Selected work

Representative Papers

Chainwash: Multi-Step Rewriting Attacks on Diffusion Language Model Watermarks

May 06, 2026

This study addresses the vulnerability of existing language model watermarks under iterative rewriting attacks, which significantly degrade detection reliability. The work systematically evaluates the robustness of watermarking schemes in diffusion-based language models subjected to multi-round, multi-style chained rewrites—including paraphrasing, simplification, and academic stylization—and reveals, for the first time, the cumulative destructive effect of sequential rewrites on watermark integrity. Experiments employ four open-source large language models (1.5B–8B parameters) within a statistical watermark detection framework, subjecting generated text to five rounds of chained rewrites. Results demonstrate a sharp decline in detection efficacy: from an initial rate of 87.9% to 14%–41% after a single rewrite, and further down to merely 4.86% after five rounds, establishing multi-step rewriting as a highly effective watermark removal attack.

0 citationsRead paper

Empirical Evidence of Complexity-Induced Limits in Large Language Models on Finite Discrete State-Space Problems with Explicit Validity Constraints

Apr 14, 2026

This study identifies and formalizes a phenomenon termed “reasoning collapse,” wherein large language models exhibit a sharp performance drop on discrete-state reasoning tasks as their complexity increases. To systematically investigate this issue, the authors introduce a controlled evaluation framework comprising nine canonical reasoning task categories, featuring parameterized task generation, deterministic solution verification, and multi-level complexity stratification. The work demonstrates that while models perform robustly at low complexity levels, their accuracy plummets by over 50% once task-specific complexity thresholds are exceeded. This abrupt decline is accompanied by systematic failures—including constraint violations, state omissions, and highly confident incorrect predictions—highlighting the critical role of dynamic complexity in undermining model reasoning capabilities.

0 citationsRead paper

Prompt Control-Flow Integrity: A Priority-Aware Runtime Defense Against Prompt Injection in LLM Systems

Mar 18, 2026

This work addresses the vulnerability of large language models to prompt injection attacks under API and retrieval-augmented generation (RAG) architectures, which can lead to policy bypasses and unsafe outputs. The authors propose PCFI, a runtime defense mechanism that introduces, for the first time, role-origin- and priority-based prompt integrity control. It models prompts as structured segments—system, developer, user, and retrieved documents—and enforces priority-aware policies through a three-stage middleware pipeline. By integrating lexical heuristics, role-switching detection, and hierarchical policy enforcement, PCFI achieves a 100% attack interception rate with 0% false positives on a custom test suite, incurring only a median processing overhead of 0.04 milliseconds.

0 citationsRead paper

Text to Automata Diagrams: Comparing TikZ Code Generation with Direct Image Synthesis

Mar 09, 2026

This study addresses the challenge of automatically processing student-drawn automaton diagrams for educational feedback, given their high variability in structure, layout, and correctness. The authors propose a two-stage approach: first, a vision-language model generates textual descriptions from scanned images, which are then manually corrected and fed into a large language model to produce TikZ code that compiles into standardized diagrams for fidelity evaluation. This work presents the first systematic comparison between text-mediated and direct image-based methods for automaton diagram reconstruction, highlighting the critical role of human-in-the-loop correction in enhancing pipeline accuracy. Experimental results demonstrate that even minimal manual refinement of the generated descriptions substantially improves the correctness of the resulting TikZ diagrams, offering a viable pathway toward automated grading and the generation of accessible instructional materials.

0 citationsRead paper

Prompt-Based Caption Generation for Single-Tooth Dental Images Using Vision-Language Models

Mar 08, 2026

Existing dental visual-language datasets primarily focus on full-mouth views or specific pathologies, lacking fine-grained annotations at the individual tooth level, which hinders the training of precise vision-language models. To address this gap, this work proposes a prompt-driven description generation framework that leverages vision-language models (VLMs) with carefully designed instructional prompts to produce semantically rich and visually aligned dental descriptions from single-tooth RGB images. Experimental results demonstrate that the proposed approach significantly improves both the quality of generated text and its alignment with image content, thereby enhancing VLM performance on single-tooth visual understanding tasks and filling a critical void in fine-grained dental visual-language data.

