arXiv subjects

Jason Li

Publications and source records attributed to Jason Li.

At least 55 records · Page 3Linked to original sources

arXiv2025

Rethinking Cross-Generator Image Forgery Detection through DINOv3

As generative models become increasingly diverse and powerful, cross-generator detection has emerged as a new challenge. Existing detection methods often memorize artifacts of specific generative models rather than learning transferable cues, leading to substantial failures on unseen generators. Surprisingly, this work finds that frozen visual foundation models, especially DINOv3, already exhibit strong cross-generator detection capability without any fine-tuning. Through systematic studies on frequency, spatial, and token perspectives, we observe that DINOv3 tends to rely on global, low-frequency structures as weak but transferable authenticity cues instead of high-frequency, generator-specific artifacts. Motivated by this insight, we introduce a simple, training-free token-ranking strategy followed by a lightweight linear probe to select a small subset of authenticity-relevant tokens. This token subset consistently improves detection accuracy across all evaluated datasets. Our study provides empirical evidence and a feasible hypothesis for understanding why foundation models generalize across diverse generators, offering a universal, efficient, and interpretable baseline for image forgery detection.

cs.CV↗

arXiv2025

Shortcutting for Negative-Weight Shortest Path

Consider the single-source shortest paths problem on a directed graph with real-valued edge weights. We solve this problem in $O(n^{2.5}\log^{4.5}n)$ time, improving on prior work of Fineman (STOC 2024) and Huang-Jin-Quanrud (SODA 2025, 2026) on dense graphs. Our main technique is an shortcutting procedure that iteratively reduces the number of negative-weight edges along shortest paths by a constant factor.

cs.DS↗

arXiv2025

Improved Directed Expander Decompositions

We obtain faster expander decomposition algorithms for directed graphs, matching the guarantees of Saranurak and Wang (SODA 2019) for expander decomposition on undirected graphs. Our algorithms are faster than prior work and also generalize almost losslessly to capacitated graphs. In particular, we obtain the first directed expander decomposition algorithm for capacitated graphs in near-linear time with optimal dependence on $ϕ$. To obtain our result, we provide the first implementation and analysis of the non-stop cut-matching game for directed, capacitated graphs. All existing directed expander decomposition algorithms instead temporarily add ''fake edges'' before pruning them away in a final cleanup step. Our result shows that the natural undirected approach applies even to directed graphs. The difficulty is in its analysis, which is technical and requires significant modifications from the original setting of undirected graphs.

cs.DS↗

arXiv2025

Faster Weak Expander Decompositions and Approximate Max Flow

We give faster algorithms for weak expander decompositions and approximate max flow on undirected graphs. First, we show that it is possible to "warm start" the cut-matching game when computing weak expander decompositions, avoiding the cost of the recursion depth. Our algorithm is also flexible enough to support weaker flow subroutines than previous algorithms. Our second contribution is to streamline the recent non-recursive approximate max flow algorithm of Li, Rao, and Wang (SODA, 2025) and adapt their framework to use our new weak expander decomposition primitive. Consequently, we give an approximate max flow algorithm within a few logarithmic factors of the limit of expander decomposition-based approaches.

cs.DS↗

arXiv2025

Faster Negative-Weight Shortest Paths and Directed Low-Diameter Decompositions

We present a faster algorithm for low-diameter decompositions on directed graphs, matching the $O(\log n\log\log n)$ loss factor from Bringmann, Fischer, Haeupler, and Latypov (ICALP 2025) and improving the running time to $O((m+n\log\log n)\log n\log\log n)$ in expectation. We then apply our faster low-diameter decomposition to obtain an algorithm for negative-weight single source shortest paths on integer-weighted graphs in $O((m+n\log\log n)\log(nW)\log n\log\log n)$ time, a nearly log-factor improvement over the algorithm of Bringmann, Cassis, and Fischer (FOCS 2023).

cs.DS↗

arXiv2025

Combinatorial Maximum Flow via Weighted Push-Relabel on Shortcut Graphs

We give a combinatorial algorithm for computing exact maximum flows in directed graphs with $n$ vertices and edge capacities from $\{1,\dots,U\}$ in $\tilde{O}(n^{2}\log U)$ time, which is near-optimal on dense graphs. This shaves an $n^{o(1)}$ factor from the recent result of [Bernstein-Blikstad-Saranurak-Tu FOCS'24] and, more importantly, greatly simplifies their algorithm. We believe that ours is by a significant margin the simplest of all algorithms that go beyond $\tilde{O}(m\sqrt{n})$ time in general graphs. To highlight this relative simplicity, we provide a full implementation of the algorithm in C++. The only randomized component of our work is the cut-matching game. Via existing tools, we show how to derandomize it for vertex-capacitated max flow and obtain a deterministic $\tilde{O}(n^2)$ time algorithm. This marks the first deterministic near-linear time algorithm for this problem (or even for the special case of bipartite matching) in any density regime.

