arXiv subjects

Zhuoran Yang

Publications and source records attributed to Zhuoran Yang.

At least 55 records · Page 3Linked to original sources

arXiv2025

Sample-Efficient Reinforcement Learning from Human Feedback via Information-Directed Sampling

We study the problem of reinforcement learning from human feedback (RLHF), a critical problem in training large language models, from a theoretical perspective. Our main contribution is the design of novel sample-efficient RLHF algorithms based on information-directed sampling (IDS), an online decision-making principle inspired by information theory. Our algorithms maximize the sum of the value function and a mutual information term that encourages exploration of the unknown environment (which quantifies the information gained about the environment through observed human feedback data). To tackle the challenge of large state spaces and improve sample efficiency, we construct a simplified \emph{surrogate environment} and introduce a novel distance measure (named the \emph{$\ell_g$-distance}), enabling our IDS-based algorithm to achieve a Bayesian regret upper bound of order $O(H^{\frac{3}{2}}\sqrt{\log(K(ε)) T})$, where $H$ is the episode length, $T$ is the number of episode and $K(ε)$ is related to the covering number of the environment. Specializing to the tabular settings, this regret bound is of order $\tilde{O}(H^2\sqrt{SAT})$, where $S$ and $A$ are the numbers of states and actions. Finally, we propose an Approximate-IDS algorithm that is computationally more efficient while maintaining nearly the same sample efficiency. The design principle of this approximate algorithm is not only effective in RLHF settings but also applicable to the standard RL framework. Moreover, our work showcases the value of information theory in reinforcement learning and in the training of large language models.

cs.LG↗

arXiv2025

Enhancing Long Video Generation Consistency without Tuning

Despite the considerable progress achieved in the long video generation problem, there is still significant room to improve the consistency of the generated videos, particularly in terms of their smoothness and transitions between scenes. We address these issues to enhance the consistency and coherence of videos generated with either single or multiple prompts. We propose the Time-frequency based temporal Attention Reweighting Algorithm (TiARA), which judiciously edits the attention score matrix based on the Discrete Short-Time Fourier Transform. This method is supported by a frequency-based analysis, ensuring that the edited attention score matrix achieves improved consistency across frames. It represents the first-of-its-kind for frequency-based methods in video diffusion models. For videos generated by multiple prompts, we further uncover key factors such as the alignment of the prompts affecting prompt interpolation quality. Inspired by our analyses, we propose PromptBlend, an advanced prompt interpolation pipeline that systematically aligns the prompts. Extensive experimental results validate the efficacy of our proposed method, demonstrating consistent and substantial improvements over multiple baselines.

cs.CV↗

arXiv2025

Kwai Keye-VL Technical Report

While Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities on static images, they often fall short in comprehending dynamic, information-dense short-form videos, a dominant medium in today's digital landscape. To bridge this gap, we introduce \textbf{Kwai Keye-VL}, an 8-billion-parameter multimodal foundation model engineered for leading-edge performance in short-video understanding while maintaining robust general-purpose vision-language abilities. The development of Keye-VL rests on two core pillars: a massive, high-quality dataset exceeding 600 billion tokens with a strong emphasis on video, and an innovative training recipe. This recipe features a four-stage pre-training process for solid vision-language alignment, followed by a meticulous two-phase post-training process. The first post-training stage enhances foundational capabilities like instruction following, while the second phase focuses on stimulating advanced reasoning. In this second phase, a key innovation is our five-mode ``cold-start'' data mixture, which includes ``thinking'', ``non-thinking'', ``auto-think'', ``think with image'', and high-quality video data. This mixture teaches the model to decide when and how to reason. Subsequent reinforcement learning (RL) and alignment steps further enhance these reasoning capabilities and correct abnormal model behaviors, such as repetitive outputs. To validate our approach, we conduct extensive evaluations, showing that Keye-VL achieves state-of-the-art results on public video benchmarks and remains highly competitive on general image-based tasks (Figure 1). Furthermore, we develop and release the \textbf{KC-MMBench}, a new benchmark tailored for real-world short-video scenarios, where Keye-VL shows a significant advantage.

cs.CV↗

arXiv2025

Taming Polysemanticity in LLMs: Provable Feature Recovery via Sparse Autoencoders

