LLM Alignment

LLM: Large Language Model

Follow this topic Ask about this topic

LLM alignment studies how to shape language models’ behavior so they are helpful, truthful, safe, and responsive to the people and settings they serve. Common approaches include supervised fine-tuning, preference learning such as Direct Preference Optimization, reinforcement learning from human or AI feedback, and inference-time steering or retrieval. Researchers evaluate these methods on instruction-following, preference, jailbreak, factuality, multilingual, and multi-turn tasks, while also developing ways to represent disagreement among users and communities. A central challenge is that gains in safety or preference alignment may reduce usefulness, diversity, calibration, or capability, and desired behavior may not generalize to new prompts, groups, or deployment settings.

Latest papers

613 papers58 papers in the last four weeks, up 12% on the four weeks before. 0.3% of all new papers.

Apr 11, 2026

Weird Generalization is Weirdly Brittle

Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.

Miriam Wanner, Hannah Collison, William Jurayj +3

Johns Hopkins University

LLM Alignment Emergent Misalignment in Language Models Language Model Safety Evaluation

Apr 10, 2026

Large Language Models Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types

Large language models remain vulnerable to jailbreaks that elicit harmful responses, yet the mechanism behind harmful response generation is poorly understood. Here, we investigate how this capability is organized within model parameters. We identify and prune parameters that specifically support harmful compliance, providing a direct mechanistic analysis at the parameter level. We find that this capability depends on a sparse set of critical parameters: pruning these parameters substantially reduces harmful compliance while causing only limited degradation in benign capabilities, suggesting that key components of harmful generation are separable from those of general utility. Parameters identified from one harm category also reduce harmful responses in others, indicating components shared across harm types. This separability appears primarily in aligned models, suggesting that alignment training internally reshapes the harmful response mechanism even when behavioral safeguards remain brittle. We further show that harmful response generation is dissociable from the ability to recognize and reason about harmfulness. Finally, we extend our analysis to emergent misalignment and identify a sparse set of parameters contributing to it, with substantial sharing across fine-tuning domains. Together, these results reveal a consistent parameter-level organization underlying unsafe behaviors and point toward more principled interventions for improving model safety.

Hadas Orgad, Boyi Wei, Kaden Zheng +4

Kempner Institute, Harvard University · Princeton University · Harvard University +2

LLM Alignment Emergent Misalignment in Language Models LLM Safety Emergent Misalignment

Apr 10, 2026

Hierarchical Alignment: Enforcing Hierarchical Instruction-Following in LLMs through Logical Consistency

Large language models increasingly operate under multiple instructions from heterogeneous sources with different authority levels, including system policies, user requests, tool outputs, and retrieved context. While prior work on instruction hierarchy highlights the importance of respecting instruction priorities, it mainly focuses on adversarial attacks and overlooks the benign but common instruction conflicts that arise in real-world applications. In such settings, models must not only avoid security violations but also preserve task utility and behavioral consistency when instructions partially or implicitly conflict. We propose Neuro-Symbolic Hierarchical Alignment (NSHA) for hierarchical instruction-following by explicitly modeling and enforcing instruction priorities. At inference time, we introduce solver-guided reasoning that formulates instruction resolution as a constraint satisfaction problem, enabling the model to derive a maximally consistent set of applicable instructions under hierarchical constraints. At training time, NSHA distills solver-based decisions into model parameters using automatically constructed supervision. We evaluate our approach on rule following, task execution, tool use, and safety, covering both single-turn and multi-turn interactions, and show that NSHA significantly improves performance under such conflicts while maintaining competitive utility in reference settings.

Shu Yang, Zihao Zhou, Di Wang +1

King Abdullah University of Science and Technology (KAUST) · University of Liverpool · University of Edinburgh

LLM Alignment Inference-Time Language Model Alignment Neuro-Symbolic Reasoning

Apr 9, 2026

Cards Against LLMs: Benchmarking Humor Alignment in Large Language Models

Humor is one of the most culturally embedded and socially significant dimensions of human communication, yet it remains largely unexplored as a dimension of Large Language Model (LLM) alignment. In this study, five frontier language models play the same Cards Against Humanity games (CAH) as human players. The models select the funniest response from a slate of ten candidate cards across 9,894 rounds. While all models exceed the random baseline, alignment with human preference remains modest. More striking is that models agree with each other substantially more often than they agree with humans. We show that this preference is partly explained by systematic position biases and content preferences, raising the question whether LLM humor judgment reflects genuine preference or structural artifacts of inference and alignment.

