, we propose IsoNet—a multimodal U-Net-based mask estimation network tailored for four-microphone arrays. IsoNet uniquely integrates complex-valued multi-channel STFT features, GCC-PHAT spatial cues, face-conditioned visual embeddings, and DOA-assisted supervision, enabling user-selectable target speaker extraction. Trained with three curriculum learning strategies, the model significantly outperforms traditional approaches across a wide SNR range from −1 to 10 dB, achieving a SI-SDR of 9.31 dB (a 4.85 dB improvement), PESQ of 2.13, and STOI of 0.84, thereby effectively mitigating performance degradation in low-SNR and small-aperture scenarios.

0 citationsRead paper

Recent publications

Latest Papers

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

Aug 12, 2026

This study addresses two critical questions in glacial lake outburst floods, landslides, and glacier-related flooding in the Himalayas: where hazards are prone to occur and when they are triggered. For the first time, it decouples deformation and meteorological signals into distinct spatial susceptibility and temporal triggering prediction tasks, leveraging only freely available satellite data (e.g., InSAR and weather observations) and topographic features to build a self-contained predictive model that does not rely on neighboring-region information. Rigorous spatiotemporal cross-validation mitigates geographic overfitting, revealing that simpler models—such as gradient-boosted trees—outperform complex deep learning approaches. Meteorological indicators achieve AUC scores of 0.73–0.83 for trigger timing prediction, whereas terrain-based features show limited discriminative power for susceptibility (AUC 0.54–0.76). These findings underpin a prioritized hazard risk inventory for Nepal.

0 citationsRead paper

iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

Aug 06, 2026

Existing 3D scene generation methods struggle to reliably satisfy task-critical functional constraints such as navigability and reachability, limiting the practical utility of synthetic data. This work proposes an iterative agent-based reinforcement learning framework that first enhances physical plausibility and layout quality through pretraining with generic rewards, then leverages a large language model (LLM) to generate executable, task-specific reward programs. These LLM-generated rewards are integrated into a feedback-driven reinforcement learning loop for iterative refinement. By uniquely combining LLM-synthesized reward functions with iterative reinforcement learning, the approach significantly improves adherence to functional constraints while preserving scene diversity, thereby enhancing downstream task performance.

0 citationsRead paper

Beyond Edge Maps: Wavelet-Domain Conditioning for Multi-Adapter Map-to-Satellite Diffusion

Jul 29, 2026

This study addresses the challenge of outdated satellite basemaps in low-resource regions, where limited collaboration with commercial mapping providers hinders timely updates. To tackle this issue, the authors propose a ControlNet-based diffusion model that generates high-quality satellite imagery using only readily available OpenStreetMap raster data and its stationary wavelet transform (SWT) subbands as conditioning inputs. A multi-adapter fusion strategy jointly guides both spatial structure and frequency-domain details without requiring model retraining for multiple conditions. Notably, this work is the first to incorporate SWT subbands into map-to-image generation. Evaluated on a newly curated Nepal dataset and the Pix2Pix benchmark, the proposed method outperforms existing approaches on six out of eight metrics, with SWT-only conditioning achieving the lowest Fréchet Inception Distance (FID) on both datasets.

0 citationsRead paper

An Evolutionary Approach for Designing Stable and Highly Expressible Low-Immunogenicity Therapeutic mRNA Sequences

May 27, 2026

This study addresses the challenge of designing therapeutic mRNA sequences that simultaneously maximize translation efficiency, structural stability, and low immunogenicity—objectives often in tension with one another. To this end, the authors propose a two-stage computational framework: first, a pre-trained CodonTransformer, a BERT-like large language model, generates candidate sequences encoding the target antigen; second, a multi-objective genetic algorithm incorporating human codon usage bias, synonymous mutations, and codon-aware crossover refines these candidates. This approach represents the first integration of large language models with evolutionary optimization for mRNA design and substantially outperforms baseline methods such as LinearDesign and BiLSTM-CRF. It achieves state-of-the-art performance across key metrics, including CAI (0.73–0.74), tAI (0.63–0.64), 5′-end unpaired proportion (0.87), global minimum free energy (−346 to −356 kcal/mol), and immunogenicity penalty (27.3).

0 citationsRead paper

IsoNet: Spatially-aware audio-visual target speech extraction in complex acoustic environments

May 14, 2026

This work addresses the challenge of target speech extraction on compact devices equipped with small-aperture microphone arrays, where conventional beamforming and monaural neural models struggle to perform effectively. To overcome this limitation, we propose IsoNet—a multimodal U-Net-based mask estimation network tailored for four-microphone arrays. IsoNet uniquely integrates complex-valued multi-channel STFT features, GCC-PHAT spatial cues, face-conditioned visual embeddings, and DOA-assisted supervision, enabling user-selectable target speaker extraction. Trained with three curriculum learning strategies, the model significantly outperforms traditional approaches across a wide SNR range from −1 to 10 dB, achieving a SI-SDR of 9.31 dB (a 4.85 dB improvement), PESQ of 2.13, and STOI of 0.84, thereby effectively mitigating performance degradation in low-SNR and small-aperture scenarios.

0 citationsRead paper