Selected work

Representative Papers

BioSentinel at EXIST 2026: Soft-Label Optimization with XLM-RoBERTa for Sexism Intent Classification in Memes

Jul 27, 2026

This work addresses the detection of gender bias in online memes by classifying instances into direct, judgmental, or non-gender-discriminatory categories, while simultaneously predicting both hard labels and soft labels in the form of probability distributions. Building upon the XLM-RoBERTa architecture, the proposed approach employs a composite loss function that integrates KL divergence with weighted cross-entropy within the Learning with Disagreement framework to explicitly model annotator disagreement. Temperature scaling is further incorporated to refine soft-label predictions. This methodology enhances modeling capacity for subjective tasks. On the official test set, the model achieves an ICM-Soft-Norm score of 0.3229, an ICM-Norm score of 0.3778, and a hard-label F1-score of 0.4236, ranking 40th out of 118 teams in soft–soft evaluation and 49th out of 187 in hard–hard evaluation.

0 citationsRead paper

SpiralFovea: Input-Adaptive Foveated Tokenization as a Third Lever of Resource-Adaptive Inference

Jul 01, 2026

This work addresses the limitation of prevailing vision foundation models that rely on fixed-grid tokenization at the input stage, overlooking the potential for content-adaptive optimization. The authors propose a parameter-free, input-adaptive tokenizer that dynamically determines the identity, position, scale, and number of tokens based on local visual entropy before the backbone network. Instead of standard ViT grid patches, it generates multi-scale, spiral-ring patches centered on content-rich regions. This approach establishes input-adaptive tokenization as a third lever—alongside model architecture and inference algorithms—for resource-efficient reasoning, enabling content-aware tokenization and precise allocation of computational resources. Experiments on four fine-grained benchmarks show a 60% reduction in input tokens, an 84% decrease in per-layer self-attention FLOPs, 18–29% higher throughput, and consistent accuracy gains of 1.7–2.1 percentage points.

0 citationsRead paper

LAMP: Lean-based Agentic framework with MCP and Proof Repair

Jun 27, 2026

This work addresses the challenge of verifying mathematical proofs generated by large language models, which stems from insufficient coverage of formal knowledge bases such as combinatorics theory. To bridge this gap, the paper presents the first systematic formalization of combinatorics theory—encompassing 93 core definitions and lemmas—and integrates it into Lean 4. Building upon this foundation, the authors propose LAMP, a multi-agent framework that, without fine-tuning, leverages the Model Context Protocol (MCP) to coordinate a planner, builder, and verifier. By explicitly supplying structured domain knowledge through a formal ontology, LAMP synthesizes proofs verifiable by the Lean kernel. Evaluated on a benchmark of 90 theorems spanning eight modules and three difficulty levels, LAMP successfully generates correct proofs for 96.7% of the problems, substantially outperforming baseline methods and existing specialized provers, with ablation studies confirming the efficacy of its architectural design.

0 citationsRead paper

Enhanced Low-Density Region Exploration in Classifier-Guided Diffusion Models Through Modified Reverse Diffusion Sampling

Jun 11, 2026

Standard classifier-guided diffusion models tend to generate samples predominantly from high-density regions of the data distribution, struggling to capture rare modes in the tails of class-conditional distributions. This work proposes a density-aware guidance approach that operates solely during sampling, requiring no additional training or architectural modifications. The method employs a dual guidance mechanism: it modulates classifier gradients to encourage exploration of low-confidence (low-density) regions while simultaneously constraining the generation trajectory to remain close to the true data manifold. Leveraging a pretrained ADM and an ImageNet classifier, the approach significantly improves recall at 64×64 resolution without degrading FID, and at 256×256 resolution achieves both high perceptual quality and diversity, effectively balancing the generation of rare samples with overall fidelity.

0 citationsRead paper

The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol

Jun 03, 2026

This study presents the first multidimensional empirical analysis of a deployed Proof-of-Useful-Work (PoUW) system—specifically, the cuPOW protocol of the Pearl blockchain—to evaluate whether it genuinely performs useful AI inference while maintaining security. Through network node mapping, open-source miner experimentation, statistical distribution testing, hardware compatibility assessments, and an ROI-based economic model, the research reveals that although Pearl nodes possess AI-capable hardware, they execute no meaningful AI computations. The verification mechanism is shown to be vulnerable to spoofing with random data, mining operations consistently incur financial losses, and computational tasks exhibit no hardware lock-in effect. This work exposes the fundamental tension between verifiability and usefulness in PoUW designs and elucidates its adverse economic implications.

