Selected work

Representative Papers

Linear space streaming lower bounds for approximating CSPs

Jun 24, 2021Electron. Colloquium Comput. Complex.

This work investigates the approximability threshold of constraint satisfaction problems (CSPs) in the streaming model. For $n$-variable CSPs over domain ${0,dots,q-1}$ with $O(n)$ constraints, we prove that any algorithm achieving approximation ratio strictly better than the trivial $1/q$ requires $Omega(n)$ space—establishing the first linear-space lower bound for approximation ratios below $1/2$. Methodologically, we extend the Kapralov–Krachun linear lower-bound technique to general CSPs (surpassing prior $Omega(sqrt{n})$ bounds) via modular-$q$ linear equation encoding, communication complexity analysis, and pseudorandom hard-instance reduction. This yields optimal $q^{-(k-1)}$ inapproximability for Max $k$-LIN mod $q$ with $k>2$, $q>2$. Our results uniformly characterize the streaming hardness of broad CSP subclasses: all nontrivial approximation requires essentially linear space, significantly advancing the theoretical understanding of streaming algorithm limitations.

15 citationsRead paper

Disentangling the Expressivity of RoPE

Aug 12, 2026

This work clarifies the fundamental distinction between two theoretical accounts of Rotary Position Embedding (RoPE)—its capacity for expressing periodic modular predicates versus its reliance on a position-anchor-based local offset mechanism—and investigates their respective implications for generalization. Through a formal analysis of RoPE’s behavior in finite-precision soft-attention Transformers, the study establishes, for the first time, a rigorous correspondence between RoPE’s expressivity and classes of languages definable in temporal logic with modular predicates. The findings reveal that periodic RoPE enables length generalization on modular language tasks, whereas standard non-repeating RoPE merely simulates precision-dependent, bounded lookback operations, exhibiting a strong locality bias that severely degrades performance on tasks requiring long-range, position-invariant memory access.

0 citationsRead paper

SABRE: Scalable and Automated Benchmarking of VLMs under Stress

Aug 07, 2026

This work addresses the limitations of existing vision-language model (VLM) benchmarks in effectively uncovering model weaknesses under stress conditions and the high cost of manually curating high-quality test samples. To this end, we propose SABRE, a novel framework that enables end-to-end automated generation of stress-test datasets directly from task design specifications. SABRE establishes a human-in-the-loop data production pipeline through structured specification parsing, image generation and editing, automated filtering, and human validation. The framework supports multidimensional stress testing—such as context, texture, and attribute manipulation—and is both extensible and reusable. Using SABRE, we construct the SABRE-Prior benchmark (600 images, 1,000 questions), on which six state-of-the-art VLMs achieve an average accuracy of only 22.6%, revealing their overreliance on world priors rather than visual evidence. The framework also demonstrates strong transferability across tasks like counting and spatial reasoning.

0 citationsRead paper

Hierarchical Domain Generalization

Jul 17, 2026

This study addresses the problem of domain generalization under arbitrary domain hierarchies, focusing on how to extrapolate from finitely observed domains to the entire instance space. Departing from the conventional i.i.d. assumption, it treats the domain partition structure as a first-class object in generalization theory, revealing that the very way training and test domains are partitioned is a fundamental source of generalization failure. By integrating theoretical analyses of hypothesis class complexity with domain structure, the work characterizes the intrinsic limits of extrapolation and proves that, even with an extremely simple hypothesis class or infinitely many training samples, certain domain partitions inevitably lead to failure in generalizing to target domains. This provides a novel theoretical perspective and structural understanding of domain generalization.

0 citationsRead paper

Autoregressive Modeling of Film with Applications in Video Montage

Jul 16, 2026

This work addresses the problem of video montage—automatically editing raw footage into coherent cinematic sequences. It proposes FilmGPT, the first framework to adapt the autoregressive modeling paradigm of large language models to video editing. Built upon a long-context Transformer architecture, FilmGPT is pretrained on a large-scale film corpus to learn the “grammatical” structure of shot sequences. The authors introduce a source-constrained decoding algorithm that selects an optimal sequence from given shots without generating new frames. Requiring no handcrafted rules, the method significantly outperforms existing approaches, achieving state-of-the-art performance on standard ordering benchmarks and in user studies. It has been successfully deployed in applications such as automatic clip trimming and human-in-the-loop collaborative editing.

