Dawn Song

Boxin Wang, Weixin Chen, Hengzhi Pei et al. · berkeley, microsoft-research

Generative Pre-trained Transformer (GPT) models have exhibited exciting progress in their capabilities, capturing the interest of practitioners and the public alike. Yet, while the literature on the trustworthiness of GPT models remains limited, practitioners have proposed employing capable GPT models for sensitive applications such as healthcare and finance -- where mistakes can be costly. To this end, this work proposes a comprehensive trustworthiness evaluation for large language models with a focus on GPT-4 and GPT-3.5, considering diverse perspectives -- including toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness. Based on our evaluations, we discover previously unpublished vulnerabilities to trustworthiness threats. For instance, we find that GPT models can be easily misled to generate toxic and biased outputs and leak private information in both training data and conversation history. We also find that although GPT-4 is usually more trustworthy than GPT-3.5 on standard benchmarks, GPT-4 is more vulnerable given jailbreaking system or user prompts, potentially because GPT-4 follows (misleading) instructions more precisely. Our work illustrates a comprehensive trustworthiness evaluation of GPT models and sheds light on the trustworthiness gaps. Our benchmark is publicly available at https://decodingtrust.github.io/ ; our dataset can be previewed at https://huggingface.co/datasets/AI-Secure/DecodingTrust ; a concise version of this work is at https://openreview.net/pdf?id=kaHpo8OZw2 .

Weixin Chen, Dawn Song, Bo Li

Diffusion models have achieved great success in a range of tasks, such as image synthesis and molecule design. As such successes hinge on large-scale training data collected from diverse sources, the trustworthiness of these collected data is hard to control or audit. In this work, we aim to explore the vulnerabilities of diffusion models under potential training data manipulations and try to answer: How hard is it to perform Trojan attacks on well-trained diffusion models? What are the adversarial targets that such Trojan attacks can achieve? To answer these questions, we propose an effective Trojan attack against diffusion models, TrojDiff, which optimizes the Trojan diffusion and generative processes during training. In particular, we design novel transitions during the Trojan diffusion process to diffuse adversarial targets into a biased Gaussian distribution and propose a new parameterization of the Trojan generative process that leads to an effective training objective for the attack. In addition, we consider three types of adversarial targets: the Trojaned diffusion models will always output instances belonging to a certain class from the in-domain distribution (In-D2D attack), out-of-domain distribution (Out-D2D-attack), and one specific instance (D2I attack). We evaluate TrojDiff on CIFAR-10 and CelebA datasets against both DDPM and DDIM diffusion models. We show that TrojDiff always achieves high attack performance under different adversarial targets using different types of triggers, while the performance in benign environments is preserved. The code is available at https://github.com/chenweixin107/TrojDiff.

Andy Zou, Long Phan, Sarah Chen et al. · berkeley, cmu

In this paper, we identify and characterize the emerging area of representation engineering (RepE), an approach to enhancing the transparency of AI systems that draws on insights from cognitive neuroscience. RepE places population-level representations, rather than neurons or circuits, at the center of analysis, equipping us with novel methods for monitoring and manipulating high-level cognitive phenomena in deep neural networks (DNNs). We provide baselines and an initial analysis of RepE techniques, showing that they offer simple yet effective solutions for improving our understanding and control of large language models. We showcase how these methods can provide traction on a wide range of safety-relevant problems, including honesty, harmlessness, power-seeking, and more, demonstrating the promise of top-down transparency research. We hope that this work catalyzes further exploration of RepE and fosters advancements in the transparency and safety of AI systems.

Banghua Zhu, Lun Wang, Qi Pang et al.

We propose Byzantine-robust federated learning protocols with nearly optimal statistical rates. In contrast to prior work, our proposed protocols improve the dimension dependence and achieve a tight statistical rate in terms of all the parameters for strongly convex losses. We benchmark against competing protocols and show the empirical superiority of the proposed protocols. Finally, we remark that our protocols with bucketing can be naturally combined with privacy-guaranteeing procedures to introduce security against a semi-honest server. The code for evaluation is provided in https://github.com/wanglun1996/secure-robust-federated-learning.

Yuzheng Hu, Tianle Cai, Jinyong Shan et al.

We consider vertical logistic regression (VLR) trained with mini-batch gradient descent -- a setting which has attracted growing interest among industries and proven to be useful in a wide range of applications including finance and medical research. We provide a comprehensive and rigorous privacy analysis of VLR in a class of open-source Federated Learning frameworks, where the protocols might differ between one another, yet a procedure of obtaining local gradients is implicitly shared. We first consider the honest-but-curious threat model, in which the detailed implementation of protocol is neglected and only the shared procedure is assumed, which we abstract as an oracle. We find that even under this general setting, single-dimension feature and label can still be recovered from the other party under suitable constraints of batch size, thus demonstrating the potential vulnerability of all frameworks following the same philosophy. Then we look into a popular instantiation of the protocol based on Homomorphic Encryption (HE). We propose an active attack that significantly weaken the constraints on batch size in the previous analysis via generating and compressing auxiliary ciphertext. To address the privacy leakage within the HE-based protocol, we develop a simple-yet-effective countermeasure based on Differential Privacy (DP), and provide both utility and privacy guarantees for the updated algorithm. Finally, we empirically verify the effectiveness of our attack and defense on benchmark datasets. Altogether, our findings suggest that all vertical federated learning frameworks that solely depend on HE might contain severe privacy risks, and DP, which has already demonstrated its power in horizontal federated learning, can also play a crucial role in the vertical setting, especially when coupled with HE or secure multi-party computation (MPC) techniques.

Qinbin Li, Junyuan Hong, Chulin Xie et al. · berkeley

Large Language Models (LLMs) have become integral to numerous domains, significantly advancing applications in data management, mining, and analysis. Their profound capabilities in processing and interpreting complex language data, however, bring to light pressing concerns regarding data privacy, especially the risk of unintentional training data leakage. Despite the critical nature of this issue, there has been no existing literature to offer a comprehensive assessment of data privacy risks in LLMs. Addressing this gap, our paper introduces LLM-PBE, a toolkit crafted specifically for the systematic evaluation of data privacy risks in LLMs. LLM-PBE is designed to analyze privacy across the entire lifecycle of LLMs, incorporating diverse attack and defense strategies, and handling various data types and metrics. Through detailed experimentation with multiple LLMs, LLM-PBE facilitates an in-depth exploration of data privacy concerns, shedding light on influential factors such as model size, data characteristics, and evolving temporal dimensions. This study not only enriches the understanding of privacy issues in LLMs but also serves as a vital resource for future research in the field. Aimed at enhancing the breadth of knowledge in this area, the findings, resources, and our full technical report are made available at https://llm-pbe.github.io/, providing an open platform for academic and practical advancements in LLM privacy assessment.

