arXiv subjects

Yixin Shen

Publications and source records attributed to Yixin Shen.

18 recordsLinked to original sources

arXiv2026

Human Cognition in Machines: A Unified Perspective of World Models

This report of world models distinguishes prior works by the cognitive functions they innovate. Many works claim an almost human-like cognitive capability in their world models. To evaluate these claims requires a proper grounding in first principles from human and machine cognition theory. In moving towards human-like world models we present a conceptual unified framework for world models that fully incorporates all the cognitive functions (i.e., memory, perception, language, reasoning, imagining, motivation, and metacognition) and identify gaps in existing research as a guide for future states of the art. In particular, we find that motivation (especially intrinsic motivation) and metacognition remain drastically under-researched, and we propose concrete directions to address these gaps informed by active inference and global workspace theory. We also introduce epistemic world models, a new category encompassing agent frameworks for scientific discovery that operate over structured knowledge. Our taxonomy, applied to video, embodied, and epistemic world models, suggests research directions where prior taxonomies have not.

cs.RO↗

arXiv2026

Towards Automated Kernel Generation in the Era of LLMs

The performance of modern AI systems is fundamentally constrained by the quality of their underlying GPU kernels, which translate high-level algorithmic semantics into low-level hardware operations. Achieving near-optimal kernels requires expert-level understanding of hardware architectures and programming models, making kernel engineering a critical but notoriously time-consuming and non-scalable process. Recent advances in large language models and LLM-based agents have opened new possibilities for automating kernel generation and optimization. LLMs are well-suited to compress expert-level kernel knowledge that is difficult to formalize, while agentic systems further enable scalable optimization by casting kernel development as an iterative, feedback-driven loop. Rapid progress has been made in this area. However, the field remains fragmented and lacks a systematic perspective for LLM-driven kernel generation. This survey addresses this gap by providing a structured overview of existing approaches, spanning LLM-based approaches and agentic optimization workflows, and systematically organizing the datasets and benchmarks that underpin learning and evaluation in this domain. Moreover, key open challenges and future research directions are further outlined, aiming to establish a comprehensive reference for the next generation of automated kernel optimization. To keep track of this field, we maintain an open-source GitHub repository at https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.

cs.LG↗

arXiv2026

Quantum algorithm for Discrete Gaussian Sampling

Discrete Gaussian Sampling on lattices is a fundamental problem in lattice-based cryptography. It appears both in basic cryptographic primitives such as digital signatures and as an important cryptanalysis building block for solving hard lattice problems. In this paper, we show a quantum algorithm based on the quantum rejection sampling technique whose complexity is asymptotically quadratically faster than its classical counterpart in [Wang & Ling, IEEE Trans. Inf. Theory 2019]. Our sampler outputs a quantum state which can either be measured to get the desired distribution or be used directly as such in other quantum algorithms. By doing so, we derive two versions of quantum dual attacks that improve upon the previous ones in [Pouly & Shen, EUROCRYPT 2024]. The two versions are incomparable, each having distinct advantages (speed vs memory requirement). The second version is particularly interesting as it requires only polynomial classical and quantum memory, excluding the classical memory used in the preprocessing step of the Discrete Gaussian sampler. Our quantum Discrete Gaussian sampler can also be used to speed up the algorithm for solving the Short Integer Solution problem, in any norm, of [Bollauf, Pouly & Shen, ePrint 2026/225].

