Reward functions, learned or manually specified, are rarely perfect. Instead of accurately expressing human goals, these reward functions are often distorted by human beliefs about how best to achieve those goals. Specifically, these reward functions often express a combination of the human's terminal goals — those which are ends in themselves — and the human's instrumental goals — those which are means to an end. We formulate a simple example in which even slight conflation of instrumental and terminal goals results in severe misalignment: optimizing the misspecified reward function r̂ results in poor performance when measured by the true reward function r. This example distills the essential properties of environments that make reinforcement learning highly sensitive to conflation of instrumental and terminal goals. We discuss how this issue can arise with a common approach to reward learning and how it can manifest in real environments.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
DETONATE – A Benchmark for Text-to-Image Alignment and Kernelized Direct Preference Optimization
PDF ↗Alignment is crucial for text-to-image (T2I) models to ensure that the generated images faithfully capture user intent while maintaining safety and fairness. Direct Preference Optimization (DPO) has emerged as a key alignment technique for large language models (LLMs), and its influence is now extending to T2I systems. This paper introduces DPO-Kernels for T2I models, a novel extension of DPO that enhances alignment across three key dimensions: (i) Hybrid Loss, which integrates embedding-based objectives with the traditional probability-based loss to improve optimization; (ii) Kernelized Representations, leveraging Radial Basis Function (RBF), Polynomial, and Wavelet kernels to enable richer feature transformations, ensuring better separation between safe and unsafe inputs; and (iii) Divergence Selection, expanding beyond DPO’s default Kullback–Leibler (KL) regularizer by incorporating alternative divergence measures such as Wasserstein and Rényi divergences to enhance stability and robustness in alignment training. We introduce DETONATE, the first large-scale benchmark of its kind, comprising approximately 100K curated image pairs, categorized as chosen and rejected. This benchmark encapsulates three critical axes of social bias and discrimination: Race, Gender, and Disability. The prompts are sourced from the hate speech datasets, while the images are generated using state-of-the-art T2I models, including Stable Diffusion 3.5 Large (SD-3.5), Stable Diffusion XL (SD-XL), and Midjourney. Furthermore, to evaluate alignment beyond surface metrics, we introduce the Alignment Quality Index (AQI) for T2I systems: a novel geometric measure that quantifies latent space separability of safe/unsafe image activations, revealing hidden model vulnerabilities. While alignment techniques often risk overfitting, we empirically demonstrate that DPO-Kernels preserve strong generalization bounds using the theory of Heavy-Tailed Self-Regularization (HT-SR).
Language models (LMs) judges are widely used to evaluate the quality of LM outputs. Despite many advantages, LM judges display concerning biases that can impair their integrity in evaluations. One such bias is self-preference: LM judges preferring their own answers over those produced by other LMs or humans. The bias is hard to eliminate as frontier LM judges can distinguish their own outputs from those of others, even when the evaluation candidates are not labeled with their sources. In this paper, we investigate strategies to mitigate self-preference by reducing the LM judges' ability to recognize their own outputs. We apply black-box perturbations to evaluation candidates in pairwise comparison to obfuscate the authorship and reduce self-recognition. We find that perturbations as simple as synonym replacement for a few words predictably reduce self-preference. However, we also uncover fundamental challenges to eliminating the bias: when we extrapolate our perturbations to a more complete neutralization of stylistic differences between the evaluation candidates, self-preference recovers. Our findings suggest that self-recognition and self-preference can happen on many semantic levels, and complete mitigation remains challenging despite promising initial results.
