arXiv subjects

Andrew Zhao

Publications and source records attributed to Andrew Zhao.

At least 37 records · Page 2Linked to original sources

arXiv2026

Are My Optimized Prompts Compromised? Exploring Vulnerabilities of LLM-based Optimizers

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

cs.LG↗

arXiv2025

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model's reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.

cs.AI↗

arXiv2025

Fara-7B: An Efficient Agentic Model for Computer Use

Progress in computer use agents (CUAs) has been constrained by the absence of large and high-quality datasets that capture how humans interact with a computer. While LLMs have thrived on abundant textual data, no comparable corpus exists for CUA trajectories. To address these gaps, we introduce FaraGen, a novel synthetic data generation system for multi-step web tasks. FaraGen can propose diverse tasks from frequently used websites, generate multiple solution attempts, and filter successful trajectories using multiple verifiers. It achieves high throughput, yield, and diversity for multi-step web tasks, producing verified trajectories at approximately $1 each. We use this data to train Fara-7B, a native CUA model that perceives the computer using only screenshots, executes actions via predicted coordinates, and is small enough to run on-device. We find that Fara-7B outperforms other CUA models of comparable size on benchmarks like WebVoyager, Online-Mind2Web, and WebTailBench -- our novel benchmark that better captures under-represented web tasks in pre-existing benchmarks. Furthermore, Fara-7B is competitive with much larger frontier models, illustrating key benefits of scalable data generation systems in advancing small efficient agentic models. We are making Fara-7B open-weight on Microsoft Foundry and HuggingFace, and we are releasing WebTailBench.

cs.AI↗

arXiv2025

Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model's entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

cs.CL↗

arXiv2025

Scaffolded Language Models with Language Supervision for Mixed-Autonomy: A Survey

This survey organizes the intricate literature on the design and optimization of emerging structures around post-trained LMs. We refer to this overarching structure as scaffolded LMs and focus on LMs that are integrated into multi-step processes with tools. We view scaffolded LMs as semi-parametric models wherein we train non-parametric variables, including the prompt, tools, and scaffold's code. In particular, they interpret instructions, use tools, and receive feedback all in language. Recent works use an LM as an optimizer to interpret language supervision and update non-parametric variables according to intricate objectives. In this survey, we refer to this paradigm as training of scaffolded LMs with language supervision. A key feature of non-parametric training is the ability to learn from language. Parametric training excels in learning from demonstration (supervised learning), exploration (reinforcement learning), or observations (unsupervised learning), using well-defined loss functions. Language-based optimization enables rich, interpretable, and expressive objectives, while mitigating issues like catastrophic forgetting and supporting compatibility with closed-source models. Furthermore, agents are increasingly deployed as co-workers in real-world applications such as Copilot in Office tools or software development. In these mixed-autonomy settings, where control and decision-making are shared between human and AI, users point out errors or suggest corrections. Accordingly, we discuss agents that continuously improve by learning from this real-time, language-based feedback and refer to this setting as streaming learning from language supervision.

cs.CL↗

arXiv2025

Absolute Zero: Reinforced Self-play Reasoning with Zero Data

Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from outcome-based rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as an unified source of verifiable reward to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.

cs.LG↗

arXiv2025

Emulating Human-like Adaptive Vision for Efficient and Flexible Machine Visual Perception

Human vision is highly adaptive, efficiently sampling intricate environments by sequentially fixating on task-relevant regions. In contrast, prevailing machine vision models passively process entire scenes at once, resulting in excessive resource demands scaling with spatial-temporal input resolution and model size, yielding critical limitations impeding both future advancements and real-world application. Here we introduce AdaptiveNN, a general framework aiming to drive a paradigm shift from 'passive' to 'active, adaptive' vision models. AdaptiveNN formulates visual perception as a coarse-to-fine sequential decision-making process, progressively identifying and attending to regions pertinent to the task, incrementally combining information across fixations, and actively concluding observation when sufficient. We establish a theory integrating representation learning with self-rewarding reinforcement learning, enabling end-to-end training of the non-differentiable AdaptiveNN without additional supervision on fixation locations. We assess AdaptiveNN on 17 benchmarks spanning 9 tasks, including large-scale visual recognition, fine-grained discrimination, visual search, processing images from real driving and medical scenarios, language-driven embodied AI, and side-by-side comparisons with humans. AdaptiveNN achieves up to 28x inference cost reduction without sacrificing accuracy, flexibly adapts to varying task demands and resource budgets without retraining, and provides enhanced interpretability via its fixation patterns, demonstrating a promising avenue toward efficient, flexible, and interpretable computer vision. Furthermore, AdaptiveNN exhibits closely human-like perceptual behaviors in many cases, revealing its potential as a valuable tool for investigating visual cognition. Code is available at https://github.com/LeapLabTHU/AdaptiveNN.