We study the challenge of achieving theoretically grounded feature recovery using Sparse Autoencoders (SAEs) for the interpretation of Large Language Models. Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations such as hyperparameter sensitivity and instability. To address these issues, we first propose a novel statistical framework for the feature recovery problem, which includes a new notion of feature identifiability by modeling polysemantic features as sparse mixtures of underlying monosemantic concepts. Building on this framework, we introduce a new SAE training algorithm based on ``bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity. We theoretically \highlight{prove that this algorithm correctly recovers all monosemantic features} when input data is sampled from our proposed statistical model. Furthermore, we develop an improved empirical variant, Group Bias Adaptation (GBA), and \highlight{demonstrate its superior performance against benchmark methods when applied to LLMs with up to 1.5 billion parameters}. This work represents a foundational step in demystifying SAE training by providing the first SAE algorithm with theoretical recovery guarantees, thereby advancing the development of more transparent and trustworthy AI systems through enhanced mechanistic interpretability.

cs.LG↗

arXiv2025

The Sample Complexity of Online Strategic Decision Making with Information Asymmetry and Knowledge Transportability

Information asymmetry is a pervasive feature of multi-agent systems, especially evident in economics and social sciences. In these settings, agents tailor their actions based on private information to maximize their rewards. These strategic behaviors often introduce complexities due to confounding variables. Simultaneously, knowledge transportability poses another significant challenge, arising from the difficulties of conducting experiments in target environments. It requires transferring knowledge from environments where empirical data is more readily available. Against these backdrops, this paper explores a fundamental question in online learning: Can we employ non-i.i.d. actions to learn about confounders even when requiring knowledge transfer? We present a sample-efficient algorithm designed to accurately identify system dynamics under information asymmetry and to navigate the challenges of knowledge transfer effectively in reinforcement learning, framed within an online strategic interaction model. Our method provably achieves learning of an $ε$-optimal policy with a tight sample complexity of $O(1/ε^2)$.

cs.LG↗

arXiv2025

Learning to Lead: Incentivizing Strategic Agents in the Dark

We study an online learning version of the generalized principal-agent model, where a principal interacts repeatedly with a strategic agent possessing private types, private rewards, and taking unobservable actions. The agent is non-myopic, optimizing a discounted sum of future rewards and may strategically misreport types to manipulate the principal's learning. The principal, observing only her own realized rewards and the agent's reported types, aims to learn an optimal coordination mechanism that minimizes strategic regret. We develop the first provably sample-efficient algorithm for this challenging setting. Our approach features a novel pipeline that combines (i) a delaying mechanism to incentivize approximately myopic agent behavior, (ii) an innovative reward angle estimation framework that uses sector tests and a matching procedure to recover type-dependent reward functions, and (iii) a pessimistic-optimistic LinUCB algorithm that enables the principal to explore efficiently while respecting the agent's incentive constraints. We establish a near optimal $\tilde{O}(\sqrt{T}) $ regret bound for learning the principal's optimal policy, where $\tilde{O}(\cdot) $ omits logarithmic factors. Our results open up new avenues for designing robust online learning algorithms for a wide range of game-theoretic settings involving private types and strategic agents.

cs.LG↗

arXiv2025

Quantile-Optimal Policy Learning under Unmeasured Confounding

We study quantile-optimal policy learning where the goal is to find a policy whose reward distribution has the largest $α$-quantile for some $α\in (0, 1)$. We focus on the offline setting whose generating process involves unobserved confounders. Such a problem suffers from three main challenges: (i) nonlinearity of the quantile objective as a functional of the reward distribution, (ii) unobserved confounding issue, and (iii) insufficient coverage of the offline dataset. To address these challenges, we propose a suite of causal-assisted policy learning methods that provably enjoy strong theoretical guarantees under mild conditions. In particular, to address (i) and (ii), using causal inference tools such as instrumental variables and negative controls, we propose to estimate the quantile objectives by solving nonlinear functional integral equations. Then we adopt a minimax estimation approach with nonparametric models to solve these integral equations, and propose to construct conservative policy estimates that address (iii). The final policy is the one that maximizes these pessimistic estimates. In addition, we propose a novel regularized policy learning method that is more amenable to computation. Finally, we prove that the policies learned by these methods are $\tilde{\mathscr{O}}(n^{-1/2})$ quantile-optimal under a mild coverage assumption on the offline dataset. Here, $\tilde{\mathscr{O}}(\cdot)$ omits poly-logarithmic factors. To the best of our knowledge, we propose the first sample-efficient policy learning algorithms for estimating the quantile-optimal policy when there exist unmeasured confounding.

stat.ML↗

arXiv2025

Policy learning "without" overlap: Pessimism and generalized empirical Bernstein's inequality