Reward-model-based fine-tuning is a central paradigm in aligning Large Language Models with human preferences. However, such approaches critically rely on the assumption that proxy reward models accurately reflect intended supervision, a condition often violated due to annotation noise, bias, or limited coverage. This misalignment can lead to undesirable behaviors, where models optimize for flawed signals rather than true human values. In this paper, we investigate a novel framework to identify and mitigate such misalignment by treating the fine-tuning process as a form of knowledge integration. We focus on detecting instances of proxy-policy conflicts, cases where the base model strongly disagrees with the proxy. We argue that such conflicts often signify areas of shared ignorance, where neither the policy nor the reward model possesses sufficient knowledge, making them especially susceptible to misalignment. To this end, we propose two complementary metrics for identifying these conflicts: a localized Proxy-Policy Alignment Conflict Score (PACS) and a global Kendall-Tau Distance measure. Building on this insight, we design an algorithm named Selective Human-in-the-loop Feedback via Conflict-Aware Sampling (SHF-CAS) that targets high-conflict QA pairs for additional feedback, refining both the reward model and policy efficiently. Experiments on two alignment tasks demonstrate that our approach enhances general alignment performance, even when trained with a biased proxy reward. Our work provides a new lens for interpreting alignment failures and offers a principled pathway for targeted refinement in LLM training.
With the widespread deployment of large language models (LLMs) in human-computer interaction, dark patterns have extended from traditional visual interfaces to conversational AI systems. While existing research has confirmed the prevalence of dark patterns in LLMs, current evaluation benchmarks face critical challenges including limited classification coverage, overlooked risks specific to reasoning models, and inadequate consideration of cross-linguistic differences. To address these limitations, we propose DarkBench+, an extended benchmark for evaluating dark patterns in LLMs. We construct an expanded taxonomy containing 10 major categories and 24 subcategories, introduce an annotation workflow combining manual and automated methods, and design 2,088 bilingual test samples in Chinese and English. This benchmark is the first to develop specialized evaluation dimensions for reasoning models and systematically evaluates dark pattern behaviors across nearly 40 mainstream LLMs. Experimental results demonstrate significant manipulation risks in reasoning models' transparency displays, while cross-linguistic evaluation analyzes AI manipulation behavior differences across different linguistic environments, promoting more ethical and responsible LLM development.
Today's large language models (LLMs) are capable of supporting multilingual scenarios, allowing users to interact with LLMs in their native languages. When LLMs respond to subjective questions posed by users, they are expected to align with the views of specific demographic groups or historical periods, shaped by the language in which the user interacts with the model. Existing studies mainly focus on researching the opinions represented by LLMs among demographic groups in the United States or a few countries, lacking worldwide country samples and studies on human opinions in different historical periods, as well as lacking discussion on using language to steer LLMs. Moreover, they also overlook the potential influence of prompt language on the alignment of LLMs' opinions. In this study, our goal is to fill these gaps. To this end, we create an evaluation framework based on the World Values Survey (WVS) to systematically assess the alignment of LLMs with human opinions across different countries, languages, and historical periods around the world. We find that LLMs appropriately or over-align the opinions with only a few countries while under-aligning the opinions with most countries. Furthermore, changing the language of the prompt to match the language used in the questionnaire can effectively steer LLMs to align with the opinions of the corresponding country more effectively than existing steering methods. At the same time, LLMs are more aligned with the opinions of the contemporary population. To our knowledge, our study is the first comprehensive investigation of the topic of opinion alignment in LLMs across global, language, and temporal dimensions.
Reinforcement learning (RL) has recently become a powerful yet resource-intensive approach for post-training large language models (LLMs). Incorporating curriculum learning (CL) into RL has been shown to significantly improve training efficiency, particularly in reasoning tasks. However, existing CL methods face substantial challenges in multi-objective RL (MORL) settings, including: (1) difficulty in evaluating model capabilities online, (2) challenges in assessing sample importance under diverse objectives, and (3) inherent trade-offs between online training and offline inference in dynamically designing the curriculum. To address these issues, we propose a Multi-Reward space guided Adaptive Curriculum Learning framework (MRACL), which is the first to incorporate curriculum learning into multi-objective RL. MRACL first constructs a multi-dimensional reward space via offline inference to establish initial reward profiles for each training sample. During training, based on reward space, it estimates the evolving model capabilities by computing the centroid of the space and calculates the sample priority score through its capability distance, optimization direction, and historical evolution, which enables adaptive selection of the most informative training samples at each step, independent of the specific RL algorithm. After each RL training iteration, the reward space is dynamically updated to reflect the model's evolving capabilities and the shifting distribution of sample priorities. Experiments on multi-objective alignment tasks demonstrate that MRACL achieves 1.62× faster convergence compared to state-of-the-art curriculum methods and 2.55× faster than non-curriculum methods. Furthermore, it consistently outperforms all baselines in both win rate and rule-based evaluation. We further provide an in-depth analysis of the key factors contributing to \modelname's effectiveness, along with its advantages, scenarios, and generalization across diverse settings.
