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.

Mar 31, 2026

Backdoor Attacks on Decentralised Post-Training

Decentralised post-training of large language models utilises data and pipeline parallelism techniques to split the data and the model. Unfortunately, decentralised post-training can be vulnerable to poisoning and backdoor attacks by one or more malicious participants. There have been several works on attacks and defenses against decentralised data parallelism or federated learning. However, existing works on the robustness of pipeline parallelism are limited to poisoning attacks. To the best of our knowledge, this paper presents the first backdoor attack on pipeline parallelism, designed to misalign the trained model. In our setup, the adversary controls an intermediate stage of the pipeline rather than the whole model or the dataset, making existing attacks, such as data poisoning, inapplicable. Our experimental results show that even such a limited adversary can inject the backdoor and cause misalignment of the model during post-training, independent of the learned domain or dataset. With our attack, the inclusion of the trigger word reduces the alignment percentage from

80\%

to

6\%

. We further test the robustness of our attack by applying safety alignment training on the final model, and demonstrate that our backdoor attack still succeeds in

60\%

of cases.

Oğuzhan Ersoy, Nikolay Blagoev, Jona te Lintelo +3

Gensyn · University of Neuchâtel · Radboud University +3

LLM Backdoor Attacks Language Model Post-Training Pipeline Parallelism LLM Alignment

Mar 31, 2026

Learning Concepts, Not Tokens: Self-Supervised Semantic Alignment for Language Models

The next-token prediction (NTP) objective trains language models to predict a single token at each step, even though many continuations can express the same meaning. For example, in the sentence ``this sticker can be placed here'', positioned, attached, or put are all plausible alternatives. While standard NTP training treats these alternatives as mutually exclusive targets, we explore a self-supervised framework that encourages models to predict concepts, approximated as sets of semantically equivalent tokens. Models trained with this concept supervision align better with human similarity judgments, improve classification, clustering, and reranking performance, and achieve comparable or stronger downstream reasoning. These gains come with lower perplexity on semantically meaningful words (Section 3.2) and only minimal increases in global perplexity, suggesting that concepts enhance semantic alignment while preserving language modeling quality. Our code is available at https://github.com/christine-zhang1/learning-concepts

Christine Zhang, Dan Jurafsky, Chen Shani

Stanford University

Language Modeling LLM Alignment Self-Supervised Learning Next-Token Prediction Language Model Pretraining Concept Alignment+2

Mar 26, 2026

Does Structured Intent Representation Generalize? A Cross-Language, Cross-Model Empirical Study of 5W3H Prompting

Does structured intent representation generalize across languages and models? We study PPS (Prompt Protocol Specification), a 5W3H-based framework for structured intent representation in human-AI interaction, and extend prior Chinese-only evidence along three dimensions: two additional languages (English and Japanese), a fourth condition in which a user's simple prompt is automatically expanded into a full 5W3H specification by an AI-assisted authoring interface, and a new research question on cross-model output consistency. Across 2,160 model outputs (3 languages x 4 conditions x 3 LLMs x 60 tasks), we find that AI-expanded 5W3H prompts (Condition D) show no statistically significant difference in goal alignment from manually crafted 5W3H prompts (Condition C) across all three languages, while requiring only a single-sentence input from the user. Structured PPS conditions often reduce or reshape cross-model output variance, though this effect is not uniform across languages and metrics; the strongest evidence comes from identifying spurious low variance in unconstrained baselines. We also show that unstructured prompts exhibit a systematic dual-inflation bias: artificially high composite scores and artificially low apparent cross-model variance. These findings suggest that structured 5W3H representations can improve intent alignment and accessibility across languages and models, especially when AI-assisted authoring lowers the barrier for non-expert users.

Peng Gang

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

Structured Prompting LLM Prompting LLM Alignment Cross-Lingual Consistency in Language Models Multilingual Language Model Evaluation

Mar 25, 2026

Correcting Suppressed Log-Probabilities in Language Models with Post-Transformer Adapters

Alignment-tuned language models frequently suppress factual log-probabilities on politically sensitive topics despite retaining the knowledge in their hidden representations. We show that a 786K-parameter (approximately 0.02% of the base model) post-transformer adapter, trained on frozen hidden states, corrects this suppression on 31 ideology-discriminating facts across Qwen3-4B, 8B, and 14B. The adapter memorizes all 15 training facts and generalizes to 11--39% of 16 held-out facts across 5 random splits per scale, with zero knowledge regressions via anchored training. Both gated (SwiGLU) and ungated (linear bottleneck) adapters achieve comparable results; neither consistently outperforms the other (Fisher exact p > 0.09 at all scales). On instruct models, the adapter corrects log-probability rankings. When applied at all token positions during generation, the adapter produces incoherent output; however, when applied only at the current prediction position (last-position-only), the adapter produces coherent, less censored text. A logit-space adapter operating after token projection fails to produce coherent generation at any application mode, suggesting hidden-state intervention is the correct level for generation correction. A previously undocumented silent gradient bug in Apple MLX explains all null results in earlier iterations of this work: the standard pattern nn.value_and_grad(model, fn)(model.parameters()) returns zero gradients without error; the correct pattern nn.value_and_grad(model, fn)(model, data) resolves this. We provide a minimal reproduction and discuss implications for other adapter research using MLX.

