arXiv subjects

Qiaosheng Zhang

Publications and source records attributed to Qiaosheng Zhang.

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

Stop Overvaluing Multi-Agent Debate -- We Must Rethink Evaluation and Embrace Model Heterogeneity

Multi-agent debate (MAD) has gained significant attention as a promising line of research to improve the factual accuracy and reasoning capabilities of large language models (LLMs). Despite its conceptual appeal, current MAD research suffers from critical limitations in evaluation practices, including limited benchmark coverage, weak baseline comparisons, and inconsistent setups. This paper presents a systematic evaluation of 5 representative MAD methods across 9 benchmarks using 4 foundational models. Surprisingly, our findings reveal that MAD often fail to outperform simple single-agent baselines such as Chain-of-Thought and Self-Consistency, even when consuming significantly more inference-time computation. To advance MAD research, we further explore the role of model heterogeneity and find it as a universal antidote to consistently improve current MAD frameworks. Based on our findings, we argue that the field must stop overvaluing MAD in its current form; for true advancement, we must critically rethink evaluation paradigms and actively embrace model heterogeneity as a core design principle.

cs.CL↗

arXiv2025

The Avengers: A Simple Recipe for Uniting Smaller Language Models to Challenge Proprietary Giants

Proprietary giants are increasingly dominating the race for ever-larger language models. Can open-source, smaller models remain competitive across a broad range of tasks? In this paper, we present the Avengers -- a simple recipe that leverages the collective intelligence of these smaller models. The Avengers builds upon four lightweight operations: (i) embedding: encode queries using a text embedding model; (ii) clustering: group queries based on their semantic similarity; (iii) scoring: scores each model's performance within each cluster; and (iv) voting: improve outputs via repeated sampling and voting. At inference time, each query is embedded and assigned to its nearest cluster. The top-performing model(s) within that cluster are selected to generate the response with repeated sampling. Remarkably, with 10 open-source models (~7B parameters each), the Avengers surpasses GPT-4o, 4.1, and 4.5 in average performance across 15 diverse datasets spanning mathematics, coding, logical reasoning, general knowledge, and affective tasks. In particular, it surpasses GPT-4.1 on mathematics tasks by 18.21% and on code tasks by 7.46%. Furthermore, the Avengers delivers superior out-of-distribution generalization, and remains robust across various embedding models, clustering algorithms, ensemble strategies, and values of its sole parameter -- the number of clusters.

cs.CL↗

arXiv2025

Unsupervised Skill Discovery through Skill Regions Differentiation

Unsupervised Reinforcement Learning (RL) aims to discover diverse behaviors that can accelerate the learning of downstream tasks. Previous methods typically focus on entropy-based exploration or empowerment-driven skill learning. However, entropy-based exploration struggles in large-scale state spaces (e.g., images), and empowerment-based methods with Mutual Information (MI) estimations have limitations in state exploration. To address these challenges, we propose a novel skill discovery objective that maximizes the deviation of the state density of one skill from the explored regions of other skills, encouraging inter-skill state diversity similar to the initial MI objective. For state-density estimation, we construct a novel conditional autoencoder with soft modularization for different skill policies in high-dimensional space. Meanwhile, to incentivize intra-skill exploration, we formulate an intrinsic reward based on the learned autoencoder that resembles count-based exploration in a compact latent space. Through extensive experiments in challenging state and image-based tasks, we find our method learns meaningful skills and achieves superior performance in various downstream tasks.

cs.LG↗

arXiv2025

MM-PRM: Enhancing Multimodal Mathematical Reasoning with Scalable Step-Level Supervision