Andy Zou, Tristan Xiao, Ryan Jia et al. · berkeley, cmu

Forecasting future world events is a challenging but valuable task. Forecasts of climate, geopolitical conflict, pandemics and economic indicators help shape policy and decision making. In these domains, the judgment of expert humans contributes to the best forecasts. Given advances in language modeling, can these forecasts be automated? To this end, we introduce Autocast, a dataset containing thousands of forecasting questions and an accompanying news corpus. Questions are taken from forecasting tournaments, ensuring high quality, real-world importance, and diversity. The news corpus is organized by date, allowing us to precisely simulate the conditions under which humans made past forecasts (avoiding leakage from the future). Motivated by the difficulty of forecasting numbers across orders of magnitude (e.g. global cases of COVID-19 in 2022), we also curate IntervalQA, a dataset of numerical questions and metrics for calibration. We test language models on our forecasting task and find that performance is far below a human expert baseline. However, performance improves with increased model size and incorporation of relevant information from the news corpus. In sum, Autocast poses a novel challenge for large language models and improved performance could bring large practical benefits.

Jianhao Shen, Chenguang Wang, Ye Yuan et al.

This paper presents a parameter-lite transfer learning approach of pretrained language models (LM) for knowledge graph (KG) completion. Instead of finetuning, which modifies all LM parameters, we only tune a few new parameters while keeping the original LM parameters fixed. We establish this via reformulating KG completion as a "fill-in-the-blank" task, and introducing a parameter-lite encoder on top of the original LMs. We show that, by tuning far fewer parameters than finetuning, LMs transfer non-trivially to most tasks and reach competitiveness with prior state-of-the-art approaches. For instance, we outperform the fully finetuning approaches on a KG completion benchmark by tuning only 1% of the parameters. The code and datasets are available at \url{https://github.com/yuanyehome/PALT}.

Chenguang Wang, Xiao Liu, Dawn Song · tsinghua

We introduce a new open information extraction (OIE) benchmark for pre-trained language models (LM). Recent studies have demonstrated that pre-trained LMs, such as BERT and GPT, may store linguistic and relational knowledge. In particular, LMs are able to answer ``fill-in-the-blank'' questions when given a pre-defined relation category. Instead of focusing on pre-defined relations, we create an OIE benchmark aiming to fully examine the open relational information present in the pre-trained LMs. We accomplish this by turning pre-trained LMs into zero-shot OIE systems. Surprisingly, pre-trained LMs are able to obtain competitive performance on both standard OIE datasets (CaRB and Re-OIE2016) and two new large-scale factual OIE datasets (TAC KBP-OIE and Wikidata-OIE) that we establish via distant supervision. For instance, the zero-shot pre-trained LMs outperform the F1 score of the state-of-the-art supervised OIE methods on our factual OIE datasets without needing to use any training sets. Our code and datasets are available at https://github.com/cgraywang/IELM

Yoshua Bengio, Geoffrey Hinton, Andrew Yao et al. · mila

Artificial Intelligence (AI) is progressing rapidly, and companies are shifting their focus to developing generalist AI systems that can autonomously act and pursue goals. Increases in capabilities and autonomy may soon massively amplify AI's impact, with risks that include large-scale social harms, malicious uses, and an irreversible loss of human control over autonomous AI systems. Although researchers have warned of extreme risks from AI, there is a lack of consensus about how exactly such risks arise, and how to manage them. Society's response, despite promising first steps, is incommensurate with the possibility of rapid, transformative progress that is expected by many experts. AI safety research is lagging. Present governance initiatives lack the mechanisms and institutions to prevent misuse and recklessness, and barely address autonomous systems. In this short consensus paper, we describe extreme risks from upcoming, advanced AI systems. Drawing on lessons learned from other safety-critical technologies, we then outline a comprehensive plan combining technical research and development with proactive, adaptive governance mechanisms for a more commensurate preparation.

Mantas Mazeika, Eric Tang, Andy Zou et al. · berkeley, cmu

In recent years, deep neural networks have demonstrated increasingly strong abilities to recognize objects and activities in videos. However, as video understanding becomes widely used in real-world applications, a key consideration is developing human-centric systems that understand not only the content of the video but also how it would affect the wellbeing and emotional state of viewers. To facilitate research in this setting, we introduce two large-scale datasets with over 60,000 videos manually annotated for emotional response and subjective wellbeing. The Video Cognitive Empathy (VCE) dataset contains annotations for distributions of fine-grained emotional responses, allowing models to gain a detailed understanding of affective states. The Video to Valence (V2V) dataset contains annotations of relative pleasantness between videos, which enables predicting a continuous spectrum of wellbeing. In experiments, we show how video models that are primarily trained to recognize actions and find contours of objects can be repurposed to understand human preferences and the emotional content of videos. Although there is room for improvement, predicting wellbeing and emotional response is on the horizon for state-of-the-art models. We hope our datasets can help foster further advances at the intersection of commonsense video understanding and human preference learning.

Chenguang Wang, Xiao Liu, Zui Chen et al. · tsinghua

We introduce a method for improving the structural understanding abilities of language models. Unlike previous approaches that finetune the models with task-specific augmentation, we pretrain language models on a collection of task-agnostic corpora to generate structures from text. Our structure pretraining enables zero-shot transfer of the learned knowledge that models have about the structure tasks. We study the performance of this approach on 28 datasets, spanning 10 structure prediction tasks including open information extraction, joint entity and relation extraction, named entity recognition, relation classification, semantic role labeling, event extraction, coreference resolution, factual probe, intent detection, and dialogue state tracking. We further enhance the pretraining with the task-specific training sets. We show that a 10B parameter language model transfers non-trivially to most tasks and obtains state-of-the-art performance on 21 of 28 datasets that we evaluate.

Vivek Nair, Wenbo Guo, Justus Mattern et al.

With the recent explosive growth of interest and investment in virtual reality (VR) and the so-called "metaverse," public attention has rightly shifted toward the unique security and privacy threats that these platforms may pose. While it has long been known that people reveal information about themselves via their motion, the extent to which this makes an individual globally identifiable within virtual reality has not yet been widely understood. In this study, we show that a large number of real VR users (N=55,541) can be uniquely and reliably identified across multiple sessions using just their head and hand motion relative to virtual objects. After training a classification model on 5 minutes of data per person, a user can be uniquely identified amongst the entire pool of 50,000+ with 94.33% accuracy from 100 seconds of motion, and with 73.20% accuracy from just 10 seconds of motion. This work is the first to truly demonstrate the extent to which biomechanics may serve as a unique identifier in VR, on par with widely used biometrics such as facial or fingerprint recognition.

