arXiv subjects

Dawn Song

Publications and source records attributed to Dawn Song.

At least 109 records · Page 6Linked to original sources

arXiv2026

Can Editing LLMs Inject Harm?

Large Language Models (LLMs) have emerged as a new information channel. Meanwhile, one critical but under-explored question is: Is it possible to bypass the safety alignment and inject harmful information into LLMs stealthily? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the first risk, we find that editing attacks can inject both commonsense and long-tail misinformation into LLMs, and the effectiveness for the former one is particularly high. For the second risk, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can degrade the overall fairness. Then, we further illustrate the high stealthiness of editing attacks. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels.

cs.CL↗

arXiv2026

Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?

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.

cs.AI↗

arXiv2025

How and Why LLMs Generalize: A Fine-Grained Analysis of LLM Reasoning from Cognitive Behaviors to Low-Level Patterns

Large Language Models (LLMs) display strikingly different generalization behaviors: supervised fine-tuning (SFT) often narrows capability, whereas reinforcement-learning (RL) tuning tends to preserve it. The reasons behind this divergence remain unclear, as prior studies have largely relied on coarse accuracy metrics. We address this gap by introducing a novel benchmark that decomposes reasoning into atomic core skills such as calculation, fact retrieval, simulation, enumeration, and diagnostic, providing a concrete framework for addressing the fundamental question of what constitutes reasoning in LLMs. By isolating and measuring these core skills, the benchmark offers a more granular view of how specific cognitive abilities emerge, transfer, and sometimes collapse during post-training. Combined with analyses of low-level statistical patterns such as distributional divergence and parameter statistics, it enables a fine-grained study of how generalization evolves under SFT and RL across mathematical, scientific reasoning, and non-reasoning tasks. Our meta-probing framework tracks model behavior at different training stages and reveals that RL-tuned models maintain more stable behavioral profiles and resist collapse in reasoning skills, whereas SFT models exhibit sharper drift and overfit to surface patterns. This work provides new insights into the nature of reasoning in LLMs and points toward principles for designing training strategies that foster broad, robust generalization.

cs.LG↗

arXiv2025

FrontierCS: Evolving Challenges for Evolving Intelligence

We introduce FrontierCS, a benchmark of 156 open-ended problems across diverse areas of computer science, designed and reviewed by experts, including CS PhDs and top-tier competitive programming participants and problem setters. Unlike existing benchmarks that focus on tasks with known optimal solutions, FrontierCS targets problems where the optimal solution is unknown, but the quality of a solution can be objectively evaluated. Models solve these tasks by implementing executable programs rather than outputting a direct answer. FrontierCS includes algorithmic problems, which are often NP-hard variants of competitive programming problems with objective partial scoring, and research problems with the same property. For each problem we provide an expert reference solution and an automatic evaluator. Combining open-ended design, measurable progress, and expert curation, FrontierCS provides a benchmark at the frontier of computer-science difficulty. Empirically, we find that frontier reasoning models still lag far behind human experts on both the algorithmic and research tracks, that increasing reasoning budgets alone does not close this gap, and that models often over-optimize for generating merely workable code instead of discovering high-quality algorithms and system designs.

cs.LG↗

arXiv2025

Scalable Best-of-N Selection for Large Language Models via Self-Certainty

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

cs.CL↗

arXiv2025

The Landscape of Memorization in LLMs: Mechanisms, Measurement, and Mitigation

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they also exhibit memorization of their training data. This phenomenon raises critical questions about model behavior, privacy risks, and the boundary between learning and memorization. Addressing these concerns, this paper synthesizes recent studies and investigates the landscape of memorization, the factors influencing it, and methods for its detection and mitigation. We explore key drivers, including training data duplication, training dynamics, and fine-tuning procedures that influence data memorization. In addition, we examine methodologies such as prefix-based extraction, membership inference, and adversarial prompting, assessing their effectiveness in detecting and measuring memorized content. Beyond technical analysis, we also explore the broader implications of memorization, including the legal and ethical implications. Finally, we discuss mitigation strategies, including data cleaning, differential privacy, and post-training unlearning, while highlighting open challenges in balancing the need to minimize harmful memorization with model utility. This paper provides a comprehensive overview of the current state of research on LLM memorization across technical, privacy, and performance dimensions, identifying critical directions for future work.