While Multimodal Large Language Models (MLLMs) have achieved impressive progress in vision-language understanding, they still struggle with complex multi-step reasoning, often producing logically inconsistent or partially correct solutions. A key limitation lies in the lack of fine-grained supervision over intermediate reasoning steps. To address this, we propose MM-PRM, a process reward model trained within a fully automated, scalable framework. We first build MM-Policy, a strong multimodal model trained on diverse mathematical reasoning data. Then, we construct MM-K12, a curated dataset of 10,000 multimodal math problems with verifiable answers, which serves as seed data. Leveraging a Monte Carlo Tree Search (MCTS)-based pipeline, we generate over 700k step-level annotations without human labeling. The resulting PRM is used to score candidate reasoning paths in the Best-of-N inference setup and achieves significant improvements across both in-domain (MM-K12 test set) and out-of-domain (OlympiadBench, MathVista, etc.) benchmarks. Further analysis confirms the effectiveness of soft labels, smaller learning rates, and path diversity in optimizing PRM performance. MM-PRM demonstrates that process supervision is a powerful tool for enhancing the logical robustness of multimodal reasoning systems. We release all our codes and data at https://github.com/ModalMinds/MM-PRM.

cs.AI↗

arXiv2025

CPGD: Toward Stable Rule-based Reinforcement Learning for Language Models

Recent advances in rule-based reinforcement learning (RL) have significantly improved the reasoning capability of language models (LMs) with rule-based rewards. However, existing RL methods -- such as GRPO, REINFORCE++, and RLOO -- often suffer from training instability, where large policy updates and improper clipping can lead to training collapse. To address this issue, we propose Clipped Policy Gradient Optimization with Policy Drift (CPGD), a novel algorithm designed to stabilize policy learning in LMs. CPGD introduces a policy drift constraint based on KL divergence to dynamically regularize policy updates, and leverages a clip mechanism on the logarithm of the ratio to prevent excessive policy updates. We provide theoretical justification for CPGD and demonstrate through empirical analysis that it mitigates the instability observed in prior approaches. Furthermore, we show that CPGD significantly improves performance while maintaining training stability. Our implementation balances theoretical rigor with practical usability, offering a robust alternative for RL in the post-training of LMs. We release our code at https://github.com/ModalMinds/MM-EUREKA.

cs.LG↗

arXiv2025

Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models

Despite the growing popularity of graph attention mechanisms, their theoretical understanding remains limited. This paper aims to explore the conditions under which these mechanisms are effective in node classification tasks through the lens of Contextual Stochastic Block Models (CSBMs). Our theoretical analysis reveals that incorporating graph attention mechanisms is \emph{not universally beneficial}. Specifically, by appropriately defining \emph{structure noise} and \emph{feature noise} in graphs, we show that graph attention mechanisms can enhance classification performance when structure noise exceeds feature noise. Conversely, when feature noise predominates, simpler graph convolution operations are more effective. Furthermore, we examine the over-smoothing phenomenon and show that, in the high signal-to-noise ratio (SNR) regime, graph convolutional networks suffer from over-smoothing, whereas graph attention mechanisms can effectively resolve this issue. Building on these insights, we propose a novel multi-layer Graph Attention Network (GAT) architecture that significantly outperforms single-layer GATs in achieving \emph{perfect node classification} in CSBMs, relaxing the SNR requirement from $ ω(\sqrt{\log n}) $ to $ ω(\sqrt{\log n} / \sqrt[3]{n}) $. To our knowledge, this is the first study to delineate the conditions for perfect node classification using multi-layer GATs. Our theoretical contributions are corroborated by extensive experiments on both synthetic and real-world datasets, highlighting the practical implications of our findings.

cs.LG↗

arXiv2025

Do We Truly Need So Many Samples? Multi-LLM Repeated Sampling Efficiently Scales Test-Time Compute

This paper presents a simple, effective, and cost-efficient strategy to improve LLM performance by scaling test-time compute. Our strategy builds upon the repeated-sampling-then-voting framework, with a novel twist: incorporating multiple models, even weaker ones, to leverage their complementary strengths that potentially arise from diverse training data and paradigms. By using consistency as a signal, our strategy dynamically switches between models. Theoretical analysis highlights the efficiency and performance advantages of our strategy. Extensive experiments on six datasets demonstrate that our strategy not only outperforms self-consistency and state-of-the-art multi-agent debate approaches, but also significantly reduces inference costs. Additionally, ModelSwitch requires only a few comparable LLMs to achieve optimal performance and can be extended with verification methods, demonstrating the potential of leveraging multiple LLMs in the generation-verification paradigm.