quant-ph↗

arXiv2026

Open-Source Multimodal Moxin Models with Moxin-VLM and Moxin-VLA

Recently, Large Language Models (LLMs) have undergone a significant transformation, marked by a rapid rise in both their popularity and capabilities. Leading this evolution are proprietary LLMs like GPT-4 and GPT-o1, which have captured widespread attention in the AI community due to their remarkable performance and versatility. Simultaneously, open-source LLMs, such as LLaMA and Mistral, have made great contributions to the ever-increasing popularity of LLMs due to the ease to customize and deploy the models across diverse applications. Moxin 7B is introduced as a fully open-source LLM developed in accordance with the Model Openness Framework, which moves beyond the simple sharing of model weights to embrace complete transparency in training, datasets, and implementation detail, thus fostering a more inclusive and collaborative research environment that can sustain a healthy open-source ecosystem. To further equip Moxin with various capabilities in different tasks, we develop three variants based on Moxin, including Moxin-VLM, Moxin-VLA, and Moxin-Chinese, which target the vision-language, vision-language-action, and Chinese capabilities, respectively. Experiments show that our models achieve superior performance in various evaluations. We adopt open-source framework and open data for the training. We release our models, along with the available data and code to derive these models.

cs.CL↗

arXiv2025

Kron-LoRA: Hybrid Kronecker-LoRA Adapters for Scalable, Sustainable Fine-tuning

Fine-tuning massive pre-trained language models across many tasks demands adapters that are both parameter-efficient and expressive. We introduce \textbf{Kron-LoRA}, a hybrid adapter that combines Kronecker-structured factorization with low-rank LoRA compression-an integration that, to our knowledge, has not been explored in parameter-efficient fine-tuning or in matrix approximation literature. Kron-LoRA achieves up to 4$\times$ fewer parameters than standard LoRA while retaining similar expressivity. Experiments on DistilBERT, Mistral-7B, LLaMA-2-7B, and LLaMA-3-8B across eight benchmarks show that Kron-LoRA matches or exceeds LoRA baselines with modest memory savings and only a 5-8\% speed overhead. In sequential fine-tuning, it also delivers competitive cross-task transfer despite using only one-quarter of the adapter parameters. Kron-LoRA thus offers a scalable, sustainable solution for multi-task adaptation of large language models.

cs.LG↗

arXiv2025

A Tight Quantum Algorithm for Multiple Collision Search

Searching for collisions in random functions is a fundamental computational problem, with many applications in symmetric and asymmetric cryptanalysis. When one searches for a single collision, the known quantum algorithms match the query lower bound. This is not the case for the problem of finding multiple collisions, despite its regular appearance as a sub-component in sieving-type algorithms. At EUROCRYPT 2019, Liu and Zhandry gave a query lower bound $Ω(2^{m/3 + 2k/3})$ for finding $2^k$ collisions in a random function with m-bit output. At EUROCRYPT 2023, Bonnetain et al. gave a quantum algorithm matching this bound for a large range of $m$ and $k$, but not all admissible values. Like many previous collision-finding algorithms, theirs is based on the MNRS quantum walk framework, but it chains the walks by reusing the state after outputting a collision. In this paper, we give a new algorithm that tackles the remaining non-optimal range, closing the problem. Our algorithm is tight (up to a polynomial factor) in queries, and also in time under a quantum RAM assumption. The idea is to extend the chained walk to a regime in which several collisions are returned at each step, and the ``walks'' themselves only perform a single diffusion layer.

quant-ph↗

arXiv2025

Improved Classical and Quantum Algorithms for the Shortest Vector Problem via Bounded Distance Decoding