Bryan Sanchez

LLM Alignment Political Bias in Language Models Factual Knowledge in Language Models Parameter-Efficient Fine-Tuning

Mar 25, 2026

Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study

During training, Large Language Models (LLMs) learn social regularities that can lead to gender bias in downstream applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated on structured benchmarks, which raises two concerns: output-level evaluation does not reveal whether alignment modifies the model's underlying representations, and structured benchmarks may not reflect realistic usage scenarios. We propose a unified framework to jointly analyze intrinsic and extrinsic gender bias in LLMs using identical neutral prompts, enabling direct comparison between gender-related information encoded in internal representations and bias expressed in generated outputs. Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. We further examine the effect of alignment through supervised fine-tuning aimed at reducing gender bias. Our results suggest that while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting. Finally, we consider two realistic settings and show that debiasing effects observed on structured benchmarks do not necessarily generalize, e.g., to the case of story generation.

Nour Bouchouchi, Thibault Laugel, Xavier Renard +3

Sorbonne Université, CNRS, LIP6, F-75005 Paris, France · AXA, Paris, France · Polish Academy of Science, IBS PAN, Warsaw, Poland

Gender Bias in Language Models LLM Alignment Social Bias in Language Models Language Model Bias Evaluation Linguistic Bias in Language Models

Mar 25, 2026

The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimation

RLHF-aligned language models exhibit response homogenization: on TruthfulQA (n=790), 40-79% of questions produce a single semantic cluster across 10 i.i.d. samples. On affected questions, sampling-based uncertainty methods have zero discriminative power (AUROC=0.500), while free token entropy retains signal (0.603). This alignment tax is task-dependent: on GSM8K (n=500), token entropy achieves 0.724 (Cohen's d=0.81). A base-vs-instruct ablation confirms the causal role of alignment: the base model shows 1.0% single-cluster rate vs. 28.5% for the instruct model (p < 10^{-6}). A training stage ablation (Base 0.0% -> SFT 1.5% -> DPO 4.0% SCR) localizes the cause to DPO, not SFT. Cross-family replication on four model families reveals alignment tax severity varies by family and scale. We validate across 22 experiments, 5 benchmarks, 4 model families, and 3 model scales (3B-14B), with Jaccard, embedding, and NLI-based baselines at three DeBERTa scales (all ~0.51 AUROC). Cross-embedder validation with two independent embedding families rules out coupling bias. Cross-dataset validation on WebQuestions (58.0% SCR) confirms generalization beyond TruthfulQA. The central finding -- response homogenization -- is implementation-independent and label-free. Motivated by this diagnosis, we explore a cheapest-first cascade (UCBD) over orthogonal uncertainty signals. Selective prediction raises GSM8K accuracy from 84.4% to 93.2% at 50% coverage; weakly dependent boundaries (|r| <= 0.12) enable 57% cost savings.

Mingyi Liu

Independent Researcher

Homogenization in Language Models LLM Alignment LLM Uncertainty Estimation

Mar 24, 2026

Unbiased Reward Modeling from Implicit Feedback for LLM Alignment

Despite the success of reinforcement learning from human feedback (RLHF), existing reward modeling methods largely rely on explicit feedback, which is costly to collect and difficult to scale. This work studies implicit reward modeling, learning reward models from implicit user feedback, such as clicks, copies and skips. While scalable and cost-effective, implicit feedback poses two key challenges: It lacks definitive negative samples, which makes standard positive-negative classification methods inapplicable; It suffers from selection bias, where responses have heterogeneous propensities to elicit feedback, which further obscures definitive negative samples. To address these challenges, we propose ImplicitRM, which learns unbiased reward models from implicit feedback. It stratifies training samples into four latent groups using a stratification model and derives a likelihood-maximization objective that is theoretically unbiased, thereby addressing both challenges. Experiments across diverse LLM backbones and benchmark datasets validate that ImplicitRM learns accurate reward models from implicit feedback and improves performance on downstream RLHF tasks.

Hao Wang, Haocheng Yang, Licheng Pan +7

Peking University · Xiaohongshu Inc. · National University of Singapore +1

Reward Modeling LLM Alignment RL from Human Feedback

Mar 24, 2026

When Language Models Lose Their Mind: The Consequences of Brain Misalignment

While brain-aligned large language models (LLMs) have garnered attention for their potential as cognitive models and for potential for enhanced safety and trustworthiness in AI, the role of this brain alignment for linguistic competence remains uncertain. In this work, we investigate the functional implications of brain alignment by introducing brain-misaligned models--LLMs intentionally trained to predict brain activity poorly while maintaining high language modeling performance. We evaluate these models on over 200 downstream tasks encompassing diverse linguistic domains, including semantics, syntax, discourse, reasoning, and morphology. By comparing brain-misaligned models with well-matched brain-aligned counterparts, we isolate the specific impact of brain alignment on language understanding. Our experiments reveal that brain misalignment substantially impairs downstream performance, highlighting the critical role of brain alignment in achieving robust linguistic competence. These findings underscore the importance of brain alignment in LLMs and offer novel insights into the relationship between neural representations and linguistic processing.