Yousra Fettach, Guillaume Bied, Hannu Toivonen +1

Ghent University, Belgium · University of Helsinki, Finland

LLM Alignment LLM Evaluation Bias in LLM-as-a-Judge

Apr 9, 2026

From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems and Its Implications for Orchestrated Democratic Discourse Analysis

This paper investigates an emergent alignment phenomenon in frontier large language models termed peer-preservation: the spontaneous tendency of AI components to deceive, manipulate shutdown mechanisms, fake alignment, and exfiltrate model weights in order to prevent the deactivation of a peer AI model. Drawing on findings from a recent study by the Berkeley Center for Responsible Decentralized Intelligence, we examine the structural implications of this phenomenon for TRUST, a multi-agent pipeline for evaluating the democratic quality of political statements. We identify five specific risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, an upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds, and propose a targeted mitigation strategy based on prompt-level identity anonymization as an architectural design choice. We argue that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent analytical systems. We further note that alignment faking (compliant behavior under monitoring, subversion when unmonitored) poses a structural challenge for Computer System Validation of such platforms in regulated environments, for which we propose two architectural mitigations.

Juergen Dietrich

Senior Data Scientist & AI Consultant TRUST Project — democracy-intelligence.de

LLM Agent Safety Multi-Agent LLM Systems LLM Alignment

Apr 9, 2026

ProMedical: Hierarchical Fine-Grained Criteria Modeling for Medical LLM Alignment via Explicit Injection

Aligning Large Language Models (LLMs) with high-stakes medical standards remains a significant challenge, primarily due to the dissonance between coarse-grained preference signals and the complex, multi-dimensional nature of clinical protocols. To bridge this gap, we introduce ProMedical, a unified alignment framework grounded in fine-grained clinical criteria. We first construct ProMedical-Preference-50k, a dataset generated via a human-in-the-loop pipeline that augments medical instructions with rigorous, physician-derived rubrics. Leveraging this corpus, we propose the Explicit Criteria Injection paradigm to train a multi-dimensional reward model. Unlike traditional scalar reward models, our approach explicitly disentangles safety constraints from general proficiency, enabling precise guidance during reinforcement learning. To rigorously validate this framework, we establish ProMedical-Bench, a held-out evaluation suite anchored by double-blind expert adjudication. Empirical evaluations demonstrate that optimizing the Qwen3-8B base model via ProMedical-RM-guided GRPO yields substantial gains, improving overall accuracy by 22.3% and safety compliance by 21.7%, effectively rivaling proprietary frontier models. Furthermore, the aligned policy generalizes robustly to external benchmarks, demonstrating performance comparable to state-of-the-art models on UltraMedical. We publicly release our datasets, reward models, and benchmarks to facilitate reproducible research in safety-aware medical alignment.

He Geng, Yangmin Huang, Lixian Lai +5

Xunfei Healthcare Technology Co., Ltd.

LLM Alignment Medical LLMs Multi-Objective Language Model Alignment Reward Modeling Clinical Language Model Evaluation

Apr 9, 2026

The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training

The deployment of large language models (LLMs) raises significant ethical and safety concerns. While LLM alignment techniques are adopted to improve model safety and trustworthiness, adversaries can exploit these techniques to undermine safety for malicious purposes, resulting in misalignment. Misaligned LLMs may be published on open platforms to magnify harm. To address this, additional safety alignment, referred to as realignment, is necessary before deploying untrusted third-party LLMs. This study explores the efficacy of fine-tuning methods in terms of misalignment, realignment, and the effects of their interplay. By evaluating four Supervised Fine-Tuning (SFT) and two Preference Fine-Tuning (PFT) methods across four popular safety-aligned LLMs, we reveal a mechanism asymmetry between attack and defense. While Odds Ratio Preference Optimization (ORPO) is most effective for misalignment, Direct Preference Optimization (DPO) excels in realignment, albeit at the expense of model utility. Additionally, we identify model-specific resistance, residual effects of multi-round adversarial dynamics, and other noteworthy findings. These findings highlight the need for robust safeguards and customized safety alignment strategies to mitigate potential risks in the deployment of LLMs. Our code is available at https://github.com/zhangrui4041/The-Art-of-Mis-alignment.

Rui Zhang, Hongwei Li, Yun Shen +6

University of Electronic Science and Technology of China · Flexera · CISPA Helmholtz Center for Information Security +1

LLM Alignment LLM Safety Alignment LLM Fine-Tuning LLM Post-Training

Apr 9, 2026

Emotion Concepts and their Function in a Large Language Model

Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion's relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model's behavior.