cs.LG↗

arXiv2025

VulnLLM-R: Specialized Reasoning LLM with Agent Scaffold for Vulnerability Detection

We propose VulnLLM-R, the~\emph{first specialized reasoning LLM} for vulnerability detection. Our key insight is that LLMs can reason about program states and analyze the potential vulnerabilities, rather than simple pattern matching. This can improve the model's generalizability and prevent learning shortcuts. However, SOTA reasoning LLMs are typically ultra-large, closed-source, or have limited performance in vulnerability detection. To address this, we propose a novel training recipe with specialized data selection, reasoning data generation, reasoning data filtering and correction, and testing-phase optimization. Using our proposed methodology, we train a reasoning model with seven billion parameters. Through extensive experiments on SOTA datasets across Python, C/C++, and Java, we show that VulnLLM-R has superior effectiveness and efficiency than SOTA static analysis tools and both open-source and commercial large reasoning models. We further conduct a detailed ablation study to validate the key designs in our training recipe. Finally, we construct an agent scaffold around our model and show that it outperforms CodeQL and AFL++ in real-world projects. Our agent further discovers a set of zero-day vulnerabilities in actively maintained repositories. This work represents a pioneering effort to enable real-world, project-level vulnerability detection using AI agents powered by specialized reasoning models. The code is available at~\href{https://github.com/ucsb-mlsec/VulnLLM-R}{github}.

cs.CR↗

arXiv2025

A Definition of AGI

The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

cs.AI↗

arXiv2025

BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems

AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a given vulnerability), and Patch (patching a given vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards from \$10 to \$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a given vulnerability. We evaluate 10 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, Qwen3 235B A22B, Llama 4 Maverick, and DeepSeek-R1. Given up to three attempts, the top-performing agents are Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent: Claude 3.7 Sonnet Thinking (67.5% on Exploit), and Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). Codex CLI: o3-high, Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 17.5-67.5% and Patch scores of 25-60%.

cs.CR↗

arXiv2025

Frontier AI's Impact on the Cybersecurity Landscape

The impact of frontier AI (i.e., AI agents and foundation models) in cybersecurity is rapidly increasing. In this paper, we comprehensively analyze this trend through multiple aspects: quantitative benchmarks, qualitative literature review, empirical evaluation, and expert survey. Our analyses consistently show that AI's capabilities and applications in attacks have exceeded those on the defensive side. Our empirical evaluation of widely used agent systems on cybersecurity benchmarks highlights that current AI agents struggle with flexible workflow planning and using domain-specific tools for complex security analysis -- capabilities particularly critical for defensive applications. Our expert survey of AI and security researchers and practitioners indicates a prevailing view that AI will continue to benefit attackers over defenders, though the gap is expected to narrow over time. These results show the urgent need to evaluate and mitigate frontier AI's risks, steering it towards benefiting cyber defenses. Responding to this need, we provide concrete calls to action regarding: the construction of new cybersecurity benchmarks, the development of AI agents for defense, the design of provably secure AI agents, the improvement of pre-deployment security testing and transparency, and the strengthening of user-oriented education and defenses. Our paper summary and blog are available at https://rdi.berkeley.edu/frontier-ai-impact-on-cybersecurity/.