The most important computational problem on lattices is the Shortest Vector Problem (SVP). In this paper, we present new algorithms that improve the state-of-the-art for provable classical/quantum algorithms for SVP. We present the following results. $\bullet$ A new algorithm for SVP that provides a smooth tradeoff between time complexity and memory requirement. For any positive integer $4\leq q\leq \sqrt{n}$, our algorithm takes $q^{13n+o(n)}$ time and requires $poly(n)\cdot q^{16n/q^2}$ memory. This tradeoff which ranges from enumeration ($q=\sqrt{n}$) to sieving ($q$ constant), is a consequence of a new time-memory tradeoff for Discrete Gaussian sampling above the smoothing parameter. $\bullet$ A quantum algorithm for SVP that runs in time $2^{0.950n+o(n)}$ and requires $2^{0.5n+o(n)}$ classical memory and poly(n) qubits. In Quantum Random Access Memory (QRAM) model this algorithm takes only $2^{0.835n+o(n)}$ time and requires a QRAM of size $2^{0.293n+o(n)}$, poly(n) qubits and $2^{0.5n}$ classical space. This improves over the previously fastest classical (which is also the fastest quantum) algorithm due to [ADRS15] that has a time and space complexity $2^{n+o(n)}$. $\bullet$ A classical algorithm for SVP that runs in time $2^{1.669n+o(n)}$ time and $2^{0.5n+o(n)}$ space. This improves over an algorithm of [CCL18] that has the same space complexity. The time complexity of our classical and quantum algorithms are obtained using a known upper bound on a quantity related to the lattice kissing number which is $2^{0.402n}$. We conjecture that for most lattices this quantity is a $2^{o(n)}$. Assuming that this is the case, our classical algorithm runs in time $2^{1.292n+o(n)}$, our quantum algorithm runs in time $2^{0.750n+o(n)}$ and our quantum algorithm in QRAM model runs in time $2^{0.667n+o(n)}$.

cs.DS↗

arXiv2025

Rattan: An Extensible and Scalable Modular Internet Path Emulator

The rapid growth of Internet paths in heterogeneity, scale, and dynamics has made existing emulators increasingly insufficient in flexibility, scalability, and usability. To address these limitations, we present Rattan, an extensible and scalable software network path emulator for modern Internet conditions. Rattan's core innovation lies in its cell-based architecture: by splitting emulation functions into modular "cells" with well-documented asynchronous interfaces, users are allowed to easily compose different cells by hierarchically linking them and easily construct new cells by using standard cell interfaces. This design enables: (1) scalability, supporting hundreds of concurrent gigabit-level paths on a single machine and cluster-level experiments composed of multiple machines; (2) extensibility, simulating new network conditions by constructing new cells. Rattan empowers developers and researchers to efficiently and confidently evaluate, validate, and diagnose diverse network transport innovations for online services.

cs.NI↗

arXiv2025

7B Fully Open Source Moxin-LLM/VLM -- From Pretraining to GRPO-based Reinforcement Learning Enhancement

Recently, Large Language Models (LLMs) have undergone a significant transformation, marked by a rapid rise in both their popularity and capabilities. Leading this evolution are proprietary LLMs like GPT-4 and GPT-o1, which have captured widespread attention in the AI community due to their remarkable performance and versatility. Simultaneously, open-source LLMs, such as LLaMA, have made great contributions to the ever-increasing popularity of LLMs due to the ease to customize and deploy the models across diverse applications. Although open-source LLMs present unprecedented opportunities for innovation and research, the commercialization of LLMs has raised concerns about transparency, reproducibility, and safety. Many open-source LLMs fail to meet fundamental transparency requirements by withholding essential components like training code and data, which may hinder further innovations on LLMs. To mitigate this issue, we introduce Moxin 7B, a fully open-source LLM developed, adhering to principles of open science, open source, open data, and open access. We release the pre-training code and configurations, training and fine-tuning datasets, and intermediate and final checkpoints, aiming to make continuous commitments to fully open-source LLMs. After pre-training the base model, we finetune the Moxin Base model with SOTA post-training framework and instruction data to obtain Moxin Instruct model. To improve the reasoning capability, we further finetune our Instruct model with chain-of-thought data distilled from DeepSeek R1, and then use Group Relative Policy Optimization (GRPO) following DeepSeek R1 to finetune our model, leading to the Moxin Reasoning model. Moreover, we develop our vision language model based on our Moxin model. Experiments show that our models achieve superior performance in various evaluations such as zero-shot evaluation, few-shot evaluation, and CoT evaluation.

cs.CL↗

arXiv2025

Optimal Sampling for Generalized Linear Model under Measurement Constraint with Surrogate Variables