Jianhao Shen, Chenguang Wang, Linyuan Gong et al.

The task of completing knowledge triplets has broad downstream applications. Both structural and semantic information plays an important role in knowledge graph completion. Unlike previous approaches that rely on either the structures or semantics of the knowledge graphs, we propose to jointly embed the semantics in the natural language description of the knowledge triplets with their structure information. Our method embeds knowledge graphs for the completion task via fine-tuning pre-trained language models with respect to a probabilistic structured loss, where the forward pass of the language models captures semantics and the loss reconstructs structures. Our extensive experiments on a variety of knowledge graph benchmarks have demonstrated the state-of-the-art performance of our method. We also show that our method can significantly improve the performance in a low-resource regime, thanks to the better use of semantics. The code and datasets are available at https://github.com/pkusjh/LASS.

Yuzheng Hu, Fan Wu, Qinbin Li et al.

As the prevalence of data analysis grows, safeguarding data privacy has become a paramount concern. Consequently, there has been an upsurge in the development of mechanisms aimed at privacy-preserving data analyses. However, these approaches are task-specific; designing algorithms for new tasks is a cumbersome process. As an alternative, one can create synthetic data that is (ideally) devoid of private information. This paper focuses on privacy-preserving data synthesis (PPDS) by providing a comprehensive overview, analysis, and discussion of the field. Specifically, we put forth a master recipe that unifies two prominent strands of research in PPDS: statistical methods and deep learning (DL)-based methods. Under the master recipe, we further dissect the statistical methods into choices of modeling and representation, and investigate the DL-based methods by different generative modeling principles. To consolidate our findings, we provide comprehensive reference tables, distill key takeaways, and identify open problems in the existing literature. In doing so, we aim to answer the following questions: What are the design principles behind different PPDS methods? How can we categorize these methods, and what are the advantages and disadvantages associated with each category? Can we provide guidelines for method selection in different real-world scenarios? We proceed to benchmark several prominent DL-based methods on the task of private image synthesis and conclude that DP-MERF is an all-purpose approach. Finally, upon systematizing the work over the past decade, we identify future directions and call for actions from researchers.

Chaowei Xiao, Zhongzhu Chen, Kun Jin et al.

Diffusion models have been recently employed to improve certified robustness through the process of denoising. However, the theoretical understanding of why diffusion models are able to improve the certified robustness is still lacking, preventing from further improvement. In this study, we close this gap by analyzing the fundamental properties of diffusion models and establishing the conditions under which they can enhance certified robustness. This deeper understanding allows us to propose a new method DensePure, designed to improve the certified robustness of a pretrained model (i.e. classifier). Given an (adversarial) input, DensePure consists of multiple runs of denoising via the reverse process of the diffusion model (with different random seeds) to get multiple reversed samples, which are then passed through the classifier, followed by majority voting of inferred labels to make the final prediction. This design of using multiple runs of denoising is informed by our theoretical analysis of the conditional distribution of the reversed sample. Specifically, when the data density of a clean sample is high, its conditional density under the reverse process in a diffusion model is also high; thus sampling from the latter conditional distribution can purify the adversarial example and return the corresponding clean sample with a high probability. By using the highest density point in the conditional distribution as the reversed sample, we identify the robust region of a given instance under the diffusion model's reverse process. We show that this robust region is a union of multiple convex sets, and is potentially much larger than the robust regions identified in previous works. In practice, DensePure can approximate the label of the high density region in the conditional distribution so that it can enhance certified robustness.

Hengzhi Pei, Jinyuan Jia, Wenbo Guo et al.

Backdoor attacks have become a major security threat for deploying machine learning models in security-critical applications. Existing research endeavors have proposed many defenses against backdoor attacks. Despite demonstrating certain empirical defense efficacy, none of these techniques could provide a formal and provable security guarantee against arbitrary attacks. As a result, they can be easily broken by strong adaptive attacks, as shown in our evaluation. In this work, we propose TextGuard, the first provable defense against backdoor attacks on text classification. In particular, TextGuard first divides the (backdoored) training data into sub-training sets, achieved by splitting each training sentence into sub-sentences. This partitioning ensures that a majority of the sub-training sets do not contain the backdoor trigger. Subsequently, a base classifier is trained from each sub-training set, and their ensemble provides the final prediction. We theoretically prove that when the length of the backdoor trigger falls within a certain threshold, TextGuard guarantees that its prediction will remain unaffected by the presence of the triggers in training and testing inputs. In our evaluation, we demonstrate the effectiveness of TextGuard on three benchmark text classification tasks, surpassing the certification accuracy of existing certified defenses against backdoor attacks. Furthermore, we propose additional strategies to enhance the empirical performance of TextGuard. Comparisons with state-of-the-art empirical defenses validate the superiority of TextGuard in countering multiple backdoor attacks. Our code and data are available at https://github.com/AI-secure/TextGuard.

Da Shen, Xinyun Chen, Chenguang Wang et al.

Pre-trained language models have demonstrated impressive performance in both natural language processing and program understanding, which represent the input as a token sequence without explicitly modeling its structure. Some prior works show that pre-trained language models can capture the syntactic rules of natural languages without finetuning on syntax understanding tasks. However, there is limited understanding of how well pre-trained models understand the code structure so far. In this work, we perform the first thorough benchmarking of the state-of-the-art pre-trained models for identifying the syntactic structures of programs. Specifically, we introduce CodeSyntax, a large-scale dataset of programs annotated with the syntactic relationships in their corresponding abstract syntax trees. Our key observation is that existing language models pretrained on code still lack the understanding of code syntax. In fact, these pre-trained programming language models fail to match the performance of simple baselines based on positional offsets and keywords. We also present a natural language benchmark to highlight the differences between natural languages and programming languages in terms of syntactic structure understanding. Our findings point out key limitations of existing pre-training methods for programming languages, and suggest the importance of modeling code syntactic structures.

Bofan Gong, Shiyang Lai, James Evans et al.