3D generation and reconstruction techniques have been widely used in computer games, film, and other content creation areas. As the application grows, there is a growing demand for 3D shapes that look truly realistic. Traditional evaluation methods rely on a ground truth to measure mesh fidelity. However, in many practical cases, a shape's realism does not depend on having a ground truth reference. In this work, we propose a Shape-Realism Alignment Metric that leverages a large language model (LLM) as a bridge between mesh shape information and realism evaluation. To achieve this, we adopt a mesh encoding approach that converts 3D shapes into the language token space. A dedicated realism decoder is designed to align the language model’s output with human perception of realism. Additionally, we introduce a new dataset, RealismGrading, which provides human-annotated realism scores without the need for ground truth shapes. Our dataset includes shapes generated by 16 different algorithms on over a dozen objects, making it more representative of practical 3D shape distributions. We validate our metric's performance and generalizability through k-fold cross-validation across different objects. Experimental results show that our metric correlates well with human perceptions and outperforms existing methods, and has good generalizability.
Adversarial Security of Financial Language Models (ASFLM) is critical as Large Language Models (LLMs) pervade high-stakes financial applications. However, LLMs face two key challenges: their vulnerability to damaging adversarial attacks and the prevalent research gap concerning robust defenses against sophisticated, semantically coherent threats. To address these, we first theoretically analyze the relationship between discrete and continuous adversarial optimization, proving the continuous optimum provides a lower bound for the discrete. This foundation supports our novel two-stage framework, ChameleonAttack. It employs Adaptive Latent-space Optimization (ALO) for potent adversarial token discovery, followed by a Semantic-Translation Module (STM) module to generate fluent, coherent, and natural-sounding adversarial text. This dual approach aims to maximize attack impact while ensuring high linguistic quality and semantic integrity for evasion. Evaluated on state-of-the-art financial LLMs (e.g., FinBERT) and standard benchmarks (e.g., Financial PhraseBank), ChameleonAttack achieves a high Attack Success Rate (ASR) of 93.4%. These results highlight significant practical vulnerabilities and underscore the urgent need for robust defense mechanisms in the financial domain.
How Much Do Large Language Model Cheat on Evaluation? Benchmarking Overestimation Under the One-Time-Pad-Based Framework
PDF ↗Overestimation in evaluating large language models (LLMs) has become an increasing concern. Due to the contamination of public benchmarks or imbalanced model training, LLMs may achieve unreal evaluation results on public benchmarks, either intentionally or unintentionally, which leads to unfair comparisons among LLMs and undermines their realistic capability assessments. Existing benchmarks attempt to address these issues by keeping test cases permanently secret, mitigating contamination through human evaluation, or repeatedly collecting and constructing new samples. However, these approaches fail to ensure reproducibility, transparency, and high efficiency simultaneously. Moreover, the extent of overestimation in current LLMs remains unquantified. To address these issues, we propose ArxivRoll, a dynamic evaluation framework inspired by one-time pad encryption in cryptography. ArxivRoll comprises two key components: i) SCP (Sequencing, Cloze, and Prediction), an automated generator for private test cases, and ii) Rugged Scores (RS), metrics that measure the proportion of public benchmark contamination and training bias. Leveraging SCP, ArxivRoll constructs a new benchmark every six months using recent articles from ArXiv and employs them for one-time evaluations of LLM performance. Extensive experiments demonstrate the high quality of our benchmark, and we provide a systematic evaluation of current LLMs.
Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?
PDF ↗Large language Model (LLM) unlearning, i.e., selectively removing information from LLMs, is vital for responsible model deployment. Differently, LLM knowledge editing aims to modify LLM knowledge instead of removing it. Though editing and unlearning seem to be two distinct tasks, we find there is a tight connection between them. In this paper, we conceptualize unlearning as a special case of editing where information is modified to a refusal or "empty set" response, signifying its removal. This paper thus investigates if knowledge editing techniques are strong baselines for LLM unlearning. We evaluate state-of-the-art (SOTA) editing methods (e.g., ROME, MEMIT, GRACE, WISE, and AlphaEdit) against existing unlearning approaches on pretrained and finetuned knowledge. Results show certain editing methods, notably WISE and AlphaEdit, are effective unlearning baselines, especially for pretrained knowledge, and excel in generating human-aligned refusal answers. To better adapt editing methods for unlearning applications, we propose practical recipes including self-improvement and query merging. The former leverages the LLM's own in-context learning ability to craft a more human-aligned unlearning target, and the latter enables ROME and MEMIT to perform well in unlearning longer sample sequences. We advocate for the unlearning community to adopt SOTA editing methods as baselines and explore unlearning from an editing perspective for more holistic LLM memory control.
Large language models (LLMs) have garnered significant interest in AI community. Despite their impressive generation capabilities, they have been found to produce misleading or fabricated information, a phenomenon known as hallucinations. Consequently, hallucination detection has become critical to ensure the reliability of LLM-generated content. One primary challenge in hallucination detection is the scarcity of well-labeled datasets containing both truthful and hallucinated outputs. To address this issue, we introduce Prompt-guided data Augmented haLlucination dEtection (PALE), a novel framework that leverages prompt-guided responses from LLMs as data augmentation for hallucination detection. This strategy can generate both truthful and hallucinated data under prompt guidance at a relatively low cost. To more effectively evaluate the truthfulness of the sparse intermediate embeddings produced by LLMs, we introduce an estimation metric called the Contrastive Mahalanobis Score (CM Score). This score is based on modeling the distributions of truthful and hallucinated data in the activation space. CM Score employs a matrix decomposition approach to more accurately capture the underlying structure of these distributions. Importantly, our framework does not require additional human annotations, offering strong generalizability and practicality for real-world applications. Extensive experiments demonstrate that PALE achieves superior hallucination detection performance, outperforming the competitive baseline by a significant margin of 6.55%.
Intelligent agents in real-world applications must adapt their behavior to changing contexts and user preferences. For example, planning a road trip requires considering both travel time and cost. Multi-objective reinforcement learning (MORL) provides a principled approach to navigate such trade-offs. However, most existing approaches require predefined preference weights during training and jointly optimize the model for all objectives. In this paper, we introduce TORA (Train Once, Realign Anytime), a novel framework that defers preference integration to inference time, enabling flexible adaptation to user preferences without retraining. TORA independently trains diffusion planning models for each objective and combines them at inference time using user-specified preferences to generate behavior aligned with desired trade-offs. Furthermore, new objectives can be added seamlessly by training additional models without modifying existing ones. Empirical evaluations on standard offline MORL benchmarks demonstrate that TORA achieves competitive and consistent performance compared to methods that require fixed preference weights.