cs.DS↗

arXiv2025

RePainter: Empowering E-commerce Object Removal via Spatial-matting Reinforcement Learning

In web data, product images are central to boosting user engagement and advertising efficacy on e-commerce platforms, yet the intrusive elements such as watermarks and promotional text remain major obstacles to delivering clear and appealing product visuals. Although diffusion-based inpainting methods have advanced, they still face challenges in commercial settings due to unreliable object removal and limited domain-specific adaptation. To tackle these challenges, we propose Repainter, a reinforcement learning framework that integrates spatial-matting trajectory refinement with Group Relative Policy Optimization (GRPO). Our approach modulates attention mechanisms to emphasize background context, generating higher-reward samples and reducing unwanted object insertion. We also introduce a composite reward mechanism that balances global, local, and semantic constraints, effectively reducing visual artifacts and reward hacking. Additionally, we contribute EcomPaint-100K, a high-quality, large-scale e-commerce inpainting dataset, and a standardized benchmark EcomPaint-Bench for fair evaluation. Extensive experiments demonstrate that Repainter significantly outperforms state-of-the-art methods, especially in challenging scenes with intricate compositions. We will release our code and weights upon acceptance.

cs.CV↗

arXiv2025

Align2Speak: Improving TTS for Low Resource Languages via ASR-Guided Online Preference Optimization

Developing high-quality text-to-speech (TTS) systems for low-resource languages is challenging due to the scarcity of paired text and speech data. In contrast, automatic speech recognition (ASR) models for such languages are often more accessible, owing to large-scale multilingual pre-training efforts. We propose a framework based on Group Relative Policy Optimization (GRPO) to adapt an autoregressive, multilingual TTS model to new languages. Our method first establishes a language-agnostic foundation for TTS synthesis by training a multilingual baseline with International Phonetic Alphabet (IPA) tokens. Next, we fine-tune this model on limited paired data of the new languages to capture the target language's prosodic features. Finally, we apply GRPO to optimize the model using only unpaired text and speaker prompts, guided by a multi-objective reward from pretrained ASR, speaker verification, and audio quality estimation models. Experiments demonstrate that this pipeline produces intelligible and speaker-consistent speech in low-resource languages, substantially outperforming fine-tuning alone. Furthermore, our GRPO-based framework also improves TTS performance in high-resource languages, surpassing offline alignment methods such as Direct Preference Optimization (DPO) yielding superior intelligibility, speaker similarity, and audio quality.

cs.AI↗

arXiv2025

HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset

This paper introduces HiFiTTS-2, a large-scale speech dataset designed for high-bandwidth speech synthesis. The dataset is derived from LibriVox audiobooks, and contains approximately 36.7k hours of English speech for 22.05 kHz training, and 31.7k hours for 44.1 kHz training. We present our data processing pipeline, including bandwidth estimation, segmentation, text preprocessing, and multi-speaker detection. The dataset is accompanied by detailed utterance and audiobook metadata generated by our pipeline, enabling researchers to apply data quality filters to adapt the dataset to various use cases. Experimental results demonstrate that our data pipeline and resulting dataset can facilitate the training of high-quality, zero-shot text-to-speech (TTS) models at high bandwidths.

eess.AS↗

arXiv2025

Error Reflection Prompting: Can Large Language Models Successfully Understand Errors?

Prompting methods for language models, such as Chain-of-thought (CoT), present intuitive step-by-step processes for problem solving. These methodologies aim to equip models with a better understanding of the correct procedures for addressing a given task. Despite these advancements, CoT lacks the ability of reflection and error correction, potentially causing a model to perpetuate mistakes and errors. Therefore, inspired by the human ability for said tasks, we propose Error Reflection Prompting (ERP) to further enhance reasoning in language models. Building upon CoT, ERP is a method comprised of an incorrect answer, error recognition, and a correct answer. This process enables the model to recognize types of errors and the steps that lead to incorrect answers, allowing the model to better discern which steps to avoid and which to take. The model is able to generate the error outlines itself with automated ERP generation, allowing for error recognition and correction to be integrated into the reasoning chain and produce scalability and reliability in the process. The results demonstrate that ERP serves as a versatile supplement to conventional CoT, ultimately contributing to more robust and capable reasoning abilities along with increased interpretability in how models ultimately reach their errors.