cs.CV↗

arXiv2025

MorphEUS: Morphable Omnidirectional Unmanned System

Omnidirectional aerial vehicles (OMAVs) have opened up a wide range of possibilities for inspection, navigation, and manipulation applications using drones. In this paper, we introduce MorphEUS, a morphable co-axial quadrotor that can control position and orientation independently with high efficiency. It uses a paired servo motor mechanism for each rotor arm, capable of pointing the vectored-thrust in any arbitrary direction. As compared to the \textit{state-of-the-art} OMAVs, we achieve higher and more uniform force/torque reachability with a smaller footprint and minimum thrust cancellations. The overactuated nature of the system also results in resiliency to rotor or servo-motor failures. The capabilities of this quadrotor are particularly well-suited for contact-based infrastructure inspection and close-proximity imaging of complex geometries. In the accompanying control pipeline, we present theoretical results for full controllability, almost-everywhere exponential stability, and thrust-energy optimality. We evaluate our design and controller on high-fidelity simulations showcasing the trajectory-tracking capabilities of the vehicle during various tasks. Supplementary details and experimental videos are available on the project webpage.

cs.RO↗

arXiv2025

Learning the structure of any Hamiltonian from minimal assumptions

We study the problem of learning an unknown quantum many-body Hamiltonian $H$ from black-box queries to its time evolution $e^{-\mathrm{i} H t}$. Prior proposals for solving this task either impose some assumptions on $H$, such as its interaction structure or locality, or otherwise use an exponential amount of computational postprocessing. In this paper, we present algorithms to learn any $n$-qubit Hamiltonian, which do not need to know the Hamiltonian terms in advance, nor are they restricted to local interactions. Our algorithms are efficient as long as the number of terms $m$ is polynomially bounded in the system size $n$. We consider two models of control over the time evolution:~the first has access to time reversal ($t < 0$), enabling an algorithm that outputs an $ε$-accurate classical description of $H$ after querying its dynamics for a total of $\widetilde{\mathcal{O}}(m/ε)$ evolution time. The second access model is more conventional, allowing only forward-time evolutions;~our algorithm requires $\widetilde{\mathcal{O}}(\|H\|^3/ε^4)$ evolution time in this setting. Central to our results is the recently introduced concept of a pseudo-Choi state of $H$. We extend the utility of this learning resource by showing how to use it to learn the Fourier spectrum of $H$, how to achieve nearly Heisenberg-limited scaling with it, and how to prepare it even under our more restricted access models.

quant-ph↗

arXiv2025

Optimizing Social Media Annotation of HPV Vaccine Skepticism and Misinformation Using Large Language Models: An Experimental Evaluation of In-Context Learning and Fine-Tuning Stance Detection Across Multiple Models

This paper leverages large-language models (LLMs) to experimentally determine optimal strategies for scaling up social media content annotation for stance detection on HPV vaccine-related tweets. We examine both conventional fine-tuning and emergent in-context learning methods, systematically varying strategies of prompt engineering across widely used LLMs and their variants (e.g., GPT4, Mistral, and Llama3, etc.). Specifically, we varied prompt template design, shot sampling methods, and shot quantity to detect stance on HPV vaccination. Our findings reveal that 1) in general, in-context learning outperforms fine-tuning in stance detection for HPV vaccine social media content; 2) increasing shot quantity does not necessarily enhance performance across models; and 3) different LLMs and their variants present differing sensitivity to in-context learning conditions. We uncovered that the optimal in-context learning configuration for stance detection on HPV vaccine tweets involves six stratified shots paired with detailed contextual prompts. This study highlights the potential and provides an applicable approach for applying LLMs to research on social media stance and skepticism detection.

cs.CL↗

arXiv2025

Towards Understanding the Benefit of Multitask Representation Learning in Decision Process

