Selected work

Representative Papers

Querying an astronomical database using large language models: the ALeRCE text-to-SQL system

Jun 16, 2026

This work addresses the high barrier posed by SQL-based querying of astronomical databases for non-expert users. The authors propose a large language model (LLM)-driven text-to-SQL system enabling natural language queries over the ALeRCE database. Their approach employs a four-module, stepwise generation framework—comprising schema linking, query classification, prompt decomposition, and self-correction—which substantially outperforms end-to-end baselines and significantly reduces execution errors. Evaluation on 110 annotated samples across 13 LLMs demonstrates strong performance: on models such as Claude Opus 4.6, exact match rates for row and column identifiers reach 0.97 and 0.94, respectively, for simple queries, while maintaining robustness on complex queries.

0 citationsRead paper

Closing the Evidence Gap: reddemcee, a Fast Adaptive Parallel Tempering Sampler

Sep 29, 2025

Traditional MCMC methods excel at posterior sampling but yield substantially less accurate marginal likelihood (evidence) estimates than nested sampling, limiting their utility in astrophysical model comparison. To address this, we propose an adaptive parallel tempering MCMC framework featuring an automated temperature ladder adjustment scheme, affine-invariant stretch moves, five exchange-rate optimization objectives, and three evidence estimators (TI+, SS+, H+). Crucially, we introduce—within an MCMC context for the first time—the joint use of thermodynamic integration with a geometric bridging path to achieve low-bias evidence estimation. On a 15-dimensional test problem, our method achieves ~7× higher sampling efficiency, evidence errors <3%, and well-calibrated uncertainties. Applied to the HD 20794 planetary system, it successfully reproduces published results while delivering tighter parameter constraints; its evidence accuracy matches or surpasses that of dynamic nested sampling.

0 citationsRead paper

Testing chatbots on the creation of encoders for audio conditioned image generation

Sep 09, 2025

This study presents the first systematic evaluation of mainstream conversational agents (e.g., Gemini, Grok) on cross-modal architecture design—specifically, their ability to autonomously generate an audio encoder that replaces the CLIP text encoder in Stable Diffusion 1.5 for end-to-end audio-to-image generation. Leveraging a dataset of over two million audio–image–text triplets, we prompt five large language models to design audio encoder architectures, then uniformly train and evaluate all generated models. Results show that while most models produce syntactically valid architectures, their audio embeddings fail to align with the original CLIP text embedding space. Gemini achieves the best quantitative performance (lowest FID, highest CLIP-Score), whereas Grok yields the most visually coherent generated images. Our work exposes fundamental structural limitations and consistent misalignment biases in current LLMs for cross-modal representation alignment tasks, establishing an empirical benchmark and methodological insights for AI-driven neural architecture search.

0 citationsRead paper

Effectively obtaining acoustic, visual and textual data from videos

Sep 06, 2025

A critical shortage exists of high-quality, large-scale, semantically aligned acoustic–visual–textual multimodal datasets. Method: This paper proposes an end-to-end framework for constructing video-based multimodal data, integrating three key components: (i) video content filtering, (ii) cross-modal synchronization triplet extraction (audio–frame–subtitle), and (iii) fine-grained description synthesis leveraging image-to-text generation models—ensuring temporal and semantic alignment across all three modalities. Contribution/Results: The resulting publicly released dataset spans diverse real-world scenarios and substantially advances performance on cross-modal retrieval and joint embedding learning tasks, achieving state-of-the-art results across multiple benchmarks. By providing a scalable, high-fidelity resource, this work establishes a new foundation for training and evaluating foundational multimodal models.

0 citationsRead paper

Vision Transformer attention alignment with human visual perception in aesthetic object evaluation

Jul 23, 2025

Understanding the alignment between vision transformer (ViT) attention mechanisms and human visual attention in fine-grained aesthetic evaluation—particularly for handicrafts—remains an open challenge. Method: We conducted eye-tracking experiments to record human fixation distributions and generated attention heatmaps using a DINO-pretrained ViT. Spatial similarity between human and model attention was quantified via Kullback–Leibler (KL) divergence. Contribution/Results: We identify that a specific attention head at layer 12 achieves statistically significant alignment with human attention (p < 0.05) under Gaussian kernel smoothing (σ = 2.4), demonstrating both global coverage and selective sensitivity to salient local features. In contrast, other heads exhibit systematic deviations, revealing inherent architectural biases. This is the first empirical evidence that ViT attention can approximate human perceptual mechanisms in domain-specific aesthetic judgment. The findings establish a foundation for interpretable, AI-driven design assessment and introduce a new paradigm for human-aligned visual modeling in computational aesthetics.