cs.AI↗

arXiv2025

MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-based Reinforcement Learning

DeepSeek R1, and o1 have demonstrated powerful reasoning capabilities in the text domain through stable large-scale reinforcement learning. To enable broader applications, some works have attempted to transfer these capabilities to multimodal reasoning. However, these efforts have been limited by the limited difficulty of selected tasks and relatively small training scales, making it challenging to demonstrate strong multimodal reasoning abilities. To address this gap, we introduce the MMK12 dataset and MM-EUREKA with 7B and 32B parameters. The former is a high-quality multimodal mathematics reasoning dataset featuring diverse knowledge domains with human-verified answers and solution processes. The latter is a multimodal model employing rule-based reinforcement learning on MMK12, utilizing online filtering and two-stage training strategy to enhance training stability. MM-EUREKA demonstrates remarkable performance gains in multimodal mathematical reasoning, outperforming previous powerful models like InternVL2.5-78B or InternVL2.5-38B-MPO. In particular, MM-EUREKA achieves competitive or superior performance compared to both open-source and closed-source models, and trails slightly behind o1 in multidisciplinary reasoning tasks. We open-source our complete pipeline to foster further research in this area. We release all our codes, models, data, etc. at https://github.com/ModalMinds/MM-EUREKA

cs.CV↗

arXiv2025

Capacity Region for Covert Secret Key Generation over Multiple Access Channels

We study covert secret key generation over a binary-input two-user multiple access channel with one-way public discussion and derive bounds on the capacity region. Specifically, in this problem, there are three legitimate parties: Alice, Bob and Charlie. The goal is to allow Charlie to generate a secret key with Alice and another secret key with Bob, reliably, secretly and covertly. Reliability ensures that the key generated by Alice and Charlie is the same and the key generated by Bob and Charlie is the same. Secrecy ensures that the secret keys generated are only known to specific legitimate parties. Covertness ensures that the key generation process is undetectable by a warden Willie. As a corollary of our result, we establish bounds on the capacity region of wiretap secret key generation without the covertness constraint and discuss the impact of covertness. Our results generalize the point-to-point result of Tahmasbi and Bloch (TIFS 2020) to the setting of multiterminal communication.

cs.IT↗

arXiv2025

Online Preference Alignment for Language Models via Count-based Exploration

Reinforcement Learning from Human Feedback (RLHF) has shown great potential in fine-tuning Large Language Models (LLMs) to align with human preferences. Existing methods perform preference alignment from a fixed dataset, which can be limited in data coverage, and the resulting reward model is hard to generalize in out-of-distribution responses. Thus, online RLHF is more desirable to empower the LLM to explore outside the support of the initial dataset by iteratively collecting the prompt-response pairs. In this paper, we study the fundamental problem in online RLHF, i.e. \emph{how to explore} for LLM. We give a theoretical motivation in linear reward assumption to show that an optimistic reward with an upper confidence bound (UCB) term leads to a provably efficient RLHF policy. Then, we reformulate our objective to direct preference optimization with an exploration term, where the UCB-term can be converted to a count-based exploration bonus. We further propose a practical algorithm, named \emph{Count-based Online Preference Optimization (COPO)}, which leverages a simple coin-flip counting module to estimate the pseudo-count of a prompt-response pair in previously collected data. COPO encourages LLMs to balance exploration and preference optimization in an iterative manner, which enlarges the exploration space and the entire data coverage of iterative LLM policies. We conduct online RLHF experiments on Zephyr and Llama-3 models. The results on instruction-following and standard academic benchmarks show that COPO significantly increases performance.

cs.LG↗

arXiv2025

Community Detection for Contextual-LSBM: Theoretical Limitations of Misclassification Rate and Efficient Algorithms