Multitask Representation Learning (MRL) has emerged as a prevalent technique to improve sample efficiency in Reinforcement Learning (RL). Empirical studies have found that training agents on multiple tasks simultaneously within online and transfer learning environments can greatly improve efficiency. Despite its popularity, a comprehensive theoretical framework that elucidates its operational efficacy remains incomplete. Prior analyses have predominantly assumed that agents either possess a pre-known representation function or utilize functions from a linear class, where both are impractical. The complexity of real-world applications typically requires the use of sophisticated, non-linear functions such as neural networks as representation function, which are not pre-existing but must be learned. Our work tries to fill the gap by extending the analysis to \textit{unknown non-linear} representations, giving a comprehensive analysis for its mechanism in online and transfer learning setting. We consider the setting that an agent simultaneously playing $M$ contextual bandits (or MDPs), developing a shared representation function $ϕ$ from a non-linear function class $Φ$ using our novel Generalized Functional Upper Confidence Bound algorithm (GFUCB). We formally prove that this approach yields a regret upper bound that outperforms the lower bound associated with learning $M$ separate tasks, marking the first demonstration of MRL's efficacy in a general function class. This framework also explains the contribution of representations to transfer learning when faced with new, yet related tasks, and identifies key conditions for successful transfer. Empirical experiments further corroborate our theoretical findings.

cs.LG↗

arXiv2025

Model Surgery: Modulating LLM's Behavior Via Simple Parameter Editing

Large Language Models (LLMs) have demonstrated great potential as generalist assistants, showcasing powerful task understanding and problem-solving capabilities. To deploy LLMs as AI assistants, it is crucial that these models exhibit desirable behavioral traits, such as non-toxicity and resilience against jailbreak attempts. Current approaches for detoxification or preventing jailbreaking usually involve Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), which requires finetuning billions of parameters through gradient descent with substantial computational cost. Furthermore, models modified through SFT and RLHF may deviate from the pretrained models, potentially leading to a degradation in foundational LLM capabilities. In this paper, we observe that surprisingly, directly editing a small subset of parameters can effectively modulate specific behaviors of LLMs, such as detoxification and resistance to jailbreaking, with only inference-level computational resources. Experiments demonstrate that in the detoxification task, our approach achieves reductions of up to 90.0% in toxicity on the RealToxicityPrompts dataset and 49.2% on ToxiGen, while maintaining the LLM's general capabilities in areas such as common sense, question answering, and mathematics

cs.AI↗

arXiv2024

ExpeL: LLM Agents Are Experiential Learners

The recent surge in research interest in applying large language models (LLMs) to decision-making tasks has flourished by leveraging the extensive world knowledge embedded in LLMs. While there is a growing demand to tailor LLMs for custom decision-making tasks, finetuning them for specific tasks is resource-intensive and may diminish the model's generalization capabilities. Moreover, state-of-the-art language models like GPT-4 and Claude are primarily accessible through API calls, with their parametric weights remaining proprietary and unavailable to the public. This scenario emphasizes the growing need for new methodologies that allow learning from agent experiences without requiring parametric updates. To address these problems, we introduce the Experiential Learning (ExpeL) agent. Our agent autonomously gathers experiences and extracts knowledge using natural language from a collection of training tasks. At inference, the agent recalls its extracted insights and past experiences to make informed decisions. Our empirical results highlight the robust learning efficacy of the ExpeL agent, indicating a consistent enhancement in its performance as it accumulates experiences. We further explore the emerging capabilities and transfer learning potential of the ExpeL agent through qualitative observations and additional experiments.

cs.LG↗

arXiv2024

DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints

Recent advances in large language model assistants have made them indispensable, raising significant concerns over managing their safety. Automated red teaming offers a promising alternative to the labor-intensive and error-prone manual probing for vulnerabilities, providing more consistent and scalable safety evaluations. However, existing approaches often compromise diversity by focusing on maximizing attack success rate. Additionally, methods that decrease the cosine similarity from historical embeddings with semantic diversity rewards lead to novelty stagnation as history grows. To address these issues, we introduce DiveR-CT, which relaxes conventional constraints on the objective and semantic reward, granting greater freedom for the policy to enhance diversity. Our experiments demonstrate DiveR-CT's marked superiority over baselines by 1) generating data that perform better in various diversity metrics across different attack success rate levels, 2) better-enhancing resiliency in blue team models through safety tuning based on collected data, 3) allowing dynamic control of objective weights for reliable and controllable attack success rates, and 4) reducing susceptibility to reward overoptimization. Overall, our method provides an effective and efficient approach to LLM red teaming, accelerating real-world deployment.