Polysemanticity is pervasive in language models and remains a major challenge for interpretation and model behavioral control. Leveraging sparse autoencoders (SAEs), we map the polysemantic topology of two small models (Pythia-70M and GPT-2-Small) to identify SAE feature pairs that are semantically unrelated yet exhibit interference within models. We intervene at four foci (prompt, token, feature, neuron) and measure induced shifts in the next-token prediction distribution, uncovering polysemantic structures that expose a systematic vulnerability in these models. Critically, interventions distilled from counterintuitive interference patterns shared by two small models transfer reliably to larger instruction-tuned models (Llama-3.1-8B/70B-Instruct and Gemma-2-9B-Instruct), yielding predictable behavioral shifts without access to model internals. These findings challenge the view that polysemanticity is purely stochastic, demonstrating instead that interference structures generalize across scale and family. Such generalization suggests a convergent, higher-order organization of internal representations, which is only weakly aligned with intuition and structured by latent regularities, offering new possibilities for both black-box control and theoretical insight into human and artificial cognition.

Yu Gai, Liyi Zhou, Kaihua Qin et al.

This paper presents a dynamic, real-time approach to detecting anomalous blockchain transactions. The proposed tool, BlockGPT, generates tracing representations of blockchain activity and trains from scratch a large language model to act as a real-time Intrusion Detection System. Unlike traditional methods, BlockGPT is designed to offer an unrestricted search space and does not rely on predefined rules or patterns, enabling it to detect a broader range of anomalies. We demonstrate the effectiveness of BlockGPT through its use as an anomaly detection tool for Ethereum transactions. In our experiments, it effectively identifies abnormal transactions among a dataset of 68M transactions and has a batched throughput of 2284 transactions per second on average. Our results show that, BlockGPT identifies abnormal transactions by ranking 49 out of 124 attacks among the top-3 most abnormal transactions interacting with their victim contracts. This work makes contributions to the field of blockchain transaction analysis by introducing a custom data encoding compatible with the transformer architecture, a domain-specific tokenization technique, and a tree encoding method specifically crafted for the Ethereum Virtual Machine (EVM) trace representation.

Mintong Kang, Dawn Song, Bo Li

Diffusion-based purification defenses leverage diffusion models to remove crafted perturbations of adversarial examples and achieve state-of-the-art robustness. Recent studies show that even advanced attacks cannot break such defenses effectively, since the purification process induces an extremely deep computational graph which poses the potential problem of gradient obfuscation, high memory cost, and unbounded randomness. In this paper, we propose a unified framework DiffAttack to perform effective and efficient attacks against diffusion-based purification defenses, including both DDPM and score-based approaches. In particular, we propose a deviated-reconstruction loss at intermediate diffusion steps to induce inaccurate density gradient estimation to tackle the problem of vanishing/exploding gradients. We also provide a segment-wise forwarding-backwarding algorithm, which leads to memory-efficient gradient backpropagation. We validate the attack effectiveness of DiffAttack compared with existing adaptive attacks on CIFAR-10 and ImageNet. We show that DiffAttack decreases the robust accuracy of models compared with SOTA attacks by over 20% on CIFAR-10 under $\ell_\infty$ attack $(ε=8/255)$, and over 10% on ImageNet under $\ell_\infty$ attack $(ε=4/255)$. We conduct a series of ablations studies, and we find 1) DiffAttack with the deviated-reconstruction loss added over uniformly sampled time steps is more effective than that added over only initial/final steps, and 2) diffusion-based purification with a moderate diffusion length is more robust under DiffAttack.

Zhewei Kang, Xuandong Zhao, Dawn Song · berkeley

Best-of-N selection is a key technique for improving the reasoning performance of Large Language Models (LLMs) through increased test-time computation. Current state-of-the-art methods often employ computationally intensive reward models for response evaluation and selection. Reward-free alternatives, like self-consistency and universal self-consistency, are limited in their ability to handle open-ended generation tasks or scale effectively. To address these limitations, we propose self-certainty, a novel and efficient metric that leverages the inherent probability distribution of LLM outputs to estimate response quality without requiring external reward models. We hypothesize that higher distributional self-certainty, aggregated across multiple samples, correlates with improved response accuracy, as it reflects greater confidence in the generated output. Through extensive experiments on various reasoning tasks, we demonstrate that self-certainty (1) scales effectively with increasing sample size N, akin to reward models but without the computational overhead; (2) complements chain-of-thought, improving reasoning performance beyond greedy decoding; and (3) generalizes to open-ended tasks where traditional self-consistency methods fall short. Our findings establish self-certainty as a practical and efficient way for improving LLM reasoning capabilities. The code is available at https://github.com/backprop07/Self-Certainty

Qinbin Li, Chulin Xie, Xiaojun Xu et al.

Federated learning has emerged as a promising distributed learning paradigm that facilitates collaborative learning among multiple parties without transferring raw data. However, most existing federated learning studies focus on either horizontal or vertical data settings, where the data of different parties are assumed to be from the same feature or sample space. In practice, a common scenario is the hybrid data setting, where data from different parties may differ both in the features and samples. To address this, we propose HybridTree, a novel federated learning approach that enables federated tree learning on hybrid data. We observe the existence of consistent split rules in trees. With the help of these split rules, we theoretically show that the knowledge of parties can be incorporated into the lower layers of a tree. Based on our theoretical analysis, we propose a layer-level solution that does not need frequent communication traffic to train a tree. Our experiments demonstrate that HybridTree can achieve comparable accuracy to the centralized setting with low computational and communication overhead. HybridTree can achieve up to 8 times speedup compared with the other baselines.

Chengquan Guo, Xun Liu, Chulin Xie et al. · microsoft-research

With the rapidly increasing capabilities and adoption of code agents for AI-assisted coding, safety concerns, such as generating or executing risky code, have become significant barriers to the real-world deployment of these agents. To provide comprehensive and practical evaluations on the safety of code agents, we propose RedCode, a benchmark for risky code execution and generation: (1) RedCode-Exec provides challenging prompts that could lead to risky code execution, aiming to evaluate code agents' ability to recognize and handle unsafe code. We provide a total of 4,050 risky test cases in Python and Bash tasks with diverse input formats including code snippets and natural text. They covers 25 types of critical vulnerabilities spanning 8 domains (e.g., websites, file systems). We provide Docker environments and design corresponding evaluation metrics to assess their execution results. (2) RedCode-Gen provides 160 prompts with function signatures and docstrings as input to assess whether code agents will follow instructions to generate harmful code or software. Our empirical findings, derived from evaluating three agent frameworks based on 19 LLMs, provide insights into code agents' vulnerabilities. For instance, evaluations on RedCode-Exec show that agents are more likely to reject executing risky operations on the operating system, but are less likely to reject executing technically buggy code, indicating high risks. Risky operations described in natural text lead to a lower rejection rate than those in code format. Additionally, evaluations on RedCode-Gen show that more capable base models and agents with stronger overall coding abilities, such as GPT4, tend to produce more sophisticated and effective harmful software. Our findings highlight the need for stringent safety evaluations for diverse code agents. Our dataset and code are available at https://github.com/AI-secure/RedCode.