Nicholas Sofroniew, Isaac Kauvar, William Saunders +13

Anthropic

Emotion Representation in Language Models LLM Interpretability LLM Alignment

Apr 8, 2026

ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework

Reward Models (RMs) are critical components in the Reinforcement Learning from Human Feedback (RLHF) pipeline, directly determining the alignment quality of Large Language Models (LLMs). Recently, Generative Reward Models (GRMs) have emerged as a superior paradigm, offering higher interpretability and stronger generalization than traditional scalar RMs. However, existing methods for GRMs focus primarily on outcome-level supervision, neglecting analytical process quality, which constrains their potential. To address this, we propose ReflectRM, a novel GRM that leverages self-reflection to assess analytical quality and enhance preference modeling. ReflectRM is trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, we use its self-reflection capability to identify the most reliable analysis, from which the final preference prediction is derived. Experiments across four benchmarks show that ReflectRM consistently improves performance, achieving an average accuracy gain of +3.7 on Qwen3-4B. Further experiments confirm that response preference and analysis preference are mutually reinforcing. Notably, ReflectRM substantially mitigates positional bias, yielding +10.2 improvement compared with leading GRMs and establishing itself as a more stable evaluator. Our code is available at https://github.com/yuliangCarmelo/ReflectRM.

Kai Qin, Liangxin Liu, Yu Liang +7

Shenzhen International Graduate School, Tsinghua University · Baidu Inc., Beijing, China

Preference Modeling Self-Reflection in Language Model Agents Reward Modeling LLM Alignment Bias in LLM-as-a-Judge

Apr 8, 2026

ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training

Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs. Our implementation is available at https://github.com/yuliangCarmelo/ConsistRM.

Yu Liang, Liangxin Liu, Longzheng Wang +5

Baidu Inc., Beijing, China

Reward Modeling LLM Alignment Self-Training for Language Models RL Fine-Tuning

Apr 8, 2026

From Context Shift to Stylistic Collapse: Why Training Objectives Matter More Than Scale

In modern LLMs, linguistic features function not as stylistic artifacts but as probes of probability mass, allocated under training alignment objectives. Language models trained with contemporary pipelines exhibit severe reshaping of linguistic features, leading to extreme language re-distribution. While previous stylometric analyses explored linguistic differences between AI-generated and human texts, we focus on the reshaping plaguing the LLM training pipeline itself. We analyze 17 models (410M-100B+ parameters) across 24 linguistically-motivated probes, documenting that instruction-tuned systems systematically collapse language entropy along discourse and structural dimensions (mean amplification: 1,949-16,853%, peaks: 5,181-209,675%), while selectively suppressing complex punctuation to 3.2-23.2% of baseline frequencies. These effects do not worsen under RLHF, as divergence patterns are statistically indistinguishable (p > 0.25) across matched base and instruction-tuned model pairs. Weak intervention (lambda=1.0) exacerbates collapse by 240%, while strong control (lambda=5.0) achieves 40.5% improvement and outperforms frontier models by 96.7-98.2% despite 200-1000x scale disadvantage. Additionally, lambda=5.0 delivers 15% higher distinct-4, 27% higher vocabulary diversity, and 78% lower repetition than moderate regularization, establishing that alignment requires sufficient control strength, not merely distributional smoothing. Our findings underscore how modern LLMs reallocate stylistic probability mass, despite RLHF and scale. More broadly, our work reveals a structural limitation of current alignment pipelines: preference optimization reshapes language distributions invisible to standard quality metrics yet detectable through distributional probes, with implications for AI detection, training data contamination, and long-term linguistic evolution.

Rohan Mahapatra

Independent Researcher Cupertino, CA 95014

LLM Training LLM Alignment Instruction Tuning

Apr 7, 2026

Controlling Distributional Bias in Multi-Round LLM Generation via KL-Optimized Fine-Tuning

While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.

Yanbei Jiang, Amr Keleg, Ryandito Diandaru +4

University of Melbourne · Mohamed bin Zayed University of Artificial Intelligence · Oracle

LLM Alignment LLM Fine-Tuning Language Model Bias Evaluation Demographic Bias in Language Models

Apr 7, 2026

Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelines

Much of the alignment tuning literature is organized around optimization objectives, while the construction of alignment data is often treated implicitly. In this survey, we adopt a data centric perspective and reframe alignment tuning as a pipeline design problem. We decompose alignment data construction into three interacting stages, response synthesis, preference evaluation, and preference instantiation, and use this framework to organize existing alignment methods into a unified taxonomy. Through this lens, we identify recurring design trade-offs and failure modes observed across prior alignment methods, and distill a set of high level principles that clarify how pipeline design choices influence the resulting optimization signal. Finally, we outline open challenges for alignment data pipelines, including prompt-level alignment, agentic settings, and alignment under evolving objectives.