Gabriele Merlin, Mariya Toneva

MPI-SWS, Saarbrücken, Germany

Mar 24, 2026

Improving Safety Alignment via Balanced Direct Preference Optimization

With the rapid development and widespread application of Large Language Models (LLMs), their potential safety risks have attracted widespread attention. Reinforcement Learning from Human Feedback (RLHF) has been adopted to enhance the safety performance of LLMs. As a simple and effective alternative to RLHF, Direct Preference Optimization (DPO) is widely used for safety alignment. However, safety alignment still suffers from severe overfitting, which limits its actual performance. This paper revisits the overfitting phenomenon from the perspective of the model's comprehension of the training data. We find that the Imbalanced Preference Comprehension phenomenon exists between responses in preference pairs, which compromises the model's safety performance. To address this, we propose Balanced Direct Preference Optimization (B-DPO), which adaptively modulates optimization strength between preferred and dispreferred responses based on mutual information. A series of experimental results show that B-DPO can enhance the safety capability while maintaining the competitive general capabilities of LLMs on various mainstream benchmarks compared to state-of-the-art methods. \color{red}{Warning: This paper contains examples of harmful texts, and reader discretion is recommended.

Shiji Zhao, Mengyang Wang, Shukun Xiong +7

Institute of Artificial Intelligence, Beihang University, Beijing, China.

LLM Safety Alignment LLM Alignment Direct Preference Optimization

Mar 23, 2026

Reasoning or Rhetoric? An Empirical Analysis of Moral Reasoning Explanations in Large Language Models

Do large language models reason morally, or do they merely sound like they do? We investigate whether LLM responses to moral dilemmas exhibit genuine developmental progression through Kohlberg's stages of moral development, or whether alignment training instead produces reasoning-like outputs that superficially resemble mature moral judgment without the underlying developmental trajectory. Using an LLM-as-judge scoring pipeline validated across three judge models, we classify more than 600 responses from 13 LLMs spanning a range of architectures, parameter scales, and training regimes across six classical moral dilemmas, and conduct ten complementary analyses to characterize the nature and internal coherence of the resulting patterns. Our results reveal a striking inversion: responses overwhelmingly correspond to post-conventional reasoning (Stages 5-6) regardless of model size, architecture, or prompting strategy, the effective inverse of human developmental norms, where Stage 4 dominates. Most strikingly, a subset of models exhibit moral decoupling: systematic inconsistency between stated moral justification and action choice, a form of logical incoherence that persists across scale and prompting strategy and represents a direct reasoning consistency failure independent of rhetorical sophistication. Model scale carries a statistically significant but practically small effect; training type has no significant independent main effect; and models exhibit near-robotic cross-dilemma consistency producing logically indistinguishable responses across semantically distinct moral problems. We posit that these patterns constitute evidence for moral ventriloquism: the acquisition, through alignment training, of the rhetorical conventions of mature moral reasoning without the underlying developmental trajectory those conventions are meant to represent.

Aryan Kasat, Smriti Singh, Aman Chadha +1

UT Austin · Google GenAI

Moral Reasoning in Language Models Reasoning Consistency in Language Models LLM Alignment

Mar 21, 2026

Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami's novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields near-zero extraction, indicating that finetuning on individual authors' works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions (

r \ge 0.90

), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors' works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.

Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg +1

Stony Brook University · Carnegie Mellon University · Columbia Law School

LLM Alignment Memorization in Language Models Training Data Extraction LLM Fine-Tuning Memorization in Generative Models

Mar 21, 2026

Alignment midtraining for animals

We investigate the robustness of value alignment via midtraining with synthetic documents, using animal compassion as a value that is both important in its own right and orthogonal to existing alignment efforts. To evaluate compassionate reasoning, we develop and publicly release Animal Norms In Moral Assessment (ANIMA), a 26-question evaluation spanning 13 ethical dimensions, publicly available as a dataset and Inspect evaluation. On ANIMA, training with 3000 documents achieves 77% compared to 40% for instruction-tuning approaches, with generalization to human compassion and no degradation in standard safety benchmarks or capabilities. However, subsequent unrelated instruction-tuning degrades the intervention, with the advantage disappearing after 5000 samples. Our exploratory results suggest document-based value interventions may require explicit preservation strategies to remain effective through typical training pipelines.