This paper studies offline policy learning, which aims at utilizing observations collected a priori (from either fixed or adaptively evolving behavior policies) to learn an optimal individualized decision rule that achieves the best overall outcomes for a given population. Existing policy learning methods rely on a uniform overlap assumption, i.e., the propensities of exploring all actions for all individual characteristics must be lower bounded. As one has no control over the data collection process, this assumption can be unrealistic in many situations, especially when the behavior policies are allowed to evolve over time with diminishing propensities for certain actions. In this paper, we propose Pessimistic Policy Learning (PPL), a new algorithm that optimizes lower confidence bounds (LCBs) -- instead of point estimates -- of the policy values. The LCBs are constructed using knowledge of the behavior policies for collecting the offline data. Without assuming any uniform overlap condition, we establish a data-dependent upper bound for the suboptimality of our algorithm, which only depends on (i) the overlap for the optimal policy, and (ii) the complexity of the policy class we optimize over. As an implication, for adaptively collected data, we ensure efficient policy learning as long as the propensities for optimal actions are lower bounded over time, while those for suboptimal ones are allowed to diminish arbitrarily fast. In our theoretical analysis, we develop a new self-normalized type concentration inequality for inverse-propensity-weighting estimators, generalizing the well-known empirical Bernstein's inequality to unbounded and non-i.i.d. data. We complement our theory with an efficient optimization algorithm via Majorization-Minimization and policy tree search, as well as extensive simulation studies and real-world applications that demonstrate the efficacy of PPL.

cs.LG↗

arXiv2025

In-Context Linear Regression Demystified: Training Dynamics and Mechanistic Interpretability of Multi-Head Softmax Attention

We study how multi-head softmax attention models are trained to perform in-context learning on linear data. Through extensive empirical experiments and rigorous theoretical analysis, we demystify the emergence of elegant attention patterns: a diagonal and homogeneous pattern in the key-query (KQ) weights, and a last-entry-only and zero-sum pattern in the output-value (OV) weights. Remarkably, these patterns consistently appear from gradient-based training starting from random initialization. Our analysis reveals that such emergent structures enable multi-head attention to approximately implement a debiased gradient descent predictor -- one that outperforms single-head attention and nearly achieves Bayesian optimality up to proportional factor. Furthermore, compared to linear transformers, the softmax attention readily generalizes to sequences longer than those seen during training. We also extend our study to scenarios with anisotropic covariates and multi-task linear regression. In the former, multi-head attention learns to implement a form of pre-conditioned gradient descent. In the latter, we uncover an intriguing regime where the interplay between head number and task number triggers a superposition phenomenon that efficiently resolves multi-task in-context learning. Our results reveal that in-context learning ability emerges from the trained transformer as an aggregated effect of its architecture and the underlying data distribution, paving the way for deeper understanding and broader applications of in-context learning.

cs.LG↗

arXiv2025

Self-Supervised Pre-training with Combined Datasets for 3D Perception in Autonomous Driving

The significant achievements of pre-trained models leveraging large volumes of data in the field of NLP and 2D vision inspire us to explore the potential of extensive data pre-training for 3D perception in autonomous driving. Toward this goal, this paper proposes to utilize massive unlabeled data from heterogeneous datasets to pre-train 3D perception models. We introduce a self-supervised pre-training framework that learns effective 3D representations from scratch on unlabeled data, combined with a prompt adapter based domain adaptation strategy to reduce dataset bias. The approach significantly improves model performance on downstream tasks such as 3D object detection, BEV segmentation, 3D object tracking, and occupancy prediction, and shows steady performance increase as the training data volume scales up, demonstrating the potential of continually benefit 3D perception models for autonomous driving. We will release the source code to inspire further investigations in the community.

cs.CV↗

arXiv2025

Reflective Planning: Vision-Language Models for Multi-Stage Long-Horizon Robotic Manipulation

Solving complex long-horizon robotic manipulation problems requires sophisticated high-level planning capabilities, the ability to reason about the physical world, and reactively choose appropriate motor skills. Vision-language models (VLMs) pretrained on Internet data could in principle offer a framework for tackling such problems. However, in their current form, VLMs lack both the nuanced understanding of intricate physics required for robotic manipulation and the ability to reason over long horizons to address error compounding issues. In this paper, we introduce a novel test-time computation framework that enhances VLMs' physical reasoning capabilities for multi-stage manipulation tasks. At its core, our approach iteratively improves a pretrained VLM with a "reflection" mechanism - it uses a generative model to imagine future world states, leverages these predictions to guide action selection, and critically reflects on potential suboptimalities to refine its reasoning. Experimental results demonstrate that our method significantly outperforms several state-of-the-art commercial VLMs as well as other post-training approaches such as Monte Carlo Tree Search (MCTS). Videos are available at https://reflect-vlm.github.io.