Measurement-constrained datasets, often encountered in semi-supervised learning, arise when data labeling is costly, time-intensive, or hindered by confidentiality or ethical concerns, resulting in a scarcity of labeled data. In certain cases, surrogate variables are accessible across the entire dataset and can serve as approximations to the true response variable; however, these surrogates often contain measurement errors and thus cannot be directly used for accurate prediction. We propose an optimal sampling strategy that effectively harnesses the available information from surrogate variables. This approach provides consistent estimators under the assumption of a generalized linear model, achieving theoretically lower asymptotic variance than existing optimal sampling algorithms that do not use surrogate data information. By employing the A-optimality criterion from optimal experimental design, our strategy maximizes statistical efficiency. Numerical studies demonstrate that our approach surpasses existing optimal sampling methods, exhibiting reduced empirical mean squared error and enhanced robustness in algorithmic performance. These findings highlight the practical advantages of our strategy in scenarios where measurement constraints exist and surrogates are available.

stat.ME↗

arXiv2024

Does quantum lattice sieving require quantum RAM?

In this paper, we study the requirement for quantum random access memory (QRAM) in quantum lattice sieving, a fundamental algorithm for lattice-based cryptanalysis. First, we obtain a lower bound on the cost of quantum lattice sieving with a bounded size QRAM. We do so in a new query model encompassing a wide range of lattice sieving algorithms similar to those in the classical sieving lower bound by Kirshanova and Laarhoven [CRYPTO 21]. This implies that, under reasonable assumptions, quantum speedups in lattice sieving require the use of QRAM. In particular, no quantum speedup is possible without QRAM. Second, we investigate the trade-off between the size of QRAM and the quantum speedup. We obtain a new interpolation between classical and quantum lattice sieving. Moreover, we show that further improvements require a novel way to use the QRAM by proving the optimality of some subroutines. An important caveat is that this trade-off requires a strong assumption on the efficient replacement of QRAM data, indicating that even speedups with a small QRAM are already challenging. Finally, we provide a circuit for quantum lattice sieving without using QRAM. Our circuit has a better depth complexity than the best classical algorithms but requires an exponential amount of qubits. To the best of our knowledge, this is the first quantum speedup for lattice sieving without QRAM in the standard quantum circuit model. We explain why this circuit does not contradict our lower bound, which considers the query complexity.

quant-ph↗

arXiv2023

Variational quantum solutions to the Shortest Vector Problem

A fundamental computational problem is to find a shortest non-zero vector in Euclidean lattices, a problem known as the Shortest Vector Problem (SVP). This problem is believed to be hard even on quantum computers and thus plays a pivotal role in post-quantum cryptography. In this work we explore how (efficiently) Noisy Intermediate Scale Quantum (NISQ) devices may be used to solve SVP. Specifically, we map the problem to that of finding the ground state of a suitable Hamiltonian. In particular, (i) we establish new bounds for lattice enumeration, this allows us to obtain new bounds (resp.~estimates) for the number of qubits required per dimension for any lattices (resp.~random q-ary lattices) to solve SVP; (ii) we exclude the zero vector from the optimization space by proposing (a) a different classical optimisation loop or alternatively (b) a new mapping to the Hamiltonian. These improvements allow us to solve SVP in dimension up to 28 in a quantum emulation, significantly more than what was previously achieved, even for special cases. Finally, we extrapolate the size of NISQ devices that is required to be able to solve instances of lattices that are hard even for the best classical algorithms and find that with approximately $10^3$ noisy qubits such instances can be tackled.