cs.CL↗

arXiv2025

NanoCodec: Towards High-Quality Ultra Fast Speech LLM Inference

Large Language Models (LLMs) have significantly advanced audio processing by leveraging audio codecs to discretize audio into tokens, enabling the application of language modeling techniques to speech data. However, existing audio codecs often operate at high frame rates, leading to slow training and inference, particularly for autoregressive models. To address this, there is growing interest in low frame-rate audio codecs, which reduce the number of autoregressive steps required to generate one second of audio. In this paper, we conduct ablation studies to examine the impact of frame rate, bitrate, and causality on codec reconstruction quality. Based on our findings, we introduce NanoCodec, a state-of-the-art audio codec that achieves high-quality compression at just 12.5 frames per second (FPS). NanoCodec outperforms related works across various bitrate ranges, establishing a new benchmark for low-latency and efficient Speech LLM training and inference.

eess.AS↗

arXiv2025

Deterministic Almost-Linear-Time Gomory-Hu Trees

Given an $m$-edge, undirected, weighted graph $G=(V,E,w)$, a Gomory-Hu tree $T$ (Gomory and Hu, 1961) is a tree over the vertex set $V$ such that all-pairs mincuts in $G$ are preserved exactly in $T$. In this article, we give the first almost-optimal $m^{1+o(1)}$-time deterministic algorithm for constructing a Gomory-Hu tree. Prior to our work, the best deterministic algorithm for this problem dated back to the original algorithm of Gomory and Hu that runs in $nm^{1+o(1)}$ time (using current maxflow algorithms). In fact, this is the first almost-linear time deterministic algorithm for even simpler problems, such as finding the $k$-edge-connected components of a graph. Our new result hinges on two separate and novel components that each introduce a distinct set of de-randomization tools of independent interest: - a deterministic reduction from the all-pairs mincuts problem to the single-souce mincuts problem incurring only subpolynomial overhead, and - a deterministic almost-linear time algorithm for the single-source mincuts problem.

cs.DS↗

arXiv2025

SALM-Duplex: Efficient and Direct Duplex Modeling for Speech-to-Speech Language Model

Spoken dialogue is an intuitive form of human-computer interaction, yet current speech language models often remain constrained to turn-based exchanges, lacking real-time adaptability such as user barge-in. We propose a novel duplex speech to speech (S2S) architecture featuring continuous user inputs and codec agent outputs with channel fusion that directly models simultaneous user and agent streams. Using a pretrained streaming encoder for user input enables the first duplex S2S model without requiring speech pretrain. Separate architectures for agent and user modeling facilitate codec fine-tuning for better agent voices and halve the bitrate (0.6 kbps) compared to previous works. Experimental results show that the proposed model outperforms previous duplex models in reasoning, turn-taking, and barge-in abilities. The model requires significantly less speech data, as speech pretrain is skipped, which markedly simplifies the process of building a duplex S2S model from any LLMs. Finally, it is the first openly available duplex S2S model with training and inference code to foster reproducibility.

cs.CL↗

arXiv2025

Koel-TTS: Enhancing LLM based Speech Generation with Preference Alignment and Classifier Free Guidance

While autoregressive speech token generation models produce speech with remarkable variety and naturalness, their inherent lack of controllability often results in issues such as hallucinations and undesired vocalizations that do not conform to conditioning inputs. We introduce Koel-TTS, a suite of enhanced encoder-decoder Transformer TTS models that address these challenges by incorporating preference alignment techniques guided by automatic speech recognition and speaker verification models. Additionally, we incorporate classifier-free guidance to further improve synthesis adherence to the transcript and reference speaker audio. Our experiments demonstrate that these optimizations significantly enhance target speaker similarity, intelligibility, and naturalness of synthesized speech. Notably, Koel-TTS directly maps text and context audio to acoustic tokens, and on the aforementioned metrics, outperforms state-of-the-art TTS models, despite being trained on a significantly smaller dataset. Audio samples and demos are available on our website.