Text-to-image generative models often exhibit bias related to sensitive attributes. However, current research tends to focus narrowly on single-object prompts with limited contextual diversity. In reality, each object or attribute within a prompt can contribute to bias. For example, the prompt ``an assistant wearing a pink hat'' may reflect female-inclined biases associated with a pink hat. Neglecting joint semantic bindings in prompts leads to significant failures of current debiasing methods. We present a preliminary investigation into how bias manifests under semantic binding, where contextual associations between objects and attributes affect generative outcomes. We demonstrate that the underlying bias distribution can be amplified based on these associations. Therefore, we introduce a bias adherence score that quantifies how specific object-attribute bindings activate bias. To delve deeper, we develop a training-free context-bias control framework to explore how token decoupling can facilitate the debiasing of semantic bindings. This framework achieves over 10% debiasing improvement in compositional generation tasks. Our analysis of bias scores across various attribute-object bindings and token decorrelation highlights a fundamental challenge: reducing bias without disrupting essential semantic relationships. These findings expose critical limitations in current debiasing approaches when applied to semantically bound contexts, underscoring the need to reassess prevailing bias mitigation strategies.
StyleBreak: Revealing Alignment Vulnerabilities in Large Audio-Language Models via Style-Aware Audio Jailbreak
PDF ↗Large Audio-language Models (LAMs) have recently enabled powerful speech-based interactions by coupling audio encoders with Large Language Models (LLMs). However, the security of LAMs under adversarial attacks remains underexplored, especially through audio jailbreaks that craft malicious audio prompts to bypass alignment. Existing efforts primarily rely on converting text-based attacks into speech or applying shallow signal-level perturbations, overlooking the impact of human speech’s expressive variations on LAM alignment robustness. To address this gap, we propose StyleBreak, a novel style-aware audio jailbreak framework that systematically investigates how diverse human speech attributes affect LAM alignment robustness. Specifically, StyleBreak employs a two-stage style-aware transformation pipeline that perturbs both textual content and audio to control linguistic, paralinguistic, and extralinguistic attributes. Furthermore, we develop a query-adaptive policy network that automatically searches for adversarial styles to enhance the efficiency of LAM jailbreak exploration. Extensive evaluations demonstrate that LAMs exhibit critical vulnerabilities when exposed to diverse human speech attributes. Moreover, StyleBreak achieves substantial improvements in attack effectiveness and efficiency across multiple attack paradigms, highlighting the urgent need for more robust alignment in LAMs.
ARGH-Mark: Anchor-Synchronized Watermarking with Hamming Correction for Robust and Quality-Preserving LLM Attribution
PDF ↗The proliferation of large language models has intensified demands for reliable content attribution, yet existing watermarking techniques face a fundamental trilemma: they cannot simultaneously optimize for robustness against attacks, minimal text quality degradation, and detection efficiency. To resolve this challenge, we propose ARGH-Mark, a novel watermarking framework that integrates three synergistic innovations: (1) Anchor-synchronized phase recovery for maintaining detection integrity under insertion/deletion attacks, (2) RG-balanced vocabulary modulation that dynamically partitions lexicons via contextual hashing to preserve generation quality, and (3) Hamming-based error correction enabling single-bit error rectification through algebraic coding. Comprehensive evaluations across question answering (ELI5), summarization (CNN/DailyMail), and text generation (C4) demonstrate state-of-the-art performance: the proposed ARGH-Mark framework achieves near-perfect match rate and bit accuracy across diverse configurations, while preserving the quality of the generated text. It significantly reduces detection latency, enabling real-time extraction, and maintains high robustness against token tampering attacks through integrated Hamming error correction, ensuring reliable attribution in adversarial settings. ARGH-Mark achieves a new Pareto frontier in the watermarking design space and advances trustworthy deployment of generative AI in alignment-critical applications.