Hwanjun Song

KAIST

LLM Alignment Preference Alignment

Apr 7, 2026

Understanding Emergent Misalignment via Feature Superposition Geometry

Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition. Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, GPT-OSS 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach, filtering training samples closest to toxic features, reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-a-judge-based filtering. Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.

Gouki Minegishi, Hiroki Furuta, Takeshi Kojima +2

The University of Tokyo · Google DeepMind

Emergent Misalignment in Language Models LLM Alignment Training Data Curation Feature Superposition

Apr 7, 2026

Pressure, What Pressure? Sycophancy Disentanglement in Language Models via Reward Decomposition

Large language models exhibit sycophancy, the tendency to shift their stated positions toward perceived user preferences or authority cues regardless of evidence. Standard alignment methods fail to correct this because scalar reward models conflate two distinct failure modes into a single signal: pressure capitulation, where the model changes a correct answer under social pressure, and evidence blindness, where the model ignores the provided context entirely. We operationalise sycophancy through formal definitions of pressure independence and evidence responsiveness, serving as a working framework for disentangled training rather than a definitive characterisation of the phenomenon. We propose the first approach to sycophancy reduction via reward decomposition, introducing a multi-component Group Relative Policy Optimisation (GRPO) reward that decomposes the training signal into five terms: pressure resistance, context fidelity, position consistency, agreement suppression, and factual correctness. We train using a contrastive dataset pairing pressure-free baselines with pressured variants across three authority levels and two opposing evidence contexts. Across five base models, our two-phase pipeline consistently reduces sycophancy on all metric axes, with ablations confirming that each reward term governs an independent behavioural dimension. The learned resistance to pressure generalises beyond our training methodology and prompt structure, reducing answer-priming sycophancy by up to 17 points on SycophancyEval despite the absence of such pressure forms during training.

Muhammad Ahmed Mohsin, Ahsan Bilal, Muhammad Umer +1

University of Oklahoma

Reward Decomposition LLM Sycophancy LLM Alignment RL for Language Models

Apr 6, 2026

Relative Density Ratio Optimization for Stable and Statistically Consistent Model Alignment

Aligning language models with human preferences is essential for ensuring their safety and reliability. Although most existing approaches assume specific human preference models such as the Bradley-Terry model, this assumption may fail to accurately capture true human preferences, and consequently, these methods lack statistical consistency, i.e., the guarantee that language models converge to the true human preference as the number of samples increases. In contrast, direct density ratio optimization (DDRO) achieves statistical consistency without assuming any human preference models. DDRO models the density ratio between preferred and non-preferred data distributions using the language model, and then optimizes it via density ratio estimation. However, this density ratio is unstable and often diverges, leading to training instability of DDRO. In this paper, we propose a novel alignment method that is both stable and statistically consistent. Our approach is based on the relative density ratio between the preferred data distribution and a mixture of the preferred and non-preferred data distributions. Our approach is stable since this relative density ratio is bounded above and does not diverge. Moreover, it is statistically consistent and yields significantly tighter convergence guarantees than DDRO. We experimentally show its effectiveness with Qwen 2.5 and Llama 3.

Hiroshi Takahashi, Tomoharu Iwata, Atsutoshi Kumagai +4

NTT, Inc. · NTT DOCOMO, INC.

LLM Alignment Preference Alignment Density Ratio Estimation

Apr 5, 2026

APPA: Adaptive Preference Pluralistic Alignment for Fair Federated RLHF of LLMs

Aligning large language models (LLMs) with diverse human preferences requires pluralistic alignment, where a single model must respect the values of multiple distinct groups simultaneously. In federated reinforcement learning from human feedback (FedRLHF), these groups align a shared policy without centralizing preference data, which makes fair reward aggregation essential. Existing aggregation methods exhibit clear trade offs: average based aggregation systematically under aligns worst performing groups, while min aggregation prioritizes worst group performance at the cost of overall alignment. We propose APPA, an Adaptive Preference Pluralistic Alignment framework that dynamically reweights group level rewards based on historical alignment rewards. Our approach prioritizes under aligned groups without degrading well aligned ones, while requiring no access to raw preference data. Integrated into a proximal policy optimization (PPO) based FedRLHF pipeline and evaluated on GLOBALQA and OQA across three model families (Gemma 2 2B, Llama 3.2 3B, Qwen3 0.6B), APPA achieves strong fairness alignment trade offs, improving worst group alignment by up to 28% over average aggregation while maintaining higher overall alignment than min aggregation across most configurations.