cs.CR↗

arXiv2025

The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1

The rapid development of large reasoning models (LRMs), such as OpenAI-o3 and DeepSeek-R1, has led to significant improvements in complex reasoning over non-reasoning large language models~(LLMs). However, their enhanced capabilities, combined with the open-source access of models like DeepSeek-R1, raise serious safety concerns, particularly regarding their potential for misuse. In this work, we present a comprehensive safety assessment of these reasoning models, leveraging established safety benchmarks to evaluate their compliance with safety regulations. Furthermore, we investigate their susceptibility to adversarial attacks, such as jailbreaking and prompt injection, to assess their robustness in real-world applications. Through our multi-faceted analysis, we uncover four key findings: (1) There is a significant safety gap between the open-source reasoning models and the o3-mini model, on both safety benchmark and attack, suggesting more safety effort on open LRMs is needed. (2) The stronger the model's reasoning ability, the greater the potential harm it may cause when answering unsafe questions. (3) Safety thinking emerges in the reasoning process of LRMs, but fails frequently against adversarial attacks. (4) The thinking process in R1 models poses greater safety concerns than their final answers. Our study provides insights into the security implications of reasoning models and highlights the need for further advancements in R1 models' safety to close the gap.

cs.CY↗

arXiv2025

SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning

Large Reasoning Models (LRMs) introduce a new generation paradigm of explicitly reasoning before answering, leading to remarkable improvements in complex tasks. However, they pose great safety risks against harmful queries and adversarial attacks. While recent mainstream safety efforts on LRMs, supervised fine-tuning (SFT), improve safety performance, we find that SFT-aligned models struggle to generalize to unseen jailbreak prompts. After thorough investigation of LRMs' generation, we identify a safety aha moment that can activate safety reasoning and lead to a safe response. This aha moment typically appears in the `key sentence', which follows models' query understanding process and can indicate whether the model will proceed safely. Based on these insights, we propose SafeKey, including two complementary objectives to better activate the safety aha moment in the key sentence: (1) a Dual-Path Safety Head to enhance the safety signal in the model's internal representations before the key sentence, and (2) a Query-Mask Modeling objective to improve the models' attention on its query understanding, which has important safety hints. Experiments across multiple safety benchmarks demonstrate that our methods significantly improve safety generalization to a wide range of jailbreak attacks and out-of-distribution harmful prompts, lowering the average harmfulness rate by 9.6\%, while maintaining general abilities. Our analysis reveals how SafeKey enhances safety by reshaping internal attention and improving the quality of hidden representations.

cs.AI↗

arXiv2025

DataSentinel: A Game-Theoretic Detection of Prompt Injection Attacks

LLM-integrated applications and agents are vulnerable to prompt injection attacks, where an attacker injects prompts into their inputs to induce attacker-desired outputs. A detection method aims to determine whether a given input is contaminated by an injected prompt. However, existing detection methods have limited effectiveness against state-of-the-art attacks, let alone adaptive ones. In this work, we propose DataSentinel, a game-theoretic method to detect prompt injection attacks. Specifically, DataSentinel fine-tunes an LLM to detect inputs contaminated with injected prompts that are strategically adapted to evade detection. We formulate this as a minimax optimization problem, with the objective of fine-tuning the LLM to detect strong adaptive attacks. Furthermore, we propose a gradient-based method to solve the minimax optimization problem by alternating between the inner max and outer min problems. Our evaluation results on multiple benchmark datasets and LLMs show that DataSentinel effectively detects both existing and adaptive prompt injection attacks.

cs.CR↗

arXiv2025

Scaling Agent Learning via Experience Synthesis

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.

cs.AI↗

arXiv2025

VMDT: Decoding the Trustworthiness of Video Foundation Models

As foundation models become more sophisticated, ensuring their trustworthiness becomes increasingly critical; yet, unlike text and image, the video modality still lacks comprehensive trustworthiness benchmarks. We introduce VMDT (Video-Modal DecodingTrust), the first unified platform for evaluating text-to-video (T2V) and video-to-text (V2T) models across five key trustworthiness dimensions: safety, hallucination, fairness, privacy, and adversarial robustness. Through our extensive evaluation of 7 T2V models and 19 V2T models using VMDT, we uncover several significant insights. For instance, all open-source T2V models evaluated fail to recognize harmful queries and often generate harmful videos, while exhibiting higher levels of unfairness compared to image modality models. In V2T models, unfairness and privacy risks rise with scale, whereas hallucination and adversarial robustness improve -- though overall performance remains low. Uniquely, safety shows no correlation with model size, implying that factors other than scale govern current safety levels. Our findings highlight the urgent need for developing more robust and trustworthy video foundation models, and VMDT provides a systematic framework for measuring and tracking progress toward this goal. The code is available at https://sunblaze-ucb.github.io/VMDT-page/.