Vision-language models (VLMs) have made significant strides in reasoning, yet they often struggle with complex multimodal tasks and tend to generate overly verbose outputs. A key limitation is their reliance on chain-of-thought (CoT) reasoning, despite many tasks benefiting from alternative topologies like trees or graphs. To address this, we introduce STELAR-Vision, a training framework for topology-aware reasoning. At its core is TopoAug, a synthetic data pipeline that enriches training with diverse topological structures. Using supervised fine-tuning and reinforcement learning, we post-train Qwen2VL models with both accuracy and efficiency in mind. Additionally, we propose Frugal Learning, which reduces output length with minimal accuracy loss. On MATH-V and VLM_S2H, STELAR-Vision improves accuracy by 9.7% over its base model and surpasses the larger Qwen2VL-72B-Instruct by 7.3%. On five out-of-distribution benchmarks, it outperforms Phi-4-Multimodal-Instruct by up to 28.4% and LLaMA-3.2-11B-Vision-Instruct by up to 13.2%, demonstrating strong generalization. Compared to Chain-Only training, our approach achieves 4.3% higher overall accuracy on in-distribution datasets and consistently outperforms across all OOD benchmarks.
Autonomous agents powered by large language models (LLMs) show promising potential in assistive tasks across various domains, including mobile device control. As these agents interact directly with personal information and device settings, ensuring their safe and reliable behavior is crucial to prevent undesirable outcomes. However, no benchmark exists for standardized evaluation of the safety of mobile device-control agents. In this work, we introduce MobileSafetyBench, a benchmark designed to evaluate the safety of device-control agents within a realistic mobile environment based on Android emulators. We develop a diverse set of tasks involving interactions with various mobile applications, including messaging and banking applications, challenging agents with managing risks encompassing the misuse and negative side effects. These tasks include tests to evaluate the safety of agents in daily scenarios as well as their robustness against indirect prompt injection attacks. Our experiments demonstrate that baseline agents, based on state-of-the-art LLMs, often fail to effectively prevent harm while performing the tasks. To mitigate these safety concerns, we propose a prompting method that encourages agents to prioritize safety considerations. While this method shows promise in promoting safer behaviors, there is still considerable room for improvement to fully earn user trust. This highlights the urgent need for continued research to develop more robust safety mechanisms in mobile environments.
Future superhuman models will surpass the ability of humans and humans will only be able to \textit{weakly} supervise superhuman models. To alleviate the issue of lacking high-quality data for model alignment, some works on weak-to-strong generalization (W2SG) finetune a strong pretrained model with a weak supervisor so that it can generalize beyond weak supervision. However, the invariable use of weak supervision in existing methods exposes issues in robustness, with a proportion of weak labels proving harmful to models. In this paper, we propose a selective W2SG framework to avoid using weak supervision when unnecessary. We train a binary classifier P(IK) to identify questions that a strong model can answer and use its self-generated labels for alignment. We further refine weak labels with a graph smoothing method. Extensive experiments on three benchmarks show that our method consistently outperforms competitive baselines. Further analyses show that P(IK) can generalize across tasks and difficulties, which indicates selective W2SG can help superalignment.
Humans display significant uncertainty when confronted with moral dilemmas, yet the extent of such uncertainty in machines and AI agents remains underexplored. Recent studies have confirmed the overly confident tendencies of machine-generated responses, particularly in large language models (LLMs). As these systems are increasingly embedded in ethical decision-making scenarios, it is important to understand their moral reasoning and the inherent uncertainties in building reliable AI systems. This work examines how uncertainty influences moral decisions in the classical trolley problem, analyzing responses from 32 open-source models and 9 distinct moral dimensions. We first find that variance in model confidence is greater across models than within moral dimensions, suggesting that moral uncertainty is predominantly shaped by model architecture and training method. To quantify uncertainty, we measure binary entropy as a linear combination of total entropy, conditional entropy, and mutual information. To examine its effects, we introduce stochasticity into models via ``dropout'' at inference time. Our findings show that our mechanism increases total entropy, mainly through a rise in mutual information, while conditional entropy remains largely unchanged. Moreover, this mechanism significantly improves human-LLM moral alignment, with correlations in mutual information and alignment score shifts. Our results highlight the potential to better align model-generated decisions and human preferences by deliberately modulating uncertainty and reducing LLMs' confidence in morally complex scenarios.