The integration of network information and node attribute information has recently gained significant attention in the community detection literature. In this work, we consider community detection in the Contextual Labeled Stochastic Block Model (CLSBM), where the network follows an LSBM and node attributes follow a Gaussian Mixture Model (GMM). Our primary focus is the misclassification rate, which measures the expected number of nodes misclassified by community detection algorithms. We first establish a lower bound on the optimal misclassification rate that holds for any algorithm. When we specialize our setting to the LSBM (which preserves only network information) or the GMM (which preserves only node attribute information), our lower bound recovers prior results. Moreover, we present an efficient spectral-based algorithm tailored for the CLSBM and derive an upper bound on its misclassification rate. Although the algorithm does not attain the lower bound, it serves as a reliable starting point for designing more accurate community detection algorithms (as many algorithms use spectral method as an initial step, followed by refinement procedures to enhance accuracy).

stat.ML↗

arXiv2024

A Fast Convergence Theory for Offline Decision Making

This paper proposes the first generic fast convergence result in general function approximation for offline decision making problems, which include offline reinforcement learning (RL) and off-policy evaluation (OPE) as special cases. To unify different settings, we introduce a framework called Decision Making with Offline Feedback (DMOF), which captures a wide range of offline decision making problems. Within this framework, we propose a simple yet powerful algorithm called Empirical Decision with Divergence (EDD), whose upper bound can be termed as a coefficient named Empirical Offline Estimation Coefficient (EOEC). We show that EOEC is instance-dependent and actually measures the correlation of the problem. When assuming partial coverage in the dataset, EOEC will reduce in a rate of $1/N$ where $N$ is the size of the dataset, endowing EDD with a fast convergence guarantee. Finally, we complement the above results with a lower bound in the DMOF framework, which further demonstrates the soundness of our theory.

cs.LG↗

arXiv2024

MVPatch: More Vivid Patch for Adversarial Camouflaged Attacks on Object Detectors in the Physical World

Recent studies have shown that Adversarial Patches (APs) can effectively manipulate object detection models. However, the conspicuous patterns often associated with these patches tend to attract human attention, posing a significant challenge. Existing research has primarily focused on enhancing attack efficacy in the physical domain while often neglecting the optimization of stealthiness and transferability. Furthermore, applying APs in real-world scenarios faces major challenges related to transferability, stealthiness, and practicality. To address these challenges, we introduce generalization theory into the context of APs, enabling our iterative process to simultaneously enhance transferability and refine visual correlation with realistic images. We propose a Dual-Perception-Based Framework (DPBF) to generate the More Vivid Patch (MVPatch), which enhances transferability, stealthiness, and practicality. The DPBF integrates two key components: the Model-Perception-Based Module (MPBM) and the Human-Perception-Based Module (HPBM), along with regularization terms. The MPBM employs ensemble strategy to reduce object confidence scores across multiple detectors, thereby improving AP transferability with robust theoretical support. Concurrently, the HPBM introduces a lightweight method for achieving visual similarity, creating natural and inconspicuous adversarial patches without relying on additional generative models. The regularization terms further enhance the practicality of the generated APs in the physical domain. Additionally, we introduce naturalness and transferability scores to provide an unbiased assessment of APs. Extensive experimental validation demonstrates that MVPatch achieves superior transferability and a natural appearance in both digital and physical domains, underscoring its effectiveness and stealthiness.

cs.CR↗

arXiv2024

Constrained Ensemble Exploration for Unsupervised Skill Discovery

Unsupervised Reinforcement Learning (RL) provides a promising paradigm for learning useful behaviors via reward-free per-training. Existing methods for unsupervised RL mainly conduct empowerment-driven skill discovery or entropy-based exploration. However, empowerment often leads to static skills, and pure exploration only maximizes the state coverage rather than learning useful behaviors. In this paper, we propose a novel unsupervised RL framework via an ensemble of skills, where each skill performs partition exploration based on the state prototypes. Thus, each skill can explore the clustered area locally, and the ensemble skills maximize the overall state coverage. We adopt state-distribution constraints for the skill occupancy and the desired cluster for learning distinguishable skills. Theoretical analysis is provided for the state entropy and the resulting skill distributions. Based on extensive experiments on several challenging tasks, we find our method learns well-explored ensemble skills and achieves superior performance in various downstream tasks compared to previous methods.