0 citationsRead paper

Recent publications

Latest Papers

Disentangling the Expressivity of RoPE

Aug 12, 2026

This work clarifies the fundamental distinction between two theoretical accounts of Rotary Position Embedding (RoPE)—its capacity for expressing periodic modular predicates versus its reliance on a position-anchor-based local offset mechanism—and investigates their respective implications for generalization. Through a formal analysis of RoPE’s behavior in finite-precision soft-attention Transformers, the study establishes, for the first time, a rigorous correspondence between RoPE’s expressivity and classes of languages definable in temporal logic with modular predicates. The findings reveal that periodic RoPE enables length generalization on modular language tasks, whereas standard non-repeating RoPE merely simulates precision-dependent, bounded lookback operations, exhibiting a strong locality bias that severely degrades performance on tasks requiring long-range, position-invariant memory access.

0 citationsRead paper

SABRE: Scalable and Automated Benchmarking of VLMs under Stress

Aug 07, 2026

This work addresses the limitations of existing vision-language model (VLM) benchmarks in effectively uncovering model weaknesses under stress conditions and the high cost of manually curating high-quality test samples. To this end, we propose SABRE, a novel framework that enables end-to-end automated generation of stress-test datasets directly from task design specifications. SABRE establishes a human-in-the-loop data production pipeline through structured specification parsing, image generation and editing, automated filtering, and human validation. The framework supports multidimensional stress testing—such as context, texture, and attribute manipulation—and is both extensible and reusable. Using SABRE, we construct the SABRE-Prior benchmark (600 images, 1,000 questions), on which six state-of-the-art VLMs achieve an average accuracy of only 22.6%, revealing their overreliance on world priors rather than visual evidence. The framework also demonstrates strong transferability across tasks like counting and spatial reasoning.

0 citationsRead paper

Hierarchical Domain Generalization

Jul 17, 2026

This study addresses the problem of domain generalization under arbitrary domain hierarchies, focusing on how to extrapolate from finitely observed domains to the entire instance space. Departing from the conventional i.i.d. assumption, it treats the domain partition structure as a first-class object in generalization theory, revealing that the very way training and test domains are partitioned is a fundamental source of generalization failure. By integrating theoretical analyses of hypothesis class complexity with domain structure, the work characterizes the intrinsic limits of extrapolation and proves that, even with an extremely simple hypothesis class or infinitely many training samples, certain domain partitions inevitably lead to failure in generalizing to target domains. This provides a novel theoretical perspective and structural understanding of domain generalization.

0 citationsRead paper

Autoregressive Modeling of Film with Applications in Video Montage

Jul 16, 2026

This work addresses the problem of video montage—automatically editing raw footage into coherent cinematic sequences. It proposes FilmGPT, the first framework to adapt the autoregressive modeling paradigm of large language models to video editing. Built upon a long-context Transformer architecture, FilmGPT is pretrained on a large-scale film corpus to learn the “grammatical” structure of shot sequences. The authors introduce a source-constrained decoding algorithm that selects an optimal sequence from given shots without generating new frames. Requiring no handcrafted rules, the method significantly outperforms existing approaches, achieving state-of-the-art performance on standard ordering benchmarks and in user studies. It has been successfully deployed in applications such as automatic clip trimming and human-in-the-loop collaborative editing.

0 citationsRead paper

Threshold Rounding and Bounded-Degree Boolean MAX 2-CSP

Jul 12, 2026

This work proposes an improved threshold rounding algorithm for Boolean MAX 2-CSP instances where each variable appears in at most $d$ constraints. By integrating semidefinite programming with a refined analysis based on graph degree constraints, the algorithm achieves—for the first time—a $\widetilde{\Omega}(1/d^4)$ improvement in approximation ratio across a broad class of such problems. Notably, for MAX 2-SAT, it attains an approximation factor of $(\beta^* + \widetilde{\Omega}(1/d^2))$, where $\beta^*$ denotes the optimal constant-factor approximation achievable without degree constraints. This result generalizes the bounded-degree MAX CUT guarantee of Hsieh and Kothari and provides a theoretical foundation for improved approximation algorithms for related problems such as MAX DI-CUT and MAX 2-AND.

0 citationsRead paper