0 citationsRead paper

Recent publications

Latest Papers

Querying an astronomical database using large language models: the ALeRCE text-to-SQL system

Jun 16, 2026

This work addresses the high barrier posed by SQL-based querying of astronomical databases for non-expert users. The authors propose a large language model (LLM)-driven text-to-SQL system enabling natural language queries over the ALeRCE database. Their approach employs a four-module, stepwise generation framework—comprising schema linking, query classification, prompt decomposition, and self-correction—which substantially outperforms end-to-end baselines and significantly reduces execution errors. Evaluation on 110 annotated samples across 13 LLMs demonstrates strong performance: on models such as Claude Opus 4.6, exact match rates for row and column identifiers reach 0.97 and 0.94, respectively, for simple queries, while maintaining robustness on complex queries.

0 citationsRead paper

Closing the Evidence Gap: reddemcee, a Fast Adaptive Parallel Tempering Sampler

Sep 29, 2025

Traditional MCMC methods excel at posterior sampling but yield substantially less accurate marginal likelihood (evidence) estimates than nested sampling, limiting their utility in astrophysical model comparison. To address this, we propose an adaptive parallel tempering MCMC framework featuring an automated temperature ladder adjustment scheme, affine-invariant stretch moves, five exchange-rate optimization objectives, and three evidence estimators (TI+, SS+, H+). Crucially, we introduce—within an MCMC context for the first time—the joint use of thermodynamic integration with a geometric bridging path to achieve low-bias evidence estimation. On a 15-dimensional test problem, our method achieves ~7× higher sampling efficiency, evidence errors <3%, and well-calibrated uncertainties. Applied to the HD 20794 planetary system, it successfully reproduces published results while delivering tighter parameter constraints; its evidence accuracy matches or surpasses that of dynamic nested sampling.

0 citationsRead paper

Testing chatbots on the creation of encoders for audio conditioned image generation

Sep 09, 2025

This study presents the first systematic evaluation of mainstream conversational agents (e.g., Gemini, Grok) on cross-modal architecture design—specifically, their ability to autonomously generate an audio encoder that replaces the CLIP text encoder in Stable Diffusion 1.5 for end-to-end audio-to-image generation. Leveraging a dataset of over two million audio–image–text triplets, we prompt five large language models to design audio encoder architectures, then uniformly train and evaluate all generated models. Results show that while most models produce syntactically valid architectures, their audio embeddings fail to align with the original CLIP text embedding space. Gemini achieves the best quantitative performance (lowest FID, highest CLIP-Score), whereas Grok yields the most visually coherent generated images. Our work exposes fundamental structural limitations and consistent misalignment biases in current LLMs for cross-modal representation alignment tasks, establishing an empirical benchmark and methodological insights for AI-driven neural architecture search.

0 citationsRead paper

Effectively obtaining acoustic, visual and textual data from videos

Sep 06, 2025

A critical shortage exists of high-quality, large-scale, semantically aligned acoustic–visual–textual multimodal datasets. Method: This paper proposes an end-to-end framework for constructing video-based multimodal data, integrating three key components: (i) video content filtering, (ii) cross-modal synchronization triplet extraction (audio–frame–subtitle), and (iii) fine-grained description synthesis leveraging image-to-text generation models—ensuring temporal and semantic alignment across all three modalities. Contribution/Results: The resulting publicly released dataset spans diverse real-world scenarios and substantially advances performance on cross-modal retrieval and joint embedding learning tasks, achieving state-of-the-art results across multiple benchmarks. By providing a scalable, high-fidelity resource, this work establishes a new foundation for training and evaluating foundational multimodal models.

0 citationsRead paper

Vision Transformer attention alignment with human visual perception in aesthetic object evaluation

Jul 23, 2025

Understanding the alignment between vision transformer (ViT) attention mechanisms and human visual attention in fine-grained aesthetic evaluation—particularly for handicrafts—remains an open challenge. Method: We conducted eye-tracking experiments to record human fixation distributions and generated attention heatmaps using a DINO-pretrained ViT. Spatial similarity between human and model attention was quantified via Kullback–Leibler (KL) divergence. Contribution/Results: We identify that a specific attention head at layer 12 achieves statistically significant alignment with human attention (p < 0.05) under Gaussian kernel smoothing (σ = 2.4), demonstrating both global coverage and selective sensitivity to salient local features. In contrast, other heads exhibit systematic deviations, revealing inherent architectural biases. This is the first empirical evidence that ViT attention can approximate human perceptual mechanisms in domain-specific aesthetic judgment. The findings establish a foundation for interpretable, AI-driven design assessment and introduce a new paradigm for human-aligned visual modeling in computational aesthetics.

0 citationsRead paper