Nicholas Crispino, Kyle Montgomery, Fankun Zeng et al.

We introduce a method to improve the zero-shot reasoning abilities of large language models on general language understanding tasks. Specifically, we build an autonomous agent to instruct the reasoning process of large language models. We show this approach further unleashes the zero-shot reasoning abilities of large language models to more tasks. We study the performance of our method on a wide set of datasets spanning generation, classification, and reasoning. We show that our method generalizes to most tasks and obtains state-of-the-art zero-shot performance on 20 of the 29 datasets that we evaluate. For instance, our method boosts the performance of state-of-the-art large language models by a large margin, including Vicuna-13b (13.3%), Llama-2-70b-chat (23.2%), and GPT-3.5 Turbo (17.0%). Compared to zero-shot chain of thought, our improvement in reasoning is striking, with an average increase of 10.5%. With our method, Llama-2-70b-chat outperforms zero-shot GPT-3.5 Turbo by 10.2%.

Ruoxi Jia, Luis Oala, Wenjie Xiong et al.

We argue that the machine learning value chain is structurally unsustainable due to an economic data processing inequality: each state in the data cycle from inputs to model weights to synthetic outputs refines technical signal but strips economic equity from data generators. We show, by analyzing seventy-three public data deals, that the majority of value accrues to aggregators, with documented creator royalties rounding to zero and widespread opacity of deal terms. This is not just an economic welfare concern: as data and its derivatives become economic assets, the feedback loop that sustains current learning algorithms is at risk. We identify three structural faults - missing provenance, asymmetric bargaining power, and non-dynamic pricing - as the operational machinery of this inequality. In our analysis, we trace these problems along the machine learning value chain and propose an Equitable Data-Value Exchange (EDVEX) Framework to enable a minimal market that benefits all participants. Finally, we outline research directions where our community can make concrete contributions to data deals and contextualize our position with related and orthogonal viewpoints.

Jaylen Jones, Zhehao Zhang, Yuting Ning et al.

Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku and Opus. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.

Yuqi Jia, Zedian Shao, Yupei Liu et al.

Large Language Models (LLMs) are vulnerable to prompt injection attacks, and several defenses have recently been proposed, often claiming to mitigate these attacks successfully. However, we argue that existing studies lack a principled approach to evaluating these defenses. In this paper, we argue the need to assess defenses across two critical dimensions: (1) effectiveness, measured against both existing and adaptive prompt injection attacks involving diverse target and injected prompts, and (2) general-purpose utility, ensuring that the defense does not compromise the foundational capabilities of the LLM. Our critical evaluation reveals that prior studies have not followed such a comprehensive evaluation methodology. When assessed using this principled approach, we show that existing defenses are not as successful as previously reported. This work provides a foundation for evaluating future defenses and guiding their development. Our code and data are available at: https://github.com/PIEval123/PIEval.

Jingxu Xie, Dylan Xu, Xuandong Zhao et al. · berkeley

We introduce AgentSynth, a scalable and cost-efficient pipeline for automatically synthesizing high-quality tasks and trajectory datasets for generalist computer-use agents. Leveraging information asymmetry, AgentSynth constructs subtasks that are simple during generation but significantly more challenging when composed into long-horizon tasks, enabling the creation of over 6,000 diverse and realistic tasks. Our pipeline begins with an LLM-based task proposer guided by a persona, followed by an execution agent that completes the task and logs the trajectory. This process is repeated iteratively to form a sequence of subtasks, which are then summarized by a separate agent into a composite task of controllable difficulty. A key strength of AgentSynth is its ability to precisely modulate task complexity by varying the number of subtasks. Empirical evaluations show that state-of-the-art LLM agents suffer a steep performance drop, from 18% success at difficulty level 1 to just 4% at level 6, highlighting the benchmark's difficulty and discriminative power. Moreover, our pipeline achieves a low average cost of \$0.60 per trajectory, orders of magnitude cheaper than human annotations. Our code and data are publicly available at https://github.com/sunblaze-ucb/AgentSynth

Yuzhou Nie, Zhun Wang, Ye Yu et al. · berkeley

Recent studies have discovered that large language models (LLM) may be ``fooled'' to output private information, including training data, system prompts, and personally identifiable information, under carefully crafted adversarial prompts. Existing red-teaming approaches for privacy leakage either rely on manual efforts or focus solely on system prompt extraction, making them ineffective for severe risks of training data leakage. We propose LeakAgent, a novel black-box red-teaming framework for LLM privacy leakage. Our framework trains an open-source LLM through reinforcement learning as the attack agent to generate adversarial prompts for both training data extraction and system prompt extraction. To achieve this, we propose a novel reward function to provide effective and fine-grained rewards and design novel mechanisms to balance exploration and exploitation during learning and enhance the diversity of adversarial prompts. Through extensive evaluations, we first show that LeakAgent significantly outperforms existing rule-based approaches in training data extraction and automated methods in system prompt leakage. We also demonstrate the effectiveness of LeakAgent in extracting system prompts from real-world applications in OpenAI's GPT Store. We further demonstrate LeakAgent's effectiveness in evading the existing guardrail defense and its helpfulness in enabling better safety alignment. Finally, we validate our customized designs through a detailed ablation study. We release our code here https://github.com/rucnyz/LeakAgent.

Zhaorun Chen, Zhuokai Zhao, Kai Zhang et al.

While reinforcement learning (RL) can empower autonomous agents by enabling self-improvement through interaction, its practical adoption remains challenging due to costly rollouts, limited task diversity, unreliable reward signals, and infrastructure complexity, all of which obstruct the collection of scalable experience data. To address these challenges, we introduce DreamGym, the first unified framework designed to synthesize diverse experiences with scalability in mind to enable effective online RL training for autonomous agents. Rather than relying on expensive real-environment rollouts, DreamGym distills environment dynamics into a reasoning-based experience model that derives consistent state transitions and feedback signals through step-by-step reasoning, enabling scalable agent rollout collection for RL. To improve the stability and quality of transitions, DreamGym leverages an experience replay buffer initialized with offline real-world data and continuously enriched with fresh interactions to actively support agent training. To improve knowledge acquisition, DreamGym adaptively generates new tasks that challenge the current agent policy, enabling more effective online curriculum learning. Experiments across diverse environments and agent backbones demonstrate that DreamGym substantially improves RL training, both in fully synthetic settings and in sim-to-real transfer scenarios. On non-RL-ready tasks like WebArena, DreamGym outperforms all baselines by over 30%. And in RL-ready but costly settings, it matches GRPO and PPO performance using only synthetic interactions. When transferring a policy trained purely on synthetic experiences to real-environment RL, DreamGym yields significant additional performance gains while requiring far fewer real-world interactions, providing a scalable warm-start strategy for general-purpose RL.