cs.RO↗

arXiv2025

DrugImproverGPT: A Large Language Model for Drug Optimization with Fine-Tuning via Structured Policy Optimization

Finetuning a Large Language Model (LLM) is crucial for generating results towards specific objectives. This research delves into the realm of drug optimization and introduce a novel reinforcement learning algorithm to finetune a drug optimization LLM-based generative model, enhancing the original drug across target objectives, while retains the beneficial chemical properties of the original drug. This work is comprised of two primary components: (1) DrugImprover: A framework tailored for improving robustness and efficiency in drug optimization. It includes a LLM designed for drug optimization and a novel Structured Policy Optimization (SPO) algorithm, which is theoretically grounded. This algorithm offers a unique perspective for fine-tuning the LLM-based generative model by aligning the improvement of the generated molecule with the input molecule under desired objectives. (2) A dataset of 1 million compounds, each with OEDOCK docking scores on 5 human proteins associated with cancer cells and 24 binding sites from SARS-CoV-2 virus. We conduct a comprehensive evaluation of SPO and demonstrate its effectiveness in improving the original drug across target properties. Our code and dataset will be publicly available at: https://github.com/xuefeng-cs/DrugImproverGPT.

cs.LG↗

arXiv2024

An Instrumental Value for Data Production and its Application to Data Pricing

How much value does a dataset or a data production process have to an agent who wishes to use the data to assist decision-making? This is a fundamental question towards understanding the value of data as well as further pricing of data. This paper develops an approach for capturing the instrumental value of data production processes, which takes two key factors into account: (a) the context of the agent's decision-making problem; (b) prior data or information the agent already possesses. We ''micro-found'' our valuation concepts by showing how they connect to classic notions of information design and signals in information economics. When instantiated in the domain of Bayesian linear regression, our value naturally corresponds to information gain. Based on our designed data value, we then study a basic monopoly pricing setting with a buyer looking to purchase from a seller some labeled data of a certain feature direction in order to improve a Bayesian regression model. We show that when the seller has the ability to fully customize any data request, she can extract the first-best revenue (i.e., full surplus) from any population of buyers, i.e., achieving first-degree price discrimination. If the seller can only sell data that are derived from an existing data pool, this limits her ability to customize, and achieving first-best revenue becomes generally impossible. However, we design a mechanism that achieves seller revenue at most $\log (κ)$ less than the first-best revenue, where $κ$ is the condition number associated with the data matrix. A corollary of this result is that the seller can extract the first-best revenue in the multi-armed bandits special case.

cs.GT↗

arXiv2024

Optimistic Policy Optimization is Provably Efficient in Non-stationary MDPs

We study episodic reinforcement learning (RL) in non-stationary linear kernel Markov decision processes (MDPs). In this setting, both the reward function and the transition kernel are linear with respect to the given feature maps and are allowed to vary over time, as long as their respective parameter variations do not exceed certain variation budgets. We propose the \underline{p}eriodically \underline{r}estarted \underline{o}ptimistic \underline{p}olicy \underline{o}ptimization algorithm (PROPO), which is an optimistic policy optimization algorithm with linear function approximation. PROPO features two mechanisms: sliding-window-based policy evaluation and periodic-restart-based policy improvement, which are tailored for policy optimization in a non-stationary environment. In addition, only utilizing the technique of sliding window, we propose a value-iteration algorithm. We establish dynamic upper bounds for the proposed methods and a minimax lower bound which shows the (near-) optimality of the proposed methods. To our best knowledge, PROPO is the first provably efficient policy optimization algorithm that handles non-stationarity.

cs.LG↗

arXiv2024

Learning Dynamic Mechanisms in Unknown Environments: A Reinforcement Learning Approach

