Institution profile

National Chengchi University

Academic institutionasia · tw
Official website
Research library29linked papers
Opportunities0open roles
Selected work

Representative Papers

Do LLMs Understand Sequential Structure? A Controlled Study of Inference and Generation

Oct 04, 2026

This study investigates whether large language models can transcend surface-level frequencies to identify latent sequential structures, thereby addressing conditional dependency failures in behavioral simulation. To this end, it proposes an evaluation framework that distinguishes inference from generation, decoupling distribution matching from rule adherence through controlled games such as Rock-Paper-Scissors and N-gram continuation tasks. Combined with Markov chain analysis, this approach systematically assesses the models’ reasoning and generative capacities regarding higher-order dependencies. The findings reveal the mechanisms by which long-context ineffectiveness and higher-order dependencies cause significant degradation in rule recovery. Furthermore, this work demonstrates that correct identification does not guarantee faithful simulation, highlighting that superficial behavioral fidelity may obscure erroneous underlying generative mechanisms.

0 citationsRead paper

FORGE: Verification-Gated Behavioral Repair for Generative Language Models

Oct 04, 2026

This study addresses the challenge of mitigating post-deployment bias and toxicity in large language models (LLMs) while providing certified repairs that preserve original capabilities. To this end, it proposes FORGE, a novel decoupled architecture that disentangles defect localization, weight editing, and behavioral verification. By integrating constrained quadratic optimization, null-space projection, and causal probing techniques with a repair abstraction mechanism for edit-agnosticism, the framework achieves per-sample correctness guarantees for autoregressive generation. Extensive evaluations across five open-source LLMs demonstrate that FORGE substantially reduces bias and toxicity with minimal perplexity degradation, outperforming conventional fine-tuning approaches particularly in few-shot scenarios.

0 citationsRead paper

Attention Capture Is Not Detection: A Two-Stage Account of How Humans Miss Localized AI Image Edits

Aug 13, 2026

This study addresses the oversight of dissociated human attention and recognition mechanisms in existing AI image editing detection. We propose a two-stage cognitive model demonstrating that edited regions drive attentional capture while semantic plausibility determines judgment accuracy. As the first work to introduce pre-attentive and recognition distinctions into this domain, we construct a generative eye-movement prediction framework. Validated through eye-tracking and mixed-effects analyses, the significant dissociation between stages is confirmed. The model achieves attention prediction correlations of 0.77–0.82, and its missed-detection behavioral prediction performance (r=0.52) significantly outperforms linear baselines (r=0.48). These findings establish a novel paradigm for understanding detection blind spots in human-AI interaction, highlighting the critical role of cognitive separation in evaluating synthetic imagery.

0 citationsRead paper
Recent publications

Latest Papers

Do LLMs Understand Sequential Structure? A Controlled Study of Inference and Generation

Oct 04, 2026

This study investigates whether large language models can transcend surface-level frequencies to identify latent sequential structures, thereby addressing conditional dependency failures in behavioral simulation. To this end, it proposes an evaluation framework that distinguishes inference from generation, decoupling distribution matching from rule adherence through controlled games such as Rock-Paper-Scissors and N-gram continuation tasks. Combined with Markov chain analysis, this approach systematically assesses the models’ reasoning and generative capacities regarding higher-order dependencies. The findings reveal the mechanisms by which long-context ineffectiveness and higher-order dependencies cause significant degradation in rule recovery. Furthermore, this work demonstrates that correct identification does not guarantee faithful simulation, highlighting that superficial behavioral fidelity may obscure erroneous underlying generative mechanisms.

0 citationsRead paper

FORGE: Verification-Gated Behavioral Repair for Generative Language Models

Oct 04, 2026

This study addresses the challenge of mitigating post-deployment bias and toxicity in large language models (LLMs) while providing certified repairs that preserve original capabilities. To this end, it proposes FORGE, a novel decoupled architecture that disentangles defect localization, weight editing, and behavioral verification. By integrating constrained quadratic optimization, null-space projection, and causal probing techniques with a repair abstraction mechanism for edit-agnosticism, the framework achieves per-sample correctness guarantees for autoregressive generation. Extensive evaluations across five open-source LLMs demonstrate that FORGE substantially reduces bias and toxicity with minimal perplexity degradation, outperforming conventional fine-tuning approaches particularly in few-shot scenarios.

0 citationsRead paper

Attention Capture Is Not Detection: A Two-Stage Account of How Humans Miss Localized AI Image Edits

Aug 13, 2026

This study addresses the oversight of dissociated human attention and recognition mechanisms in existing AI image editing detection. We propose a two-stage cognitive model demonstrating that edited regions drive attentional capture while semantic plausibility determines judgment accuracy. As the first work to introduce pre-attentive and recognition distinctions into this domain, we construct a generative eye-movement prediction framework. Validated through eye-tracking and mixed-effects analyses, the significant dissociation between stages is confirmed. The model achieves attention prediction correlations of 0.77–0.82, and its missed-detection behavioral prediction performance (r=0.52) significantly outperforms linear baselines (r=0.48). These findings establish a novel paradigm for understanding detection blind spots in human-AI interaction, highlighting the critical role of cognitive separation in evaluating synthetic imagery.

0 citationsRead paper