cs.LG↗

arXiv2024

Expanding the reach of quantum optimization with fermionic embeddings

Quadratic programming over orthogonal matrices encompasses a broad class of hard optimization problems that do not have an efficient quantum representation. Such problems are instances of the little noncommutative Grothendieck problem (LNCG), a generalization of binary quadratic programs to continuous, noncommutative variables. In this work, we establish a natural embedding for this class of LNCG problems onto a fermionic Hamiltonian, thereby enabling the study of this classical problem with the tools of quantum information. This embedding is accomplished by a new representation of orthogonal matrices as fermionic quantum states, which we achieve through the well-known double covering of the orthogonal group. Correspondingly, the embedded LNCG Hamiltonian is a two-body fermion model. Determining extremal states of this Hamiltonian provides an outer approximation to the original problem, a quantum analogue to classical semidefinite relaxations. In particular, when optimizing over the \emph{special} orthogonal group our quantum relaxation obeys additional, powerful constraints based on the convex hull of rotation matrices. The classical size of this convex-hull representation is exponential in matrix dimension, whereas our quantum representation requires only a linear number of qubits. Finally, to project the relaxed solution back into the feasible space, we propose rounding procedures which return orthogonal matrices from appropriate measurements of the quantum state. Through numerical experiments we provide evidence that this rounded quantum relaxation can produce high-quality approximations.

quant-ph↗

arXiv2024

Empowering Interdisciplinary Insights with Dynamic Graph Embedding Trajectories

We developed DyGETViz, a novel framework for effectively visualizing dynamic graphs (DGs) that are ubiquitous across diverse real-world systems. This framework leverages recent advancements in discrete-time dynamic graph (DTDG) models to adeptly handle the temporal dynamics inherent in dynamic graphs. DyGETViz effectively captures both micro- and macro-level structural shifts within these graphs, offering a robust method for representing complex and massive dynamic graphs. The application of DyGETViz extends to a diverse array of domains, including ethology, epidemiology, finance, genetics, linguistics, communication studies, social studies, and international relations. Through its implementation, DyGETViz has revealed or confirmed various critical insights. These include the diversity of content sharing patterns and the degree of specialization within online communities, the chronological evolution of lexicons across decades, and the distinct trajectories exhibited by aging-related and non-related genes. Importantly, DyGETViz enhances the accessibility of scientific findings to non-domain experts by simplifying the complexities of dynamic graphs. Our framework is released as an open-source Python package for use across diverse disciplines. Our work not only addresses the ongoing challenges in visualizing and analyzing DTDG models but also establishes a foundational framework for future investigations into dynamic graph representation and analysis across various disciplines.

cs.LG↗

arXiv2024

Group-theoretic error mitigation enabled by classical shadows and symmetries

Estimating expectation values is a key subroutine in quantum algorithms. Near-term implementations face two major challenges: a limited number of samples required to learn a large collection of observables, and the accumulation of errors in devices without quantum error correction. To address these challenges simultaneously, we develop a quantum error-mitigation strategy called ``symmetry-adjusted classical shadows,'' by adjusting classical-shadow tomography according to how symmetries are corrupted by device errors. As a concrete example, we highlight global $\mathrm{U(1)}$ symmetry, which manifests in fermions as particle number and in spins as total magnetization, and illustrate their group-theoretic unification with respective classical-shadow protocols. We establish rigorous sampling bounds under readout errors obeying minimal assumptions, and perform numerical experiments with a more comprehensive model of gate-level errors derived from existing quantum processors. Our results reveal symmetry-adjusted classical shadows as a low-cost strategy to mitigate errors from noisy quantum experiments in the ubiquitous presence of symmetry.

quant-ph↗

arXiv2024

Exploring Text-to-Motion Generation with Human Preference

This paper presents an exploration of preference learning in text-to-motion generation. We find that current improvements in text-to-motion generation still rely on datasets requiring expert labelers with motion capture systems. Instead, learning from human preference data does not require motion capture systems; a labeler with no expertise simply compares two generated motions. This is particularly efficient because evaluating the model's output is easier than gathering the motion that performs a desired task (e.g. backflip). To pioneer the exploration of this paradigm, we annotate 3,528 preference pairs generated by MotionGPT, marking the first effort to investigate various algorithms for learning from preference data. In particular, our exploration highlights important design choices when using preference data. Additionally, our experimental results show that preference learning has the potential to greatly improve current text-to-motion generative models. Our code and dataset are publicly available at https://github.com/THU-LYJ-Lab/InstructMotion}{https://github.com/THU-LYJ-Lab/InstructMotion to further facilitate research in this area.

cs.LG↗