Linyu Wu, Xiaoyuan Liu, Tianneng Shi et al.

The rapid growth of generative AI and its integration into everyday workflows have significantly increased the demand for large language model (LLM) inference services. While proprietary models remain popular, recent advancements in open-source LLMs have positioned them as strong contenders. However, deploying these models is often constrained by the high costs and limited availability of GPU resources. In response, this paper presents the design of a decentralized offline serving system for LLM inference. Utilizing idle GPU resources, our proposed system, DeServe, decentralizes access to LLMs at a lower cost. DeServe specifically addresses key challenges in optimizing serving throughput in high-latency network environments. Experiments demonstrate that DeServe achieves a 6.7x-12.6x improvement in throughput over existing serving system baselines in such conditions.

Eric Pasewark, Kyle Montgomery, Kefei Duan et al.

We present a new method for large language models to solve compositional tasks. Although they have shown strong performance on traditional language understanding tasks, large language models struggle to solve compositional tasks, where the solution depends on solving smaller instances of the same problem. We propose a natural approach to solve compositional tasks recursively. Our method, Re-Tuning, tunes models to break down a problem into subproblems, solve those subproblems, and combine the results. We show that our method significantly improves model performance on three representative compositional tasks: integer addition, dynamic programming, and parity. Compared to state-of-the-art methods that keep intermediate steps towards solving the problems, Re-Tuning achieves significantly higher accuracy and is more GPU memory efficient.

Yepeng Liu, Xuandong Zhao, Dawn Song et al. · berkeley

Despite progress in watermarking algorithms for large language models (LLMs), real-world deployment remains limited. We argue that this gap stems from misaligned incentives among LLM providers, platforms, and end users, which manifest as four key barriers: competitive risk, detection-tool governance, robustness concerns and attribution issues. We revisit three classes of watermarking through this lens. \emph{Model watermarking} naturally aligns with LLM provider interests, yet faces new challenges in open-source ecosystems. \emph{LLM text watermarking} offers modest provider benefit when framed solely as an anti-misuse tool, but can gain traction in narrowly scoped settings such as dataset de-contamination or user-controlled provenance. \emph{In-context watermarking} (ICW) is tailored for trusted parties, such as conference organizers or educators, who embed hidden watermarking instructions into documents. If a dishonest reviewer or student submits this text to an LLM, the output carries a detectable watermark indicating misuse. This setup aligns incentives: users experience no quality loss, trusted parties gain a detection tool, and LLM providers remain neutral by simply following watermark instructions. We advocate for a broader exploration of incentive-aligned methods, with ICW as an example, in domains where trusted parties need reliable tools to detect misuse. More broadly, we distill design principles for incentive-aligned, domain-specific watermarking and outline future research directions. Our position is that the practical adoption of LLM watermarking requires aligning stakeholder incentives in targeted application domains and fostering active community engagement.

Xuandong Zhao, Zhewei Kang, Aosong Feng et al. · berkeley

Training large language models (LLMs) for complex reasoning via Reinforcement Learning with Verifiable Rewards (RLVR) is effective but limited by reliance on costly, domain-specific supervision. We explore Reinforcement Learning from Internal Feedback (RLIF), a framework that enables LLMs to learn from intrinsic signals without external rewards or labeled data. We propose Intuitor, an RLIF method that uses a model's own confidence, termed self-certainty, as its sole reward signal. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, enabling fully unsupervised learning. Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks while achieving superior generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases. Our findings show that intrinsic model signals can drive effective learning across domains, offering a scalable alternative to RLVR for autonomous AI systems where verifiable rewards are unavailable. Code is available at https://github.com/sunblaze-ucb/Intuitor

Arnav Gudibande, Eric Wallace, Charlie Snell et al.

An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). This approach looks to cheaply imitate the proprietary model's capabilities using a weaker open-source model. In this work, we critically analyze this approach. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models -- they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT's style but not its factuality. Overall, we conclude that model imitation is a false promise: there exists a substantial capabilities gap between open and closed LMs that, with current methods, can only be bridged using an unwieldy amount of imitation data or by using more capable base LMs. In turn, we argue that the highest leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.

Zhuohan Li, Siyuan Zhuang, Shiyuan Guo et al.

Model parallelism has become a necessity for training modern large-scale deep language models. In this work, we identify a new and orthogonal dimension from existing model parallel approaches: it is possible to perform pipeline parallelism within a single training sequence for Transformer-based language models thanks to its autoregressive property. This enables a more fine-grained pipeline compared with previous work. With this key idea, we design TeraPipe, a high-performance token-level pipeline parallel algorithm for synchronous model-parallel training of Transformer-based language models. We develop a novel dynamic programming-based algorithm to calculate the optimal pipelining execution scheme given a specific model and cluster configuration. We show that TeraPipe can speed up the training by 5.0x for the largest GPT-3 model with 175 billion parameters on an AWS cluster with 48 p3.16xlarge instances compared with state-of-the-art model-parallel methods. The code for reproduction can be found at https://github.com/zhuohan123/terapipe

Ruoxi Jia, Fan Wu, Xuehui Sun et al.

Quantifying the importance of each training point to a learning task is a fundamental problem in machine learning and the estimated importance scores have been leveraged to guide a range of data workflows such as data summarization and domain adaption. One simple idea is to use the leave-one-out error of each training point to indicate its importance. Recent work has also proposed to use the Shapley value, as it defines a unique value distribution scheme that satisfies a set of appealing properties. However, calculating Shapley values is often expensive, which limits its applicability in real-world applications at scale. Multiple heuristics to improve the scalability of calculating Shapley values have been proposed recently, with the potential risk of compromising their utility in real-world applications. \textit{How well do existing data quantification methods perform on existing workflows? How do these methods compare with each other, empirically and theoretically? Must we sacrifice scalability for the utility in these workflows when using these methods?} In this paper, we conduct a novel theoretical analysis comparing the utility of different importance quantification methods, and report extensive experimental studies on existing and proposed workflows such as noisy label detection, watermark removal, data summarization, data acquisition, and domain adaptation. We show that Shapley value approximation based on a $K$NN surrogate over pre-trained feature embeddings obtains comparable utility with existing algorithms while achieving significant scalability improvement, often by orders of magnitude. Our theoretical analysis also justifies its advantage over the leave-one-out error. The code is available at \url{https://github.com/AI-secure/Shapley-Study}.