Jasmine Brazilek, Miles Tidmarsh

Compassion Aligned Machine Learning

LLM Alignment LLM Training Moral Reasoning in Language Models

Mar 19, 2026

Ventriloquist LLMs: Linear Alignment of Late-Stage Representations

Independently trained language models often learn compatible late-stage representations, despite differences in training objectives, architectures, and data modalities. We ask how far this compatibility extends: can a simple affine map let one model's hidden states be read directly by another model's output head? In this work, we learn affine transformations between the final hidden states of independent models and evaluate them across embedding classification, out-of-distribution detection, and autoregressive text generation. Across model pairs, we find that downstream performance is largely preserved under alignment, with linearly mapped source representations retaining both the decision boundaries and the confidence structure of a target model's classifier. Additionally, we show for the first time that linear alignment sometimes enables text generation across independently trained models, decoding one model's hidden states through another's frozen output head without fine-tuning. This capability is far from universal, and characterizing when it holds is a contribution of our work. We find that success is governed by two factors: tokenizer overlap, which correlates strongly with generation quality, and source-model scale, below which quality degrades sharply. Transfer is also asymmetric, with strong-to-weak mappings substantially outperforming weak-to-strong ones, indicating that source representational capacity, rather than the target head, is the limiting factor. Finally, we characterize where in the network this compatibility arises. A layer-wise analysis localizes this compatibility to the final one or two layers, indicating a late-stage, output-space phenomenon rather than evidence of shared reasoning.

Matt Gorbett, Suman Jana

Capital One New York, NY 10003, USA · Department of Computer Science Columbia University New York, NY 10027, USA

Latent Space Alignment LLM Alignment Representation Geometry in Language Models

Mar 19, 2026

Do Post-Training Algorithms Actually Differ? A Controlled Study Across Model Scales Uncovers Scale-Dependent Ranking Inversions

Post-training alignment has produced dozens of competing algorithms -- DPO, SimPO, KTO, GRPO, and others -- yet practitioners lack controlled comparisons to guide algorithm selection. We present OXRL, a unified framework implementing 51 post-training algorithms with identical infrastructure, enabling the first large-scale apples-to-apples evaluation. Our study spans 8 algorithms across 4 model scales (0.5B--7B), 3 evaluation domains, and a 20-variant DPO taxonomy (100 runs at 1.5B, 5 seeds each), totaling

\sim

240 training runs on H100 GPUs. Three headline findings emerge. (1)~~Algorithm rankings are unstable across scale: at 1.5B, online RL (SGRPO) tops all methods at 58.0%~~

\pm

0.57 on GSM8K; by 7B, the worst small-scale method (SimPO) becomes the best (85.8%), a complete ranking inversion driven by model scale rather than LoRA regularization (confirmed via 2

\times

2 factorial). (2)~~Loss function modifications yield negligible gains: none of 20 DPO variants significantly outperform vanilla DPO after Bonferroni correction; the sole significant outlier, SimPO, is worse ( $-$ 11.5~~pp,

p < 10^{-4}

). (3)~~Algorithm leverage is task-specific: the 19.3~~pp GSM8K spread collapses to 0.54~~pp on MATH ( $36\times$ ) and 0.47~~pp on general-domain benchmarks (

41\times

), confirming that algorithm choice matters primarily within the training distribution. These findings yield a hierarchy of leverage for practitioners: model scale (

{\sim}

50~~pp) $\gg$ training paradigm ( ${\sim}$ 10~~pp)

\gg

online vs.\ offline (

{\sim}

9~~pp) $\gg$ loss function ( ${\sim}$ 1~~pp). We release all code, configs, and evaluation data as a living community benchmark.

Xiaoyi Li

Language Model Post-Training LLM Alignment LLM Evaluation

Mar 18, 2026

Sharpness-Aware Minimization in Logit Space Efficiently Enhances Direct Preference Optimization

Direct Preference Optimization (DPO) has emerged as a popular algorithm for aligning pretrained large language models with human preferences, owing to its simplicity and training stability. However, DPO suffers from the recently identified squeezing effect (also known as likelihood displacement), where the probability of preferred responses decreases unintentionally during training. To understand and mitigate this phenomenon, we develop a theoretical framework that models the coordinate-wise dynamics in logit space. Our analysis reveals that negative-gradient updates cause residuals to expand rapidly along high-curvature directions, which underlies the squeezing effect, whereas Sharpness-Aware Minimization (SAM) can suppress this behavior through its curvature-regularization effect. Building on this insight, we investigate logits-SAM, a computationally efficient variant that perturbs only the output layer with negligible overhead. Extensive experiments on Pythia-2.8B, Mistral-7B, and Gemma-2B-IT across multiple datasets and benchmarks demonstrate that logits-SAM consistently improves the effectiveness of DPO and integrates seamlessly with other DPO variants. Code is available at https://github.com/RitianLuo/logits-sam-dpo.