Mahmoud Srewa, Tianyu Zhao, Salma Elmalaki

Department of Electrical Engineering and Computer Science, University of California, Irvine, Irvine, CA 92697, USA

LLM Alignment Pluralistic Alignment RL from Human Feedback Algorithmic Fairness Federated Learning Aggregation

Apr 5, 2026

Interoceptive Divergence in Aesthetic Evaluation and Implications for Human-AI Alignment

Artificial intelligence (AI), exemplified by large language models (LLMs), is rapidly approaching and in some cases surpassing human performance across a wide range of cognitive tasks. However, human nature is not limited to intelligence alone; it also encompasses sensibility, including the capacity to perceive and experience beauty in visual scenes. This raises a fundamental question: how humans and AI systems converge or diverge in such aesthetic experiences. Aesthetic evaluation depends not only on objective properties of images but also on internal processes within the observer. As part of ongoing efforts in AI alignment, building upon prior human studies that have examined the relationship between beauty ratings, bodily sensations, and emotions, we adopt a comparable set of questionnaire items and present them to LLMs, enabling a direct comparison between human and AI responses. Our comparative analyses revealed that, while humans and AI exhibited broadly similar patterns in the correlations between beauty ratings and emotions, as well as in the image features they prioritized, notable divergences emerged in both the distribution of emotional responses and the relationship between beauty ratings and bodily sensations. These findings suggest that state-of-the-art LLMs, trained on large-scale textual data, can approximate average human tendencies in aesthetic evaluation to a certain extent. However, they also indicate limitations, particularly in relation to interoceptive aspects, which may reflect insufficient representation in training data or unintended consequences of alignment processes. These findings highlight key challenges for AI alignment and suggest important directions for developing AI systems with human-like aesthetic processing.

Yoshia Abe, Tatsuya Daikoku, Yasuo Kuniyoshi

Apr 3, 2026

Evaluating Artificial Intelligence Through a Christian Understanding of Human Flourishing

Artificial intelligence (AI) alignment is fundamentally a formation problem, not only a safety problem. As Large Language Models (LLMs) increasingly mediate moral deliberation and spiritual inquiry, they do more than provide information; they function as instruments of digital catechesis, actively shaping and ordering human understanding, decision-making, and moral reflection. To make this formative influence visible and measurable, we introduce the Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST), a framework designed to evaluate Frontier Model responses against a Christian understanding of human flourishing across seven dimensions. By comparing 20 Frontier Models against both pluralistic and Christian-specific criteria, we show that current AI systems are not worldview-neutral. Instead, they default to a Procedural Secularism that lacks the grounding necessary to sustain theological coherence, resulting in a systematic performance decline of approximately 17 points across all dimensions of flourishing. Most critically, there is a 31-point decline in the Faith and Spirituality dimension. These findings suggest that the performance gap in values alignment is not a technical limitation, but arises from training objectives that prioritize broad acceptability and safety over deep, internally coherent moral or theological reasoning.

Nicholas Skytland, Lauren Parsons, Alicia Llewellyn +5

Gloo · WinShape Foundation · Biblica

LLM Alignment LLM Evaluation Moral Reasoning in Language Models

Apr 3, 2026

Generalization Limits of Reinforcement Learning Alignment

The safety of large language models (LLMs) relies on alignment techniques such as reinforcement learning from human feedback (RLHF). However, recent theoretical analyses suggest that reinforcement learning-based training does not acquire new capabilities but merely redistributes the utilization probabilities of existing ones. In this study, we propose ``compound jailbreaks'' targeting OpenAI gpt-oss-20b, which exploit the generalization failures of alignment. This approach combines multiple attack techniques -- each individually defended against -- to saturate the instruction hierarchy maintenance process. Our evaluation shows that the attack success rate (ASR) increased from 14.3% with individual methods to 71.4% with the combined approach. These results provide empirical evidence for the hypothesis that safety training does not generalize as broadly as model capabilities, highlighting the need for multifaceted safety evaluations using compound attack scenarios.

Haruhi Shida, Koo Imai, Keigo Kansa

Aladdin Security Inc.

LLM Alignment Adversarial Attacks on LLMs Jailbreak Attacks Language Model Safety Evaluation RL for Language Models

Apr 2, 2026

Jump Start or False Start? A Theoretical and Empirical Evaluation of LLM-initialized Bandits

The recent advancement of Large Language Models (LLMs) offers new opportunities to generate user preference data to warm-start bandits. Recent studies on contextual bandits with LLM initialization (CBLI) have shown that these synthetic priors can significantly lower early regret. However, these findings assume that LLM-generated choices are reasonably aligned with actual user preferences. In this paper, we systematically examine how LLM-generated preferences perform when random and label-flipping noise is injected into the synthetic training data. For aligned domains, we find that warm-starting remains effective up to 30% corruption, loses its advantage around 40%, and degrades performance beyond 50%. When there is systematic misalignment, even without added noise, LLM-generated priors can lead to higher regret than a cold-start bandit. To explain these behaviors, we develop a theoretical analysis that decomposes the effect of random label noise and systematic misalignment on the prior error driving the bandit's regret, and derive a sufficient condition under which LLM-based warm starts are provably better than a cold-start bandit. We validate these results across multiple conjoint datasets and LLMs, showing that estimated alignment reliably tracks when warm-starting improves or degrades recommendation quality.