0 citationsRead paper

Recent publications

Latest Papers

BioSentinel at EXIST 2026: Soft-Label Optimization with XLM-RoBERTa for Sexism Intent Classification in Memes

Jul 27, 2026

This work addresses the detection of gender bias in online memes by classifying instances into direct, judgmental, or non-gender-discriminatory categories, while simultaneously predicting both hard labels and soft labels in the form of probability distributions. Building upon the XLM-RoBERTa architecture, the proposed approach employs a composite loss function that integrates KL divergence with weighted cross-entropy within the Learning with Disagreement framework to explicitly model annotator disagreement. Temperature scaling is further incorporated to refine soft-label predictions. This methodology enhances modeling capacity for subjective tasks. On the official test set, the model achieves an ICM-Soft-Norm score of 0.3229, an ICM-Norm score of 0.3778, and a hard-label F1-score of 0.4236, ranking 40th out of 118 teams in soft–soft evaluation and 49th out of 187 in hard–hard evaluation.

0 citationsRead paper

SpiralFovea: Input-Adaptive Foveated Tokenization as a Third Lever of Resource-Adaptive Inference

Jul 01, 2026

This work addresses the limitation of prevailing vision foundation models that rely on fixed-grid tokenization at the input stage, overlooking the potential for content-adaptive optimization. The authors propose a parameter-free, input-adaptive tokenizer that dynamically determines the identity, position, scale, and number of tokens based on local visual entropy before the backbone network. Instead of standard ViT grid patches, it generates multi-scale, spiral-ring patches centered on content-rich regions. This approach establishes input-adaptive tokenization as a third lever—alongside model architecture and inference algorithms—for resource-efficient reasoning, enabling content-aware tokenization and precise allocation of computational resources. Experiments on four fine-grained benchmarks show a 60% reduction in input tokens, an 84% decrease in per-layer self-attention FLOPs, 18–29% higher throughput, and consistent accuracy gains of 1.7–2.1 percentage points.

0 citationsRead paper

LAMP: Lean-based Agentic framework with MCP and Proof Repair

Jun 27, 2026

This work addresses the challenge of verifying mathematical proofs generated by large language models, which stems from insufficient coverage of formal knowledge bases such as combinatorics theory. To bridge this gap, the paper presents the first systematic formalization of combinatorics theory—encompassing 93 core definitions and lemmas—and integrates it into Lean 4. Building upon this foundation, the authors propose LAMP, a multi-agent framework that, without fine-tuning, leverages the Model Context Protocol (MCP) to coordinate a planner, builder, and verifier. By explicitly supplying structured domain knowledge through a formal ontology, LAMP synthesizes proofs verifiable by the Lean kernel. Evaluated on a benchmark of 90 theorems spanning eight modules and three difficulty levels, LAMP successfully generates correct proofs for 96.7% of the problems, substantially outperforming baseline methods and existing specialized provers, with ablation studies confirming the efficacy of its architectural design.

0 citationsRead paper

Enhanced Low-Density Region Exploration in Classifier-Guided Diffusion Models Through Modified Reverse Diffusion Sampling

Jun 11, 2026

Standard classifier-guided diffusion models tend to generate samples predominantly from high-density regions of the data distribution, struggling to capture rare modes in the tails of class-conditional distributions. This work proposes a density-aware guidance approach that operates solely during sampling, requiring no additional training or architectural modifications. The method employs a dual guidance mechanism: it modulates classifier gradients to encourage exploration of low-confidence (low-density) regions while simultaneously constraining the generation trajectory to remain close to the true data manifold. Leveraging a pretrained ADM and an ImageNet classifier, the approach significantly improves recall at 64×64 resolution without degrading FID, and at 256×256 resolution achieves both high perceptual quality and diversity, effectively balancing the generation of rare samples with overall fidelity.

0 citationsRead paper

The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol

Jun 03, 2026

This study presents the first multidimensional empirical analysis of a deployed Proof-of-Useful-Work (PoUW) system—specifically, the cuPOW protocol of the Pearl blockchain—to evaluate whether it genuinely performs useful AI inference while maintaining security. Through network node mapping, open-source miner experimentation, statistical distribution testing, hardware compatibility assessments, and an ROI-based economic model, the research reveals that although Pearl nodes possess AI-capable hardware, they execute no meaningful AI computations. The verification mechanism is shown to be vulnerable to spoofing with random data, mining operations consistently incur financial losses, and computational tasks exhibit no hardware lock-in effect. This work exposes the fundamental tension between verifiability and usefulness in PoUW designs and elucidates its adverse economic implications.

0 citationsRead paper