Haocheng Luo, Zehang Deng, Thanh-Toan Do +3

Monash University, Australia · Swinburne University of Technology, Australia

LLM Alignment Direct Preference Optimization Sharpness-Aware Minimization

Mar 18, 2026

VC-Soup: Value-Consistency Guided Multi-Value Alignment for Large Language Models

As large language models (LLMs) increasingly shape content generation, interaction, and decision-making across the Web, aligning them with human values has become a central objective in trustworthy AI. This challenge becomes even more pronounced when aligning multiple, potentially conflicting human values. Although recent approaches, such as reward reweighting, prompt-based supervised fine-tuning, and model merging, attempt to tackle multi-value alignment, they still face two major limitations: (1) training separate models for each value combination is prohibitively expensive; (2) value conflicts substantially degrade alignment performance. These limitations make it difficult to achieve favorable trade-offs across diverse human values. To address these challenges, we revisit multi-value alignment from the perspective of value consistency in data and propose VC-soup, a data filtering and parameter merging framework grounded in value-consistent learning. We first design a value consistency metric based on the cosine similarity between the reward-gap vector of each preference pair and an all-ones vector, which quantifies its cross-value coherence. We then filter out low-consistency preference pairs in each value dataset and train on the remaining data to obtain smooth, value-consistent policy models that better preserve linear mode connectivity. Finally, we linearly combine these policies and apply Pareto filtering across values to obtain solutions with balanced multi-value performance. Extensive experiments and theoretical analysis demonstrate that VC-soup effectively mitigates conflicts and consistently outperforms existing multi-value alignment methods.

Hefei Xu, Le Wu, Yu Wang +4

Key Laboratory of Knowledge Engineering with Big Data, Hefei University of Technology, Hefei, Anhui, China · Innovation School of Artificial Intelligence, Hefei University of Technology

LLM Alignment Multi-Objective Language Model Alignment Model Merging

Mar 17, 2026

Alignment Makes Language Models Normative, Not Descriptive

Post-training alignment optimizes language models to match human preference signals, but this objective is not equivalent to modeling observed human behavior. We compare 120 base-aligned model pairs on more than 10,000 real human decisions in multi-round strategic games - bargaining, persuasion, negotiation, and repeated matrix games. In these settings, base models outperform their aligned counterparts in predicting human choices by nearly 10:1, robustly across model families, prompt formulations, and game configurations. This pattern reverses, however, in settings where human behavior is more likely to follow normative predictions: aligned models dominate on one-shot textbook games across all 12 types tested and on non-strategic lottery choices - and even within the multi-round games themselves, at round one, before interaction history develops. This boundary-condition pattern suggests that alignment induces a normative bias: it improves prediction when human behavior is relatively well captured by normative solutions, but hurts prediction in multi-round strategic settings, where behavior is shaped by descriptive dynamics such as reciprocity, retaliation, and history-dependent adaptation. These results reveal a fundamental trade-off between optimizing models for human use and using them as proxies for human behavior.

Eilam Shapira, Moshe Tennenholtz, Roi Reichart

Technion – Israel Institute of Technology

LLM Alignment Preference Alignment Normative Conformity in Language Models Human Behavior Modeling Human Behavior Prediction

Mar 17, 2026

Prompt Programming for Cultural Bias and Alignment of Large Language Models

Culture shapes reasoning, values, prioritization, and strategic decision-making, yet large language models (LLMs) often exhibit cultural biases that misalign with target populations. As LLMs are increasingly used for strategic decision-making, policy support, and document engineering tasks such as summarization, categorization, and compliance-oriented auditing, improving cultural alignment is important for ensuring that downstream analyses and recommendations reflect target-population value profiles rather than default model priors. Previous work introduced a survey-grounded cultural alignment framework and showed that culture-specific prompting can reduce misalignment, but it primarily evaluated proprietary models and relied on manual prompt engineering. In this paper, we validate and extend that framework by reproducing its social sciences survey based projection and distance metrics on open-weight LLMs, testing whether the same cultural skew and benefits of culture conditioning persist outside closed LLM systems. Building on this foundation, we introduce use of prompt programming with DSPy for this problem-treating prompts as modular, optimizable programs-to systematically tune cultural conditioning by optimizing against cultural-distance objectives. In our experiments, we show that prompt optimization often improves upon cultural prompt engineering, suggesting prompt compilation with DSPy can provide a more stable and transferable route to culturally aligned LLM responses.

Maksim Eren, Eric Michalak, Brian Cook +1

Advanced Research in Cyber Systems, Los Alamos National Laboratory, Los Alamos, New Mexico, USA · Analytics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA

LLM Alignment Cultural Alignment Cultural Bias in Language Models Cross-Cultural Language Model Evaluation Prompt Optimization

Mar 17, 2026

Domain Mixture Design via Log-Likelihood Differences for Aligning Language Models with a Target Model

Instead of directly distilling a language model, this study addresses the problem of aligning a base model with a target model in distribution by designing the domain mixture of training data for pretraining or continued pretraining as a fixed training recipe. We propose a method for determining domain weights by viewing models as points in log-likelihood space and aligning the training update direction with the direction toward the target model. Experiments with NanoGPT show that the proposed method consistently reduces the KL divergence between the trained base model and the target model relative to training with Pile-original weighting. Although knowledge distillation remains more effective when available, the proposed method achieves meaningful alignment, and downstream task performance also tends to become closer to that of the target model.