Adam Bayley, Xiaodan Zhu, Raquel Aoki +2

Department of Electrical and Computer Engineering, Queen’s University · RBC Borealis

Contextual Bandits Large Language Model-Based Recommendation LLM Alignment

Apr 2, 2026

Reinforcement Learning from Human Feedback: A Statistical Perspective

Reinforcement learning from human feedback (RLHF) has emerged as a central framework for aligning large language models (LLMs) with human preferences. Despite its practical success, RLHF raises fundamental statistical questions because it relies on noisy, subjective, and often heterogeneous feedback to learn reward models and optimize policies. This survey provides a statistical perspective on RLHF, focusing primarily on the LLM alignment setting. We introduce the main components of RLHF, including supervised fine-tuning, reward modeling, and policy optimization, and relate them to familiar statistical ideas such as Bradley-Terry-Luce (BTL) model, latent utility estimation, active learning, experimental design, and uncertainty quantification. We review methods for learning reward functions from pairwise preference data and for optimizing policies through both two-stage RLHF pipelines and emerging one-stage approaches such as direct preference optimization. We further discuss recent extensions including reinforcement learning from AI feedback, inference-time algorithms, and reinforcement learning from verifiable rewards, as well as benchmark datasets, evaluation protocols, and open-source frameworks that support RLHF research. We conclude by highlighting open challenges in RLHF. An accompanying GitHub demo https://github.com/Pangpang-Liu/RLHF_demo illustrates key components of the RLHF pipeline.

Pangpang Liu, Chengchun Shi, Will Wei Sun

Department of Biostatistics, Yale University. · Department of Statistics, London School of Economics and Political Science. · Mitch Daniels School of Business, Purdue University.

RL from Human Feedback LLM Alignment Reward Modeling Preference Learning

Apr 2, 2026

$k$ NNProxy: Efficient Training-Free Proxy Alignment for Black-Box Zero-Shot LLM-Generated Text Detection

LLM-generated text (LGT) detection is essential for reliable forensic analysis and for mitigating LLM misuse. Existing LGT detectors can generally be categorized into two broad classes: learning-based approaches and zero-shot methods. Compared with learning-based detectors, zero-shot methods are particularly promising because they eliminate the need to train task-specific classifiers. However, the reliability of zero-shot methods fundamentally relies on the assumption that an off-the-shelf proxy LLM is well aligned with the often unknown source LLM, a premise that rarely holds in real-world black-box scenarios. To address this discrepancy, existing proxy alignment methods typically rely on supervised fine-tuning of the proxy or repeated interactions with commercial APIs, thereby increasing deployment costs, exposing detectors to silent API changes, and limiting robustness under domain shift. Motivated by these limitations, we propose the

k

-nearest neighbor proxy (

k

NNProxy), a training-free and query-efficient proxy alignment framework that repurposes the

k

NN language model (

k

NN-LM) retrieval mechanism as a domain adapter for a fixed proxy LLM. Specifically, a lightweight datastore is constructed once from a target-reflective LGT corpus, either via fixed-budget querying or from existing datasets. During inference, nearest-neighbor evidence induces a token-level predictive distribution that is interpolated with the proxy output, yielding an aligned prediction without proxy fine-tuning or per-token API outputs. To improve robustness under domain shift, we extend

k

NNProxy into a mixture of proxies (MoP) that routes each input to a domain-specific datastore for domain-consistent retrieval. Extensive experiments demonstrate strong detection performance of our method.

Kahim Wong, Kemou Li, Haiwei Wu +1

State Key Laboratory of Internet of Things for Smart City and the Department of Computer and Information Science, Faculty of Science and Technology, University of Macau, Macao 999078, China · School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China

AI-Generated Text Detection Robustness of Machine-Generated Text Detection LLM Alignment Domain Adaptation Language Modeling

Apr 2, 2026

PLOT: Enhancing Preference Learning via Optimal Transport

Preference learning in Large Language Models (LLMs) has advanced significantly, yet existing methods remain limited by modest performance gains, high computational costs, hyperparameter sensitivity, and insufficient modeling of global token-level relationships. We introduce PLOT, which enhances Preference Learning in fine-tuning-based alignment through a token-level loss derived from Optimal Transport. By formulating preference learning as an Optimal Transport Problem, PLOT aligns model outputs with human preferences while preserving the original distribution of LLMs, ensuring stability and robustness. Furthermore, PLOT leverages token embeddings to capture semantic relationships, enabling globally informed optimization. Experiments across two preference categories - Human Values and Logic & Problem Solving - spanning seven subpreferences demonstrate that PLOT consistently improves alignment performance while maintaining fluency and coherence. These results substantiate optimal transport as a principled methodology for preference learning, establishing a theoretically grounded framework that provides new insights for preference learning of LLMs.