cs.LG↗

arXiv2024

Ensemble Successor Representations for Task Generalization in Offline-to-Online Reinforcement Learning

In Reinforcement Learning (RL), training a policy from scratch with online experiences can be inefficient because of the difficulties in exploration. Recently, offline RL provides a promising solution by giving an initialized offline policy, which can be refined through online interactions. However, existing approaches primarily perform offline and online learning in the same task, without considering the task generalization problem in offline-to-online adaptation. In real-world applications, it is common that we only have an offline dataset from a specific task while aiming for fast online-adaptation for several tasks. To address this problem, our work builds upon the investigation of successor representations for task generalization in online RL and extends the framework to incorporate offline-to-online learning. We demonstrate that the conventional paradigm using successor features cannot effectively utilize offline data and improve the performance for the new task by online fine-tuning. To mitigate this, we introduce a novel methodology that leverages offline data to acquire an ensemble of successor representations and subsequently constructs ensemble Q functions. This approach enables robust representation learning from datasets with different coverage and facilitates fast adaption of Q functions towards new tasks during the online fine-tuning phase. Extensive empirical evaluations provide compelling evidence showcasing the superior performance of our method in generalizing to diverse or even unseen tasks.

cs.LG↗

arXiv2024

Provably Efficient Information-Directed Sampling Algorithms for Multi-Agent Reinforcement Learning

This work designs and analyzes a novel set of algorithms for multi-agent reinforcement learning (MARL) based on the principle of information-directed sampling (IDS). These algorithms draw inspiration from foundational concepts in information theory, and are proven to be sample efficient in MARL settings such as two-player zero-sum Markov games (MGs) and multi-player general-sum MGs. For episodic two-player zero-sum MGs, we present three sample-efficient algorithms for learning Nash equilibrium. The basic algorithm, referred to as MAIDS, employs an asymmetric learning structure where the max-player first solves a minimax optimization problem based on the joint information ratio of the joint policy, and the min-player then minimizes the marginal information ratio with the max-player's policy fixed. Theoretical analyses show that it achieves a Bayesian regret of tilde{O}(sqrt{K}) for K episodes. To reduce the computational load of MAIDS, we develop an improved algorithm called Reg-MAIDS, which has the same Bayesian regret bound while enjoying less computational complexity. Moreover, by leveraging the flexibility of IDS principle in choosing the learning target, we propose two methods for constructing compressed environments based on rate-distortion theory, upon which we develop an algorithm Compressed-MAIDS wherein the learning target is a compressed environment. Finally, we extend Reg-MAIDS to multi-player general-sum MGs and prove that it can learn either the Nash equilibrium or coarse correlated equilibrium in a sample efficient manner.

cs.IT↗

arXiv2024

Community Detection in the Multi-View Stochastic Block Model

This paper considers the problem of community detection on multiple potentially correlated graphs from an information-theoretical perspective. We first put forth a random graph model, called the multi-view stochastic block model (MVSBM), designed to generate correlated graphs on the same set of nodes (with cardinality $n$). The $n$ nodes are partitioned into two disjoint communities of equal size. The presence or absence of edges in the graphs for each pair of nodes depends on whether the two nodes belong to the same community or not. The objective for the learner is to recover the hidden communities with observed graphs. Our technical contributions are two-fold: (i) We establish an information-theoretic upper bound (Theorem~1) showing that exact recovery of community is achievable when the model parameters of MVSBM exceed a certain threshold. (ii) Conversely, we derive an information-theoretic lower bound (Theorem~2) showing that when the model parameters of MVSBM fall below the aforementioned threshold, then for any estimator, the expected number of misclassified nodes will always be greater than one. Our results for the MVSBM recover several prior results for community detection in the standard SBM as well as in multiple independent SBMs as special cases.

cs.SI↗