quant-ph↗

arXiv2023

Quantum Augmented Dual Attack

We present a quantum augmented variant of the dual lattice attack on the Learning with Errors (LWE) problem, using classical memory with quantum random access (QRACM). Applying our results to lattice parameters from the literature, we find that our algorithm outperforms previous algorithms, assuming unit cost access to a QRACM. On a technical level, we show how to obtain a quantum speedup on the search for Fast Fourier Transform (FFT) coefficients above a given threshold by leveraging the relative sparseness of the FFT and using quantum amplitude estimation. We also discuss the applicability of the Quantum Fourier Transform in this context. Furthermore, we give a more rigorous analysis of the classical and quantum expected complexity of guessing part of the secret vector where coefficients follow a discrete Gaussian (mod $q$).

quant-ph↗

arXiv2022

Finding many Collisions via Reusable Quantum Walks

Given a random function $f$ with domain $[2^n]$ and codomain $[2^m]$, with $m \geq n$, a collision of $f$ is a pair of distinct inputs with the same image. Collision finding is an ubiquitous problem in cryptanalysis, and it has been well studied using both classical and quantum algorithms. Indeed, the quantum query complexity of the problem is well known to be $Θ(2^{m/3})$, and matching algorithms are known for any value of $m$. The situation becomes different when one is looking for multiple collision pairs. Here, for $2^k$ collisions, a query lower bound of $Θ(2^{(2k+m)/3})$ was shown by Liu and Zhandry (EUROCRYPT~2019). A matching algorithm is known, but only for relatively small values of $m$, when many collisions exist. In this paper, we improve the algorithms for this problem and, in particular, extend the range of admissible parameters where the lower bound is met. Our new method relies on a chained quantum walk algorithm, which might be of independent interest. It allows to extract multiple solutions of an MNRS-style quantum walk, without having to recompute it entirely: after finding and outputting a solution, the current state is reused as the initial state of another walk. As an application, we improve the quantum sieving algorithms for the shortest vector problem (SVP), with a complexity of $2^{0.2563d + o(d)}$ instead of the previous $2^{0.2570d + o(d)}$.

quant-ph↗

arXiv2021

Fast Classical and Quantum Algorithms for Online $k$-server Problem on Trees

We consider online algorithms for the $k$-server problem on trees. Chrobak and Larmore proposed a $k$-competitive algorithm for this problem that has the optimal competitive ratio. However, a naive implementation of their algorithm has $O(n)$ time complexity for processing each query, where $n$ is the number of nodes in the tree. We propose a new time-efficient implementation of this algorithm that has $O(n\log n)$ time complexity for preprocessing and $O\left(k^2 + k\cdot \log n\right)$ time for processing a query. We also propose a quantum algorithm for the case where the nodes of the tree are presented using string paths. In this case, no preprocessing is needed, and the time complexity for each query is $O(k^2\sqrt{n}\log n)$. When the number of queries is $o\left(\frac{\sqrt{n}}{k^2\log n}\right)$, we obtain a quantum speed-up on the total runtime compared to our classical algorithm. We also give a simple quantum algorithm to find the first marked element in a collection of $m$ objects, that works even in the presence of two-sided bounded errors on the input oracle. It has worst-case complexity $O(\sqrt{m})$. In the particular case of one-sided errors on the input, it has expected time complexity $O(\sqrt{x})$ where $x$ is the position of the first marked element. Compare with previous work, our algorithm can handle errors in the input oracle.