cs.SD↗

arXiv2025

Fast Algorithms for Graph Arboricity and Related Problems

We give an algorithm for finding the arboricity of a weighted, undirected graph, defined as the minimum number of spanning forests that cover all edges of the graph, in $\sqrt{n} m^{1+o(1)}$ time. This improves on the previous best bound of $\tilde{O}(nm)$ for weighted graphs and $\tilde{O}(m^{3/2}) $ for unweighted graphs (Gabow 1995) for this problem. The running time of our algorithm is dominated by a logarithmic number of calls to a directed global minimum cut subroutine -- if the running time of the latter problem improves to $m^{1+o(1)}$ (thereby matching the running time of maximum flow), the running time of our arboricity algorithm would improve further to $m^{1+o(1)}$. We also give a new algorithm for computing the entire cut hierarchy -- laminar multiway cuts with minimum cut ratio in recursively defined induced subgraphs -- in $m n^{1+o(1)}$ time. The cut hierarchy yields the ideal edge loads (Thorup 2001) in a fractional spanning tree packing of the graph which, we show, also corresponds to a max-entropy solution in the spanning tree polytope. For the cut hierarchy problem, the previous best bound was $\tilde{O}(n^2 m)$ for weighted graphs and $\tilde{O}(n m^{3/2})$ for unweighted graphs.

cs.DS↗

arXiv2025

Matroid-Based TSP Rounding for Half-Integral Solutions

We show how to round any half-integral solution to the subtour-elimination relaxation for the TSP, while losing a less-than-1.5 factor. Such a rounding algorithm was recently given by Karlin, Klein, and Oveis Gharan based on sampling from max-entropy distributions. We build on an approach of Haddadan and Newman to show how sampling from the matroid intersection polytope, and a new use of max-entropy sampling, can give better guarantees.

cs.DS↗

arXiv2025

Distributed Treewidth Computation and Courcelle's Theorem in the CONGEST Model

Algorithmic meta-theorems, stating that graph properties expressible in some particular logic can be decided efficiently in graph classes having some specific structural properties, are now standard in sequential graph algorithms. One of the most classic examples is Courcelle's theorem: all properties expressible in Monadic Second-Order logic (MSO) are decidable in linear time in graphs of bounded treewidth. We provide here a distributed version of Courcelle's theorem, in the standard CONGEST model for distributed computing: For any MSO formula $φ$ and any constant $k$, there is a CONGEST algorithm that, given an input communication network $G$ of treewidth at most $k$ and of diameter $D$, decides if $G$ satisfies property $φ$ in $\tilde O(D)$ rounds. Simple examples show that the dependency on $D$ is unavoidable. Also, if we drop the assumption of bounded treewidth, deciding MSO properties such as 3-colorability are known to require $\tildeΩ(n^2)$ rounds in the CONGEST model. Our results extend to optimization problems (e.g., computing a maximum size independent set, or a minimum dominating set) and counting (e.g. triangle counting). As usual, the $\tilde{O}$ notation hides polylogarithmic factors in $n$; here it also hides a constant factor depending on $k$ and on the MSO formula $φ$. We also give a distributed algorithm producing a linear approximation for treewidth: For any $k$, it decides that the treewidth of the input network $G$ is larger than $k$ or computes a tree decomposition of width $O(k)$ and depth $O(\log n)$, in $\tilde O(k^{O(k)} D)$ rounds in CONGEST. Our algorithms make use of the low-congestion shortcuts framework introduced by Ghaffari and Haeupler [SODA 2016], and our main technical tool is an $\tilde O(k^4 D)$ algorithm for computing $(s,t)$-vertex separators of size at most $k+1$ in graphs of treewidth at most $k$.

cs.DS↗

arXiv2025

Spectral Codecs: Improving Non-Autoregressive Speech Synthesis with Spectrogram-Based Audio Codecs

Historically, most speech models in machine-learning have used the mel-spectrogram as a speech representation. Recently, discrete audio tokens produced by neural audio codecs have become a popular alternate speech representation for speech synthesis tasks such as text-to-speech (TTS). However, the data distribution produced by such codecs is too complex for some TTS models to predict, typically requiring large autoregressive models to get good quality. Most existing audio codecs use Residual Vector Quantization (RVQ) to compress and reconstruct the time-domain audio signal. We propose a spectral codec which uses Finite Scalar Quantization (FSQ) to compress the mel-spectrogram and reconstruct the time-domain audio signal. A study of objective audio quality metrics and subjective listening tests suggests that our spectral codec has comparable perceptual quality to equivalent audio codecs. We show that FSQ, and the use of spectral speech representations, can both improve the performance of parallel TTS models.

eess.AS↗