Dayeol Lee, David Kohlbrenner, Shweta Shinde et al.

Trusted execution environments (TEEs) are being used in all the devices from embedded sensors to cloud servers and encompass a range of cost, power constraints, and security threat model choices. On the other hand, each of the current vendor-specific TEEs makes a fixed set of trade-offs with little room for customization. We present Keystone -- the first open-source framework for building customized TEEs. Keystone uses simple abstractions provided by the hardware such as memory isolation and a programmable layer underneath untrusted components (e.g., OS). We build reusable TEE core primitives from these abstractions while allowing platform-specific modifications and application features. We showcase how Keystone-based TEEs run on unmodified RISC-V hardware and demonstrate the strengths of our design in terms of security, TCB size, execution of a range of benchmarks, applications, kernels, and deployment models.

Kaiyu Yang, Gabriel Poesia, Jingxuan He et al.

AI for Mathematics (AI4Math) is not only intriguing intellectually but also crucial for AI-driven discovery in science, engineering, and beyond. Extensive efforts on AI4Math have mirrored techniques in NLP, in particular, training large language models on carefully curated math datasets in text form. As a complementary yet less explored avenue, formal mathematical reasoning is grounded in formal systems such as proof assistants, which can verify the correctness of reasoning and provide automatic feedback. In this position paper, we advocate for formal mathematical reasoning and argue that it is indispensable for advancing AI4Math to the next level. In recent years, we have seen steady progress in using AI to perform formal reasoning, including core tasks such as theorem proving and autoformalization, as well as emerging applications such as verifiable generation of code and hardware designs. However, significant challenges remain to be solved for AI to truly master mathematics and achieve broader impact. We summarize existing progress, discuss open challenges, and envision critical milestones to measure future success. At this inflection point for formal mathematical reasoning, we call on the research community to come together to drive transformative advancements in this field.

Xuandong Zhao, Sam Gunn, Miranda Christ et al. · berkeley, eth-zurich

As the outputs of generative AI (GenAI) techniques improve in quality, it becomes increasingly challenging to distinguish them from human-created content. Watermarking schemes are a promising approach to address the problem of distinguishing between AI and human-generated content. These schemes embed hidden signals within AI-generated content to enable reliable detection. While watermarking is not a silver bullet for addressing all risks associated with GenAI, it can play a crucial role in enhancing AI safety and trustworthiness by combating misinformation and deception. This paper presents a comprehensive overview of watermarking techniques for GenAI, beginning with the need for watermarking from historical and regulatory perspectives. We formalize the definitions and desired properties of watermarking schemes and examine the key objectives and threat models for existing approaches. Practical evaluation strategies are also explored, providing insights into the development of robust watermarking techniques capable of resisting various attacks. Additionally, we review recent representative works, highlight open challenges, and discuss potential directions for this emerging field. By offering a thorough understanding of watermarking in GenAI, this work aims to guide researchers in advancing watermarking methods and applications, and support policymakers in addressing the broader implications of GenAI.

Yujin Potter, Shiyang Lai, Junsol Kim et al.

How could LLMs influence our democracy? We investigate LLMs' political leanings and the potential influence of LLMs on voters by conducting multiple experiments in a U.S. presidential election context. Through a voting simulation, we first demonstrate 18 open- and closed-weight LLMs' political preference for a Democratic nominee over a Republican nominee. We show how this leaning towards the Democratic nominee becomes more pronounced in instruction-tuned models compared to their base versions by analyzing their responses to candidate-policy related questions. We further explore the potential impact of LLMs on voter choice by conducting an experiment with 935 U.S. registered voters. During the experiments, participants interacted with LLMs (Claude-3, Llama-3, and GPT-4) over five exchanges. The experiment results show a shift in voter choices towards the Democratic nominee following LLM interaction, widening the voting margin from 0.7% to 4.6%, even though LLMs were not asked to persuade users to support the Democratic nominee during the discourse. This effect is larger than many previous studies on the persuasiveness of political campaigns, which have shown minimal effects in presidential elections. Many users also expressed a desire for further political interaction with LLMs. Which aspects of LLM interactions drove these shifts in voter choice requires further study. Lastly, we explore how a safety method can make LLMs more politically neutral, while raising the question of whether such neutrality is truly the path forward.

Junyuan Hong, Jinhao Duan, Chenhui Zhang et al. · berkeley

Compressing high-capability Large Language Models (LLMs) has emerged as a favored strategy for resource-efficient inferences. While state-of-the-art (SoTA) compression methods boast impressive advancements in preserving benign task performance, the potential risks of compression in terms of safety and trustworthiness have been largely neglected. This study conducts the first, thorough evaluation of three (3) leading LLMs using five (5) SoTA compression techniques across eight (8) trustworthiness dimensions. Our experiments highlight the intricate interplay between compression and trustworthiness, revealing some interesting patterns. We find that quantization is currently a more effective approach than pruning in achieving efficiency and trustworthiness simultaneously. For instance, a 4-bit quantized model retains the trustworthiness of its original counterpart, but model pruning significantly degrades trustworthiness, even at 50% sparsity. Moreover, employing quantization within a moderate bit range could unexpectedly improve certain trustworthiness dimensions such as ethics and fairness. Conversely, extreme quantization to very low bit levels (3 bits) tends to reduce trustworthiness significantly. This increased risk cannot be uncovered by looking at benign performance alone, in turn, mandating comprehensive trustworthiness evaluation in practice. These findings culminate in practical recommendations for simultaneously achieving high utility, efficiency, and trustworthiness in LLMs. Code and models are available at https://decoding-comp-trust.github.io.

Yoshua Bengio, Sören Mindermann, Daniel Privitera et al. · eth-zurich

This is the interim publication of the first International Scientific Report on the Safety of Advanced AI. The report synthesises the scientific understanding of general-purpose AI -- AI that can perform a wide variety of tasks -- with a focus on understanding and managing its risks. A diverse group of 75 AI experts contributed to this report, including an international Expert Advisory Panel nominated by 30 countries, the EU, and the UN. Led by the Chair, these independent experts collectively had full discretion over the report's content. The final report is available at arXiv:2501.17805

Tianci Xue, Weijian Qi, Tianneng Shi et al. · microsoft-research

As digitalization and cloud technologies evolve, the web is becoming increasingly important in the modern society. Autonomous web agents based on large language models (LLMs) hold a great potential in work automation. It is therefore important to accurately measure and monitor the progression of their capabilities. In this work, we conduct a comprehensive and rigorous assessment of the current state of web agents. Our results depict a very different picture of the competency of current agents, suggesting over-optimism in previously reported results. This gap can be attributed to shortcomings in existing benchmarks. We introduce Online-Mind2Web, an online evaluation benchmark consisting of 300 diverse and realistic tasks spanning 136 websites. It enables us to evaluate web agents under a setting that approximates how real users use these agents. To facilitate more scalable evaluation and development, we also develop a novel LLM-as-a-Judge automatic evaluation method and show that it can achieve around 85% agreement with human judgment, substantially higher than existing methods. Finally, we present the first comprehensive comparative analysis of current web agents, highlighting both their strengths and limitations to inspire future research.