cs.DS↗

arXiv2020

Improved Classical and Quantum Algorithms for Subset-Sum

We present new classical and quantum algorithms for solving random subset-sum instances. First, we improve over the Becker-Coron-Joux algorithm (EUROCRYPT 2011) from $\tilde{\mathcal{O}}(2^{0.291 n})$ downto $\tilde{\mathcal{O}}(2^{0.283 n})$, using more general representations with values in $\{-1,0,1,2\}$. Next, we improve the state of the art of quantum algorithms for this problem in several directions. By combining the Howgrave-Graham-Joux algorithm (EUROCRYPT 2010) and quantum search, we devise an algorithm with asymptotic cost $\tilde{\mathcal{O}}(2^{0.236 n})$, lower than the cost of the quantum walk based on the same classical algorithm proposed by Bernstein, Jeffery, Lange and Meurer (PQCRYPTO 2013). This algorithm has the advantage of using \emph{classical} memory with quantum random access, while the previously known algorithms used the quantum walk framework, and required \emph{quantum} memory with quantum random access. We also propose new quantum walks for subset-sum, performing better than the previous best time complexity of $\tilde{\mathcal{O}}(2^{0.226 n})$ given by Helm and May (TQC 2018). We combine our new techniques to reach a time $\tilde{\mathcal{O}}(2^{0.216 n})$. This time is dependent on a heuristic on quantum walk updates, formalized by Helm and May, that is also required by the previous algorithms. We show how to partially overcome this heuristic, and we obtain an algorithm with quantum time $\tilde{\mathcal{O}}(2^{0.218 n})$ requiring only the standard classical subset-sum heuristics.

quant-ph↗

arXiv2020

Quantum Lower and Upper Bounds for 2D-Grid and Dyck Language

We study the quantum query complexity of two problems. First, we consider the problem of determining if a sequence of parentheses is a properly balanced one (a Dyck word), with a depth of at most $k$. We call this the $Dyck_{k,n}$ problem. We prove a lower bound of $Ω(c^k \sqrt{n})$, showing that the complexity of this problem increases exponentially in $k$. Here $n$ is the length of the word. When $k$ is a constant, this is interesting as a representative example of star-free languages for which a surprising $\tilde{O}(\sqrt{n})$ query quantum algorithm was recently constructed by Aaronson et al. Their proof does not give rise to a general algorithm. When $k$ is not a constant, $Dyck_{k,n}$ is not context-free. We give an algorithm with $O\left(\sqrt{n}(\log{n})^{0.5k}\right)$ quantum queries for $Dyck_{k,n}$ for all $k$. This is better than the trival upper bound $n$ for $k=o\left(\frac{\log(n)}{\log\log n}\right)$. Second, we consider connectivity problems on grid graphs in 2 dimensions, if some of the edges of the grid may be missing. By embedding the "balanced parentheses" problem into the grid, we show a lower bound of $Ω(n^{1.5-ε})$ for the directed 2D grid and $Ω(n^{2-ε})$ for the undirected 2D grid. The directed problem is interesting as a black-box model for a class of classical dynamic programming strategies including the one that is usually used for the well-known edit distance problem. We also show a generalization of this result to more than 2 dimensions.

cs.DS↗

arXiv2020

Quantum Query Complexity of Dyck Languages with Bounded Height

We consider the problem of determining if a sequence of parentheses is well parenthesized, with a depth of at most h. We denote this language as $Dyck_h$. We study the quantum query complexity of this problem for different h as function of the length n of the word. It has been known from a recent paper by Aaronson et al. that, for any constant h, since $Dyck_h$ is star-free, it has quantum query complexity $\tildeΘ(\sqrt{n})$, where the hidden logarithm factors in $\tildeΘ$ depend on h. Their proof does not give rise to an algorithm. When h is not a constant, $Dyck_h$ is not even context-free. We give an algorithm with $O\left(\sqrt{n}\log(n)^{0.5h}\right)$ quantum queries for $Dyck_h$ for all h. This is better than the trival upper bound $n$ when $h=o(\frac{\log(n)}{\log\log n})$. We also obtain lower bounds: we show that for every $0<ε\leq 0.37$, there exists $c>0$ such that $Q(\text{Dyck}_{c\log(n)}(n))=Ω(n^{1-ε})$. When $h=ω(\log(n))$, the quantum query complexity is close to $n$, i.e. $Q(\text{Dyck}_h(n))=ω(n^{1-ε})$ for all $ε>0$. Furthermore when $h=Ω(n^ε)$ for some $ε>0$, $Q(\text{Dyck}_{h}(n))=Θ(n)$.

quant-ph↗