0 citationsRead paper

Recent publications

Latest Papers

Chainwash: Multi-Step Rewriting Attacks on Diffusion Language Model Watermarks

May 06, 2026

This study addresses the vulnerability of existing language model watermarks under iterative rewriting attacks, which significantly degrade detection reliability. The work systematically evaluates the robustness of watermarking schemes in diffusion-based language models subjected to multi-round, multi-style chained rewrites—including paraphrasing, simplification, and academic stylization—and reveals, for the first time, the cumulative destructive effect of sequential rewrites on watermark integrity. Experiments employ four open-source large language models (1.5B–8B parameters) within a statistical watermark detection framework, subjecting generated text to five rounds of chained rewrites. Results demonstrate a sharp decline in detection efficacy: from an initial rate of 87.9% to 14%–41% after a single rewrite, and further down to merely 4.86% after five rounds, establishing multi-step rewriting as a highly effective watermark removal attack.

0 citationsRead paper

Empirical Evidence of Complexity-Induced Limits in Large Language Models on Finite Discrete State-Space Problems with Explicit Validity Constraints

Apr 14, 2026

This study identifies and formalizes a phenomenon termed “reasoning collapse,” wherein large language models exhibit a sharp performance drop on discrete-state reasoning tasks as their complexity increases. To systematically investigate this issue, the authors introduce a controlled evaluation framework comprising nine canonical reasoning task categories, featuring parameterized task generation, deterministic solution verification, and multi-level complexity stratification. The work demonstrates that while models perform robustly at low complexity levels, their accuracy plummets by over 50% once task-specific complexity thresholds are exceeded. This abrupt decline is accompanied by systematic failures—including constraint violations, state omissions, and highly confident incorrect predictions—highlighting the critical role of dynamic complexity in undermining model reasoning capabilities.

0 citationsRead paper

Prompt Control-Flow Integrity: A Priority-Aware Runtime Defense Against Prompt Injection in LLM Systems

Mar 18, 2026

This work addresses the vulnerability of large language models to prompt injection attacks under API and retrieval-augmented generation (RAG) architectures, which can lead to policy bypasses and unsafe outputs. The authors propose PCFI, a runtime defense mechanism that introduces, for the first time, role-origin- and priority-based prompt integrity control. It models prompts as structured segments—system, developer, user, and retrieved documents—and enforces priority-aware policies through a three-stage middleware pipeline. By integrating lexical heuristics, role-switching detection, and hierarchical policy enforcement, PCFI achieves a 100% attack interception rate with 0% false positives on a custom test suite, incurring only a median processing overhead of 0.04 milliseconds.

0 citationsRead paper

Text to Automata Diagrams: Comparing TikZ Code Generation with Direct Image Synthesis

Mar 09, 2026

This study addresses the challenge of automatically processing student-drawn automaton diagrams for educational feedback, given their high variability in structure, layout, and correctness. The authors propose a two-stage approach: first, a vision-language model generates textual descriptions from scanned images, which are then manually corrected and fed into a large language model to produce TikZ code that compiles into standardized diagrams for fidelity evaluation. This work presents the first systematic comparison between text-mediated and direct image-based methods for automaton diagram reconstruction, highlighting the critical role of human-in-the-loop correction in enhancing pipeline accuracy. Experimental results demonstrate that even minimal manual refinement of the generated descriptions substantially improves the correctness of the resulting TikZ diagrams, offering a viable pathway toward automated grading and the generation of accessible instructional materials.

0 citationsRead paper

Prompt-Based Caption Generation for Single-Tooth Dental Images Using Vision-Language Models

Mar 08, 2026

Existing dental visual-language datasets primarily focus on full-mouth views or specific pathologies, lacking fine-grained annotations at the individual tooth level, which hinders the training of precise vision-language models. To address this gap, this work proposes a prompt-driven description generation framework that leverages vision-language models (VLMs) with carefully designed instructional prompts to produce semantically rich and visually aligned dental descriptions from single-tooth RGB images. Experimental results demonstrate that the proposed approach significantly improves both the quality of generated text and its alignment with image content, thereby enhancing VLM performance on single-tooth visual understanding tasks and filling a critical void in fine-grained dental visual-language data.

0 citationsRead paper