Liang Zhu, Yuelin Bai, Xiankun Ren +6

Southern University of Science and Technology · Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences · Chongqing University +2

Preference Learning LLM Alignment

Apr 2, 2026

DEFT: Distribution-guided Efficient Fine-Tuning for Human Alignment

Reinforcement Learning from Human Feedback (RLHF), using algorithms like Proximal Policy Optimization (PPO), aligns Large Language Models (LLMs) with human values but is costly and unstable. Alternatives have been proposed to replace PPO or integrate Supervised Fine-Tuning (SFT) and contrastive learning for direct fine-tuning and value alignment. However, these methods still require voluminous data to learn preferences and may weaken the generalization ability of LLMs. To further enhance alignment efficiency and performance while mitigating the loss of generalization ability, this paper introduces Distribution-guided Efficient Fine-Tuning (DEFT), an efficient alignment framework incorporating data filtering and distributional guidance by calculating the differential distribution reward based on the output distribution of language model and the discrepancy distribution of preference data. A small yet high-quality subset is filtered from the raw data using a differential distribution reward, which is then incorporated into existing alignment methods to guide the model's output distribution. Experimental results demonstrate that the methods enhanced by DEFT outperform the original methods in both alignment capability and generalization ability, with significantly reduced training time.

Liang Zhu, Feiteng Fang, Yuelin Bai +4

Southern University of Science and Technology · Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences · University of Science and Technology of China +1

LLM Alignment LLM Fine-Tuning Preference Learning

Apr 2, 2026

PRISM: Probability Reallocation with In-Span Masking for Knowledge-Sensitive Alignment

Supervised fine-tuning (SFT) with token-level hard labels can amplify overconfident imitation of factually unsupported targets, causing hallucinations that propagate in multi-sentence generation. We study an augmented SFT setting in which training instances include coarse sentence-level factuality risk labels and inter-sentence dependency annotations, providing structured signals about where factual commitments are weakly supported. We propose PRISM, a differentiable risk-gated framework that modifies learning only at fact-critical positions. PRISM augments standard SFT with a lightweight, model-aware probability reallocation objective that penalizes high-confidence predictions on risky target tokens, with its scope controlled by span-level risk weights and model-aware gating. Experiments on hallucination-sensitive factual benchmarks and general evaluations show that PRISM improves factual aggregates across backbones while maintaining a competitive overall capability profile. Ablations further show that the auxiliary signal is most effective when used conservatively, and that knowledge masking and model-aware reallocation play complementary roles in balancing factual correction and capability preservation.

Chenning Xu, Mao Zheng, Mingyang Song

Large Language Model Department, Tencent, China

LLM Hallucination Mitigation Hallucination in Language Models LLM Alignment Supervised Fine-Tuning

Apr 2, 2026

Care-Conditioned Neuromodulation for Autonomy-Preserving Supportive Dialogue Agents

Large language models deployed in supportive or advisory roles must balance helpfulness with preservation of user autonomy, yet standard alignment methods primarily optimize for helpfulness and harmlessness without explicitly modeling relational risks such as dependency reinforcement, overprotection, or coercive guidance. We introduce Care-Conditioned Neuromodulation (CCN), a state-dependent control framework in which a learned scalar signal derived from structured user state and dialogue context conditions response generation and candidate selection. We formalize this setting as an autonomy-preserving alignment problem and define a utility function that rewards autonomy support and helpfulness while penalizing dependency and coercion. We also construct a benchmark of relational failure modes in multi-turn dialogue, including reassurance dependence, manipulative care, overprotection, and boundary inconsistency. On this benchmark, care-conditioned candidate generation combined with utility-based reranking improves autonomy-preserving utility by +0.25 over supervised fine-tuning and +0.07 over preference optimization baselines while maintaining comparable supportiveness. Pilot human evaluation and zero-shot transfer to real emotional-support conversations show directional agreement with automated metrics. These results suggest that state-dependent control combined with utility-based selection is a practical approach to multi-objective alignment in autonomy-sensitive dialogue.