Dynamic mechanism design studies how mechanism designers should allocate resources among agents in a time-varying environment. We consider the problem where the agents interact with the mechanism designer according to an unknown Markov Decision Process (MDP), where agent rewards and the mechanism designer's state evolve according to an episodic MDP with unknown reward functions and transition kernels. We focus on the online setting with linear function approximation and propose novel learning algorithms to recover the dynamic Vickrey-Clarke-Grove (VCG) mechanism over multiple rounds of interaction. A key contribution of our approach is incorporating reward-free online Reinforcement Learning (RL) to aid exploration over a rich policy space to estimate prices in the dynamic VCG mechanism. We show that the regret of our proposed method is upper bounded by $\tilde{\mathcal{O}}(T^{2/3})$ and further devise a lower bound to show that our algorithm is efficient, incurring the same $Ω(T^{2 / 3})$ regret as the lower bound, where $T$ is the total number of rounds. Our work establishes the regret guarantee for online RL in solving dynamic mechanism design problems without prior knowledge of the underlying model.

cs.LG↗

arXiv2024

Physical Informed Driving World Model

Autonomous driving requires robust perception models trained on high-quality, large-scale multi-view driving videos for tasks like 3D object detection, segmentation and trajectory prediction. While world models provide a cost-effective solution for generating realistic driving videos, challenges remain in ensuring these videos adhere to fundamental physical principles, such as relative and absolute motion, spatial relationship like occlusion and spatial consistency, and temporal consistency. To address these, we propose DrivePhysica, an innovative model designed to generate realistic multi-view driving videos that accurately adhere to essential physical principles through three key advancements: (1) a Coordinate System Aligner module that integrates relative and absolute motion features to enhance motion interpretation, (2) an Instance Flow Guidance module that ensures precise temporal consistency via efficient 3D flow extraction, and (3) a Box Coordinate Guidance module that improves spatial relationship understanding and accurately resolves occlusion hierarchies. Grounded in physical principles, we achieve state-of-the-art performance in driving video generation quality (3.96 FID and 38.06 FVD on the Nuscenes dataset) and downstream perception tasks. Our project homepage: https://metadrivescape.github.io/papers_project/DrivePhysica/page.html

cs.CV↗

arXiv2024

AV-Odyssey Bench: Can Your Multimodal LLMs Really Understand Audio-Visual Information?

Recently, multimodal large language models (MLLMs), such as GPT-4o, Gemini 1.5 Pro, and Reka Core, have expanded their capabilities to include vision and audio modalities. While these models demonstrate impressive performance across a wide range of audio-visual applications, our proposed DeafTest reveals that MLLMs often struggle with simple tasks humans find trivial: 1) determining which of two sounds is louder, and 2) determining which of two sounds has a higher pitch. Motivated by these observations, we introduce AV-Odyssey Bench, a comprehensive audio-visual benchmark designed to assess whether those MLLMs can truly understand the audio-visual information. This benchmark encompasses 4,555 carefully crafted problems, each incorporating text, visual, and audio components. To successfully infer answers, models must effectively leverage clues from both visual and audio inputs. To ensure precise and objective evaluation of MLLM responses, we have structured the questions as multiple-choice, eliminating the need for human evaluation or LLM-assisted assessment. We benchmark a series of closed-source and open-source models and summarize the observations. By revealing the limitations of current models, we aim to provide useful insight for future dataset collection and model development.

cs.CV↗

arXiv2024

A Mean-Field Analysis of Neural Stochastic Gradient Descent-Ascent for Functional Minimax Optimization

This paper studies minimax optimization problems defined over infinite-dimensional function classes of overparameterized two-layer neural networks. In particular, we consider the minimax optimization problem stemming from estimating linear functional equations defined by conditional expectations, where the objective functions are quadratic in the functional spaces. We address (i) the convergence of the stochastic gradient descent-ascent algorithm and (ii) the representation learning of the neural networks. We establish convergence under the mean-field regime by considering the continuous-time and infinite-width limit of the optimization dynamics. Under this regime, the stochastic gradient descent-ascent corresponds to a Wasserstein gradient flow over the space of probability measures defined over the space of neural network parameters. We prove that the Wasserstein gradient flow converges globally to a stationary point of the minimax objective at a $O(T^{-1} + α^{-1})$ sublinear rate, and additionally finds the solution to the functional equation when the regularizer of the minimax objective is strongly convex. Here $T$ denotes the time and $α$ is a scaling parameter of the neural networks. In terms of representation learning, our results show that the feature representation induced by the neural networks is allowed to deviate from the initial one by the magnitude of $O(α^{-1})$, measured in terms of the Wasserstein distance. Finally, we apply our general results to concrete examples including policy evaluation, nonparametric instrumental variable regression, asset pricing, and adversarial Riesz representer estimation.

cs.LG↗