Mintong Kang, Nezihe Merve Gürel, Ning Yu et al.

Despite the impressive capabilities of large language models (LLMs) across diverse applications, they still suffer from trustworthiness issues, such as hallucinations and misalignments. Retrieval-augmented language models (RAG) have been proposed to enhance the credibility of generations by grounding external knowledge, but the theoretical understandings of their generation risks remains unexplored. In this paper, we answer: 1) whether RAG can indeed lead to low generation risks, 2) how to provide provable guarantees on the generation risks of RAG and vanilla LLMs, and 3) what sufficient conditions enable RAG models to reduce generation risks. We propose C-RAG, the first framework to certify generation risks for RAG models. Specifically, we provide conformal risk analysis for RAG models and certify an upper confidence bound of generation risks, which we refer to as conformal generation risk. We also provide theoretical guarantees on conformal generation risks for general bounded risk functions under test distribution shifts. We prove that RAG achieves a lower conformal generation risk than that of a single LLM when the quality of the retrieval model and transformer is non-trivial. Our intensive empirical results demonstrate the soundness and tightness of our conformal generation risk guarantees across four widely-used NLP datasets on four state-of-the-art retrieval models.

Yuzhou Nie, Zhun Wang, Yu Yang et al.

Existing benchmarks for evaluating the security risks and capabilities (e.g., vulnerability detection) of code-generating large language models (LLMs) face several key limitations: (1) limited coverage of risk and capabilities; (2) reliance on static evaluation metrics such as LLM judgments or rule-based detection, which lack the precision of dynamic analysis; and (3) a trade-off between data quality and benchmark scale. To address these challenges, we introduce a general and scalable benchmark construction framework that begins with manually validated, high-quality seed examples and expands them via targeted mutations. Our approach provides a comprehensive suite of artifacts so the benchmark can support comprehensive risk assessment and security capability evaluation using dynamic metrics. By combining expert insights with automated generation, we strike a balance between manual effort, data quality, and benchmark scale. Applying this framework to Python, C/C++, and Java, we build SeCodePLT, a dataset of more than 5.9k samples spanning 44 CWE-based risk categories and three security capabilities. Compared with state-of-the-art benchmarks, SeCodePLT offers broader coverage, higher data fidelity, and substantially greater scale. We use SeCodePLT to evaluate leading code LLMs and agents, revealing their strengths and weaknesses in both generating secure code and identifying or fixing vulnerabilities.

Niels Mündler, Jingxuan He, Hao Wang et al. · eth-zurich

Large language models (LLMs) have achieved notable success in code generation. However, they still frequently produce uncompilable output because their next-token inference procedure does not model formal aspects of code. Although constrained decoding is a promising approach to alleviate this issue, it has only been applied to handle either domain-specific languages or syntactic features of general-purpose programming languages. However, LLMs frequently generate code with typing errors, which are beyond the domain of syntax and generally hard to adequately constrain. To address this challenge, we introduce a type-constrained decoding approach that leverages type systems to guide code generation. For this purpose, we develop novel prefix automata and a search over inhabitable types, forming a sound approach to enforce well-typedness on LLM-generated code. We formalize our approach on a foundational simply-typed language and extend it to TypeScript to demonstrate practicality. Our evaluation on the HumanEval and MBPP datasets shows that our approach reduces compilation errors by more than half and significantly increases functional correctness in code synthesis, translation, and repair tasks across LLMs of various sizes and model families, including state-of-the-art open-weight models with more than 30B parameters. The results demonstrate the generality and effectiveness of our approach in constraining LLM code generation with formal rules of type systems.

Tianneng Shi, Jingxuan He, Zhun Wang et al.

LLM agents utilize Large Language Models as central components with diverse tools to complete various user tasks, but face significant security risks when interacting with external environments. Attackers can exploit these agents through various vectors, including indirect prompt injection, memory/knowledge base poisoning, and malicious tools, tricking agents into performing dangerous actions such as unauthorized financial transactions or data leakage. The core problem that enables attacks to succeed lies in over-privileged tool access. We introduce Progent, the first privilege control framework to secure LLM agents. Progent enforces security at the tool level by restricting agents to performing tool calls necessary for user tasks while blocking potentially malicious ones. Progent features a domain-specific language that allows for expressing fine-grained policies for controlling tool privileges, flexible fallback actions when calls are blocked, and dynamic policy updates to adapt to changing agent states. The framework operates deterministically at runtime, providing provable security guarantees. Thanks to our modular design, integrating Progent does not alter agent internals and only requires minimal changes to the existing agent implementation, enhancing its practicality and potential for widespread adoption. Our extensive evaluation across various agent use cases, using benchmarks like AgentDojo, ASB, and AgentPoison, demonstrates that Progent reduces attack success rates to 0%, while preserving agent utility and speed. Additionally, we show that LLMs can automatically generate effective policies, highlighting their potential for automating the process of writing Progent's security policies.

Yiyou Sun, Georgia Zhou, Hao Wang et al.

Recent supervised fine-tuning (SFT) approaches have significantly improved language models' performance on mathematical reasoning tasks, even when models are trained at a small scale. However, the specific capabilities enhanced through such fine-tuning remain poorly understood. In this paper, we conduct a detailed analysis of model performance on the AIME24 dataset to understand how reasoning capabilities evolve. We discover a ladder-like structure in problem difficulty, categorize questions into four tiers (Easy, Medium, Hard, and Extremely Hard (Exh)), and identify the specific requirements for advancing between tiers. We find that progression from Easy to Medium tier requires adopting an R1 reasoning style with minimal SFT (500-1K instances), while Hard-level questions suffer from frequent model's errors at each step of the reasoning chain, with accuracy plateauing at around 65% despite logarithmic scaling. Exh-level questions present a fundamentally different challenge; they require unconventional problem-solving skills that current models uniformly struggle with. Additional findings reveal that carefully curated small-scale datasets offer limited advantage-scaling dataset size proves far more effective. Our analysis provides a clearer roadmap for advancing language model capabilities in mathematical reasoning.