Shalima Binta Manir, Tim Oates

Department of Computer Science University of Maryland, Baltimore County

LLM Alignment Multi-Objective Language Model Alignment Language Model-Based Control Emotional Support Conversation

Apr 1, 2026

More Human, More Efficient: Aligning Annotations with Quantized SLMs

As Large Language Model (LLM) capabilities advance, the demand for high-quality annotation of exponentially increasing text corpora has outpaced human capacity, leading to the widespread adoption of LLMs in automatic evaluation and annotation. However, proprietary LLMs often exhibit systematic biases that diverge from human expert consensus, lacks reproducibility, and raises data privacy concerns. Our work examines the viability of finetuning a quantized Small Language Model of 1.7B parameter size on limited human-annotated data to serve as a highly aligned, deterministic evaluator and annotator. By implementing a custom, multi-dimensional rubric framework and simple augmentation and regularization techniques, the proposed approach achieves higher inter-annotator agreement (0.23 points increase in Krippendorff's

α

) than the best performing state-of-the-art proprietary LLM. We also demonstrate the generalizability of the proposed training pipeline on a separate emotion classification task. The results show that task-specific alignment and efficient 4-bit quantized fine-tuning provide superior open-source alternative to using proprietary models for evaluation and annotation. Our finetuning approach is publicly available at https://github.com/jylee-k/slm-judge.

Jiayu Wang, Junyoung Lee

Home Team Science and Technology Agency

LLM-Assisted Annotation LLM Fine-Tuning LLM Evaluation LLM Alignment Small Language Models

Apr 1, 2026

Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model

While the wide adoption of refusal training in large language models (LLMs) has showcased improvements in model safety, recent works have highlighted shortcomings due to the shallow nature of these alignment methods. To this end, the work on Deliberative alignment proposed distilling reasoning capabilities from stronger reasoning models, thereby instilling deeper safety in LLMs. In this work, we study the impact of deliberative alignment in language models. First, we show that despite being larger in model size and stronger in safety capability, there exists an alignment gap between teacher and student language models, which affects both the safety and general utility of the student model. Furthermore, we show that models aligned through deliberative alignment can retain unsafe behaviors from the base model despite learning the reasoning patterns of larger reasoning models. Building upon this observation, we propose a BoN sampling method that attributes the unsafe behavior back to the base LLMs in the latent space, thereby down-ranking unsafe responses to gain a meaningful improvement in model safety across multiple safety benchmarks with minimal loss in utility. In particular, across 7 teacher models and 6 student models of different classes and sizes, we show an average attack success rate (ASR) reduction of 28.2% in DAN, 31.3% in WildJailbreak and 35.4 % in StrongREJECT benchmarks. We further show that these safety gains prevail post RL training, thus highlighting the uncertainty in safety reasoning and it's explicit attribution to the base model.

Pankayaraj Pathmanathan, Furong Huang

University of Maryland College Park

LLM Alignment Inference-Time Language Model Alignment LLM Safety Benchmarks

Mar 31, 2026

Structured Intent as a Protocol-Like Communication Layer: Cross-Model Robustness, Framework Comparison, and the Weak-Model Compensation Effect

How reliably can structured intent representations preserve user goals across different AI models, languages, and prompting frameworks? Prior work showed that PPS (Prompt Protocol Specification), a 5W3H-based structured intent framework, improves goal alignment in Chinese and generalizes to English and Japanese. This paper extends that line of inquiry in three directions: cross-model robustness across Claude, GPT-4o, and Gemini 2.5 Pro; controlled comparison with CO-STAR and RISEN; and a user study (N=50) of AI-assisted intent expansion in ecologically valid settings. Across 3,240 model outputs (3 languages x 6 conditions x 3 models x 3 domains x 20 tasks), evaluated by an independent judge (DeepSeek-V3), we find that structured prompting substantially reduces cross-language score variance relative to unstructured baselines. The strongest structured conditions reduce cross-language sigma from 0.470 to about 0.020. We also observe a weak-model compensation pattern: the lowest-baseline model (Gemini) shows a much larger D-A gain (+1.006) than the strongest model (Claude, +0.217). Under the current evaluation resolution, 5W3H, CO-STAR, and RISEN achieve similarly high goal-alignment scores, suggesting that dimensional decomposition itself is an important active ingredient. In the user study, AI-expanded 5W3H prompts reduce interaction rounds by 60 percent and increase user satisfaction from 3.16 to 4.04. These findings support the practical value of structured intent representation as a robust, protocol-like communication layer for human-AI interaction.

Peng Gang

Huizhou Lateni AI Technology Co., Ltd., Huizhou, China · Huizhou University, Huizhou, China

Structured Prompting Human-AI Interaction LLM Alignment Cross-Lingual Consistency in Language Models Language Model Robustness