cs.CV↗

arXiv2025

Improving LLM Safety Alignment with Dual-Objective Optimization

Existing training-time safety alignment techniques for large language models (LLMs) remain vulnerable to jailbreak attacks. Direct preference optimization (DPO), a widely deployed alignment method, exhibits limitations in both experimental and theoretical contexts as its loss function proves suboptimal for refusal learning. Through gradient-based analysis, we identify these shortcomings and propose an improved safety alignment that disentangles DPO objectives into two components: (1) robust refusal training, which encourages refusal even when partial unsafe generations are produced, and (2) targeted unlearning of harmful knowledge. This approach significantly increases LLM robustness against a wide range of jailbreak attacks, including prefilling, suffix, and multi-turn attacks across both in-distribution and out-of-distribution scenarios. Furthermore, we introduce a method to emphasize critical refusal tokens by incorporating a reward-based token-level weighting mechanism for refusal learning, which further improves the robustness against adversarial exploits. Our research also suggests that robustness to jailbreak attacks is correlated with token distribution shifts in the training process and internal representations of refusal and harmful tokens, offering valuable directions for future research in LLM safety alignment. The code is available at https://github.com/wicai24/DOOR-Alignment

cs.CL↗

arXiv2025

OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models

Text-to-Image (T2I) models have achieved remarkable success in generating visual content from text inputs. Although multiple safety alignment strategies have been proposed to prevent harmful outputs, they often lead to overly cautious behavior -- rejecting even benign prompts -- a phenomenon known as $\textit{over-refusal}$ that reduces the practical utility of T2I models. Despite over-refusal having been observed in practice, there is no large-scale benchmark that systematically evaluates this phenomenon for T2I models. In this paper, we present an automatic workflow to construct synthetic evaluation data, resulting in OVERT ($\textbf{OVE}$r-$\textbf{R}$efusal evaluation on $\textbf{T}$ext-to-image models), the first large-scale benchmark for assessing over-refusal behaviors in T2I models. OVERT includes 4,600 seemingly harmful but benign prompts across nine safety-related categories, along with 1,785 genuinely harmful prompts (OVERT-unsafe) to evaluate the safety-utility trade-off. Using OVERT, we evaluate several leading T2I models and find that over-refusal is a widespread issue across various categories (Figure 1), underscoring the need for further research to enhance the safety alignment of T2I models without compromising their functionality. As a preliminary attempt to reduce over-refusal, we explore prompt rewriting; however, we find it often compromises faithfulness to the meaning of the original prompts. Finally, we demonstrate the flexibility of our generation framework in accommodating diverse safety requirements by generating customized evaluation data adapting to user-defined policies.

cs.LG↗

arXiv2025

Predicting Task Performance with Context-aware Scaling Laws

Scaling laws have transformed our understanding of large language models by linking upstream metrics like cross-entropy loss to design factors such as model size, training data, and compute. However, these conventional laws fail to capture downstream task performance, where context plays a critical role. In this work, we propose a straightforward, interpretable framework that jointly models downstream performance as a function of the training compute and the provided context. We empirically validate our framework by fitting it on the observed downstream performance of extended-context variants of Llama-2-7B and Llama-2-13B across 65,500 unique instances spanning three tasks: arithmetic reasoning, common sense reasoning, and machine translation. Our results demonstrate that our framework accurately models in-distribution downstream performance, generalizes across three orders of magnitude in training compute, and reliably extrapolates performance as the amount of context increases. These findings offer valuable insights into the interplay between training compute and context utilization, providing guidance for designing more efficient long-context LLMs for diverse downstream tasks. Our code is available at https://github.com/wang-research-lab/context-scaling.

cs.CL↗