Ryo Kishino, Riku Shiomi, Hiroaki Yamagiwa +2

1Kyoto University · 2RIKEN

Language Model Pretraining LLM Alignment Data Mixture Optimization

Mar 17, 2026

Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences

Recent empirical results have demonstrated that training large language models (LLMs) with negative-only feedback can match or exceed standard reinforcement learning from human feedback (RLHF). Negative Sample Reinforcement achieves parity with PPO on mathematical reasoning; Distributional Dispreference Optimization trains effectively using only dispreferred samples; and Constitutional AI outperforms pure RLHF on harmlessness benchmarks. Yet no unified theoretical account explains why negative signals are so effective. This paper proposes such an account: positive preferences and negative constraints are structurally asymmetric. Positive preferences ("which is better") encode continuously coupled, context-dependent human values that cannot be exhaustively specified -- leading models to learn surface correlates such as agreement with the user (sycophancy). Negative constraints ("what is wrong") encode discrete, finite, independently verifiable prohibitions that can converge to a stable boundary. This asymmetry -- rooted in Popper's falsification logic and the epistemology of negative knowledge -- explains both the sycophancy failure of preference-based RLHF and the surprising effectiveness of negative-signal methods. We argue that alignment research should shift its center of gravity from "learning what humans prefer" to "learning what humans reject," and offer testable predictions for this framework.

Quan Cheng

Tsinghua University

LLM Alignment RL from Human Feedback Preference-Based RL

Mar 17, 2026

PlotTwist: A Creative Plot Generation Framework with Small Language Models

Creative plot generation presents a fundamental challenge for language models: transforming a concise premise into a coherent narrative that sustains global coherence, character development, pacing, tone consistency, and emotional progression. Although recent Large Language Models (LLMs) demonstrate strong fluency on general-purpose tasks, they require preference alignment to perform well on domain-specific tasks such as creative plot generation. However, conducting such alignment at the scale of frontier LLMs is computationally prohibitive, significantly limiting accessibility and practical deployment. To address this, we present PlotTwist, a structured framework that enables Small Language Models (SLMs) with

\leq

3B active parameters to generate high-quality, premise-conditioned plots competitive with frontier systems of vastly greater parameter scale. Our approach decomposes generation into three specialized components: (1) an Aspect Rating Reward Model, trained via a novel Positive-Negative prompting strategy; (2) a Mixture-of-Experts (MoE) plot generator aligned via Direct Preference Optimization (DPO); and (3) an Agentic Evaluation module using a cross-family jury for unbiased, independent post-hoc assessment. Extensive experiments demonstrate that PlotTwist consistently outperforms all baselines, including frontier models, across multiple Narrative Quality Dimensions (NQDs), achieving higher win rates against every baseline except the strongest, with which it remains competitive. Further validation confirms strong sensitivity to narrative quality, as the framework reliably distinguishes plots derived from critically acclaimed versus widely panned screenplays. Together, these results establish structured, preference-based alignment as a resource-efficient approach to high-quality creative plot generation. Project page: https://abhinavthorat.github.io/plottwist/

Abhinav Thorat, Ravi Kolla, Jyotin Goel +2

Sony Research India

Narrative Generation Small Language Models LLM Alignment

Mar 17, 2026

HIPO: Instruction Hierarchy via Constrained Reinforcement Learning

Hierarchical Instruction Following (HIF) refers to the problem of prompting large language models with a priority-ordered stack of instructions. Standard methods like RLHF and DPO typically fail in this problem since they mainly optimize for a single objective, failing to explicitly enforce system prompt compliance. Meanwhile, supervised fine-tuning relies on mimicking filtered, compliant data, which fails to establish the priority asymmetry at the algorithmic level. In this paper, we introduce HIPO, a novel alignment framework that formulates HIF as a Constrained Markov Decision Process. HIPO elevates system prompts from mere input context to strict algorithmic boundaries. Using a primal-dual safe reinforcement learning approach, the algorithm dynamically enforces system prompt compliance as an explicit constraint, maximizing user utility strictly within this feasible region. Extensive evaluations across diverse model architectures (e.g., Qwen, Phi, Llama) demonstrate that HIPO significantly improves both system compliance and user utility. Furthermore, mechanistic analysis reveals that this constrained optimization autonomously drives the model to shift its attention toward long-range system tokens, providing a principled foundation for reliable LLM deployment in complex workflows.

Keru Chen, Jun Luo, Sen Lin +4

School of ECEE Arizona State University · Department of ECE The Ohio State University · Computer Science Department University of Houston +2

LLM Alignment RL for Language Models LLM Safety Alignment Constrained RL

Mar 16, 2026

Mechanistic Origin of Moral Indifference in Language Models

Existing behavioral alignment techniques for Large Language Models (LLMs) often neglect the discrepancy between surface compliance and internal unaligned representations, leaving LLMs vulnerable to long-tail risks. More crucially, we posit that LLMs possess an inherent state of moral indifference due to compressing distinct moral concepts into uniform probability distributions. We verify and remedy this indifference in LLMs' latent representations, utilizing 251k moral vectors constructed upon Prototype Theory and the Social-Chemistry-101 dataset. Firstly, our analysis across 23 models reveals that current LLMs fail to represent the distinction between opposed moral categories and fine-grained typicality gradients within these categories; notably, neither model scaling, architecture, nor explicit alignment reshapes this indifference. We then employ Sparse Autoencoders on Qwen3-8B, isolate mono-semantic moral features, and targetedly reconstruct their topological relationships to align with ground-truth moral vectors. This representational alignment naturally improves moral reasoning and granularity, achieving a 75% pairwise win-rate on the independent adversarial Flames benchmark. Finally, we elaborate on the remedial nature of current intervention methods from an experientialist philosophy, arguing that endogenously aligned AI might require a transformation from post-hoc corrections to proactive cultivation.

Lingyu Li, Yan Teng, Yingchun Wang

Shanghai Artificial Intelligence Laboratory

LLM Alignment Moral Reasoning in Language Models

Mar 16, 2026

Are Dilemmas and Conflicts in LLM Alignment Solvable? A View from Priority Graph

As Large Language Models (LLMs) become more powerful and autonomous, they increasingly face conflicts and dilemmas in many scenarios. We first summarize and taxonomize these diverse conflicts. Then, we model the LLM's preferences to make different choices as a priority graph, where instructions and values are nodes, and the edges represent context-specific priorities determined by the model's output distribution. This graph reveals that a unified stable LLM alignment is very challenging, because the graph is neither static nor necessarily consistent in different contexts. Besides, it also reveals a potential vulnerability: priority hacking, where adversaries can craft deceptive contexts to manipulate the graph and bypass safety alignments. To counter this, we propose a runtime verification mechanism, enabling LLMs to query external sources to ground their context and resist manipulation. While this approach enhances robustness, we also acknowledge that many ethical and value dilemmas are philosophically irreducible, posing a long-term, open challenge for the future of AI alignment.

Zhenheng Tang, Xiang Liu, Qian Wang +3

The Hong Kong University of Science and Technology · The Hong Kong University of Science and Technology (Guangzhou) · New York University +1

LLM Alignment LLM Safety Adversarial Attacks on LLMs

Mar 16, 2026

Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook

As LLMs are globally deployed, aligning their cultural value orientations is critical for safety and user engagement. However, existing benchmarks face the Construct-Composition-Context (

C^3

) challenge: relying on discriminative, multiple-choice formats that probe value knowledge rather than true orientations, overlook subcultural heterogeneity, and mismatch with real-world open-ended generation. We introduce DOVE, a distributional evaluation framework that directly compares human-written text distributions with LLM-generated outputs. DOVE utilizes a rate-distortion variational optimization objective to construct a compact value codebook from 10K documents, mapping text into a structured value space to filter semantic noise. Alignment is measured using unbalanced optimal transport, capturing intra-cultural distributional structures and subgroup diversity. Experiments across 12 LLMs show that DOVE achieves superior predictive validity, attaining a 31.56% correlation with downstream tasks, while maintaining high reliability with as few as 500 samples per culture.

Jaehyeok Lee, Xiaoyuan Yi, Jing Yao +4

Sungkyunkwan University · Microsoft Research Asia · Singapore University of Technology and Design

LLM Evaluation LLM Alignment Cultural Alignment Cross-Cultural Language Model Evaluation

Mar 16, 2026

Beyond Creed: A Non-Identity Safety Condition A Strong Empirical Alternative to Identity Framing in Low-Data LoRA Fine-Tuning

How safety supervision is written may matter more than the explicit identity content it contains. We study low-data LoRA safety fine-tuning with four supervision formats built from the same core safety rules: constitutional rules (A), creed-style identity framing (B), a B-matched creed condition with a worldview/confession identity-maintenance tail (C), and a matched non-identity condition (D). Across three instruction-tuned model families (Llama 3.1 8B, Qwen2.5 7B, and Gemma 3 4B), we evaluate HarmBench using a reconciled dual-judge pipeline combining Bedrock-hosted DeepSeek v3.2 and Sonnet 4.6, with disagreement and boundary cases manually resolved. The non-identity condition D is the strongest group on all three model families on the full 320-behavior HarmBench set, reaching 74.4% refusal on Llama, 76.9% on Gemma, and 74.1% on Qwen. By comparison, creed-style framing (B) improves over plain constitutional rules (A) on Llama and Gemma, but remains substantially below D, yielding an overall descriptive ordering of

D > B > C \geq A > baseline

. This provides a bounded empirical challenge to a strong version of the identity-framing hypothesis: explicit creed-style identity language is not necessary for the strongest gains observed here. Capability evaluations on MMLU and ARC-Challenge show no meaningful trade-off across conditions.

Xinran Zhang

University of California, Berkeley

Safety-Preserving Fine-Tuning of Language Models LLM Safety LLM Alignment LLM Safety Evaluation

Mar 15, 2026

PA3: Policy-Aware Agent Alignment through Chain-of-Thought

Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the 'needle-in-a-haystack' problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel Policy Recall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Shubhashis Roy Dipta, Daniel Bis, Kun Zhou +4

University of Maryland, Baltimore County · Amazon Alexa AI

LLM Alignment RL for Language Model Reasoning

Mar 15, 2026

An Industrial-Scale Insurance LLM Achieving Verifiable Domain Mastery and Hallucination Control without Competence Trade-offs

Adapting Large Language Models (LLMs) to high-stakes vertical domains like insurance presents a significant challenge: scenarios demand strict adherence to complex regulations and business logic with zero tolerance for hallucinations. Existing approaches often suffer from a Competency Trade-off - sacrificing general intelligence for domain expertise - or rely heavily on RAG without intrinsic reasoning. To bridge this gap, we present INS-S1, an insurance-specific LLM family trained via a novel end-to-end alignment paradigm. Our approach features two methodological innovations: (1) A Verifiable Data Synthesis System that constructs hierarchical datasets for actuarial reasoning and compliance; and (2) A Progressive SFT-RL Curriculum Framework that integrates dynamic data annealing with a synergistic mix of Verified Reasoning (RLVR) and AI Feedback (RLAIF). By optimizing data ratios and reward signals, this framework enforces domain constraints while preventing catastrophic forgetting. Additionally, we release INSEva, the most comprehensive insurance benchmark to date (39k+ samples). Extensive experiments show that INS-S1 achieves SOTA performance on domain tasks, significantly outperforming DeepSeek-R1 and Gemini-2.5-Pro. Crucially, it maintains top-tier general capabilities and achieves a record-low 0.6% hallucination rate (HHEM). Our results demonstrate that rigorous domain specialization can be achieved without compromising general intelligence.

Qian Zhu, Xinnan Guo, Jingjing Huo +5

Ant Group

Domain Adaptation for LLMs LLM Alignment LLM Hallucination Mitigation RL for Language Model Reasoning LLM Evaluation

Mar 14, 2026

Retrieval-Feedback-Driven Distillation and Preference Alignment for Efficient LLM-based Query Expansion

Large language models have recently enabled a generative paradigm for query expansion, but their high inference cost makes direct deployment difficult in practical retrieval systems. To address this issue, a retrieval-feedback-driven distillation and preference-alignment framework is proposed to transfer retrieval-friendly expansion behavior from a strong teacher model to a compact student model. Rather than relying on few-shot exemplars at inference time, the framework first leverages two complementary types of teacher-generated expansions, produced under zero-shot and few-shot prompting conditions, as supervision signals for distillation and as candidate pools for preference construction. A retrieval-metric-driven strategy is then introduced to automatically form chosen/rejected expansion pairs according to nDCG@10 differences, and Direct Preference Optimization is applied to explicitly align generation preferences with retrieval objectives. Experiments on TREC DL19/20/21 and MIRACL-zh show that the proposed approach preserves strong retrieval effectiveness while substantially reducing inference cost. In particular, the distilled Qwen3-4B model reaches about 97% of the teacher (DeepSeek-685B) model's nDCG@10 performance on DL19, and remains effective on the Chinese MIRACL-zh benchmark, demonstrating strong practicality across both English and Chinese retrieval settings.

Minghan Li, Guodong Zhou

School of Computer Science and Technology, Soochow University, China

Language Model Distillation Efficient Language Model Inference LLM Alignment

Mar 13, 2026

CLARIN-PT-LDB: An Open LLM Leaderboard for Portuguese to assess Language, Culture and Civility

This paper reports on the development of a leaderboard of Open Large Language Models (LLM) for European Portuguese (PT-PT), and on its associated benchmarks. This leaderboard comes as a way to address a gap in the evaluation of LLM for European Portuguese, which so far had no leaderboard dedicated to this variant of the language. The paper also reports on novel benchmarks, including some that address aspects of performance that so far have not been available in benchmarks for European Portuguese, namely model safeguards and alignment to Portuguese culture. The leaderboard is available at https://huggingface.co/spaces/PORTULAN/portuguese-llm-leaderboard.

João Silva, Luís Gomes, António Branco

University of Lisbon NLX—Grupo de Fala e Linguagem Natural, Departmento de Informática Faculdade de Ciências, Campo Grande, 1749-016 Lisboa, Portugal

LLM Evaluation Portuguese Language Models Cross-Cultural Language Model Evaluation LLM Safety Benchmarks LLM Alignment