Daily Papers

Procedural content generation has enabled vast virtual worlds through levels, maps, and quests, but large-scale character generation remains underexplored. We identify two alignment-induced biases in existing methods: a positive moral bias, where characters uniformly adopt agreeable stances (e.g. always saying lying is bad), and a helpful assistant bias, where characters invariably answer questions directly (e.g. never refusing or deflecting). While such tendencies suit instruction-following systems, they suppress dramatic tension and yield predictable characters, stemming from maximum likelihood training and assistant fine-tuning. To address this, we introduce PersonaWeaver, a framework that disentangles world-building (roles, demographics) from behavioral-building (moral stances, interactional styles), yielding characters with more diverse reactions and moral stances, as well as second-order diversity in stylistic markers like length, tone, and punctuation. Code: https://github.com/mqraitem/Persona-Weaver

Multimodal Large Language Models (MLLMs) have emerged as a central focus in both industry and academia, but often suffer from biases introduced by visual and language priors, which can lead to multimodal hallucination. These biases arise from the visual encoder and the Large Language Model (LLM) backbone, affecting the attention mechanism responsible for aligning multimodal inputs. Existing decoding-based mitigation methods focus on statistical correlations and overlook the causal relationships between attention mechanisms and model output, limiting their effectiveness in addressing these biases. To tackle this issue, we propose a causal inference framework termed CausalMM that applies structural causal modeling to MLLMs, treating modality priors as a confounder between attention mechanisms and output. Specifically, by employing backdoor adjustment and counterfactual reasoning at both the visual and language attention levels, our method mitigates the negative effects of modality priors and enhances the alignment of MLLM's inputs and outputs, with a maximum score improvement of 65.3% on 6 VLind-Bench indicators and 164 points on MME Benchmark compared to conventional methods. Extensive experiments validate the effectiveness of our approach while being a plug-and-play solution. Our code is available at: https://github.com/The-Martyr/CausalMM

We introduce MENAValues, a novel benchmark designed to evaluate the cultural alignment and multilingual biases of large language models (LLMs) with respect to the beliefs and values of the Middle East and North Africa (MENA) region, an underrepresented area in current AI evaluation efforts. Drawing from large-scale, authoritative human surveys, we curate a structured dataset that captures the sociocultural landscape of MENA with population-level response distributions from 16 countries. To probe LLM behavior, we evaluate diverse models across multiple conditions formed by crossing three perspective framings (neutral, personalized, and third-person/cultural observer) with two language modes (English and localized native languages: Arabic, Persian, Turkish). Our analysis reveals three critical phenomena: "Cross-Lingual Value Shifts" where identical questions yield drastically different responses based on language, "Reasoning-Induced Degradation" where prompting models to explain their reasoning worsens cultural alignment, and "Logit Leakage" where models refuse sensitive questions while internal probabilities reveal strong hidden preferences. We further demonstrate that models collapse into simplistic linguistic categories when operating in native languages, treating diverse nations as monolithic entities. MENAValues offers a scalable framework for diagnosing cultural misalignment, providing both empirical insights and methodological tools for developing more culturally inclusive AI.

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.

Despite significant research efforts, deep neural networks are still vulnerable to biases: this raises concerns about their fairness and limits their generalization. In this paper, we propose a bias-agnostic approach to mitigate the impact of bias in deep neural networks. Unlike traditional debiasing approaches, we rely on a metric to quantify ``bias alignment/misalignment'' on target classes, and use this information to discourage the propagation of bias-target alignment information through the network. We conduct experiments on several commonly used datasets for debiasing and compare our method to supervised and bias-specific approaches. Our results indicate that the proposed method achieves comparable performance to state-of-the-art supervised approaches, although it is bias-agnostic, even in presence of multiple biases in the same sample.

Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: https://alignment-tampering.github.io/

Pretraining corpora contain extensive discourse about AI systems, yet the causal influence of this discourse on downstream alignment remains poorly understood. If prevailing descriptions of AI behaviour are predominantly negative, LLMs may internalise corresponding behavioural priors, giving rise to self-fulfilling misalignment. This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse. We find that discussion of AI contributes to misalignment. Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour. Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%. We consider this evidence of self-fulfilling alignment. These effects are dampened, but persist through post-training. Our findings establish the study of how pretraining data shapes alignment priors, or alignment pretraining, as a complement to post-training. We recommend practitioners pretrain for alignment as well as capabilities. Our models and datasets are available at alignmentpretraining.ai

From the perspective of content safety issues, alignment has shown to limit large language models' (LLMs) harmful content generation. This intentional method of reinforcing models to not respond to certain user inputs seem to be present in many modern open-source instruction tuning datasets such as OpenAssistant or Guanaco. We introduce a novel insight to an instruction-tuned model's performance affected by the presence of alignment in supervised fine-tuning dataset. To be specific, we noticed that alignment acts as if it is poisoning the instruction dataset. Experimentally, we demonstrate that aligned answers significantly worsen the performance of the resulting fine-tuned model's on various reasoning benchmarks such as Big Bench (BBH), Massive Multitask Language Understanding (MMLU), Human Eval, and Discrete Reasoning Over Paragraphs (DROP), performing worse than the counterpart tuned without alignment by 4-33%.

Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned, giving stereotypically `evil' responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases and find that models can just learn the narrow dataset task, but that the general solution appears to be more stable and more efficient. To establish this, we build on the result that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. We find a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and preliminary metrics for investigating how inductive biases shape generalisation in LLMs. We open-source all code, datasets and model finetunes.

Advances in unsupervised probes such as Contrast-Consistent Search (CCS), which reveal latent beliefs without relying on token outputs, raise the question of whether these methods can reliably assess model alignment. We investigate this by examining the sensitivity of CCS to harmful vs. safe statements and by introducing Polarity-Aware CCS (PA-CCS), a method for evaluating whether a model's internal representations remain consistent under polarity inversion. We propose two alignment-oriented metrics, Polar-Consistency and the Contradiction Index, to quantify the semantic robustness of a model's latent knowledge. To validate PA-CCS, we curate two main datasets and one control dataset containing matched harmful-safe sentence pairs constructed using different methodologies (concurrent and antagonistic statements). We apply PA-CCS to 16 language models. Our results show that PA-CCS identifies both architectural and layer-specific differences in the encoding of latent harmful knowledge. Notably, replacing the negation token with a meaningless marker degrades PA-CCS scores for models with well-aligned internal representations, while models lacking robust internal calibration do not exhibit this degradation. Our findings highlight the potential of unsupervised probing for alignment evaluation and emphasize the need to incorporate structural robustness checks into interpretability benchmarks. Code and datasets are available at: https://github.com/SadSabrina/polarity-probing. WARNING: This paper contains potentially sensitive, harmful, and offensive content.

We generalize the notion of social biases from language embeddings to grounded vision and language embeddings. Biases are present in grounded embeddings, and indeed seem to be equally or more significant than for ungrounded embeddings. This is despite the fact that vision and language can suffer from different biases, which one might hope could attenuate the biases in both. Multiple ways exist to generalize metrics measuring bias in word embeddings to this new setting. We introduce the space of generalizations (Grounded-WEAT and Grounded-SEAT) and demonstrate that three generalizations answer different yet important questions about how biases, language, and vision interact. These metrics are used on a new dataset, the first for grounded bias, created by augmenting extending standard linguistic bias benchmarks with 10,228 images from COCO, Conceptual Captions, and Google Images. Dataset construction is challenging because vision datasets are themselves very biased. The presence of these biases in systems will begin to have real-world consequences as they are deployed, making carefully measuring bias and then mitigating it critical to building a fair society.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated impressive performance on standard visual reasoning benchmarks. However, there is growing concern that these models rely excessively on linguistic shortcuts rather than genuine visual grounding, a phenomenon we term Text Bias. In this paper, we investigate the fundamental tension between visual perception and linguistic priors. We decouple the sources of this bias into two dimensions: Internal Corpus Bias, stemming from statistical correlations in pretraining, and External Instruction Bias, arising from the alignment-induced tendency toward sycophancy. To quantify this effect, we introduce V-FAT (Visual Fidelity Against Text-bias), a diagnostic benchmark comprising 4,026 VQA instances across six semantic domains. V-FAT employs a Three-Level Evaluation Framework that systematically increases the conflict between visual evidence and textual information: (L1) internal bias from atypical images, (L2) external bias from misleading instructions, and (L3) synergistic bias where both coincide. We introduce the Visual Robustness Score (VRS), a metric designed to penalize "lucky" linguistic guesses and reward true visual fidelity. Our evaluation of 12 frontier MLLMs reveals that while models excel in existing benchmarks, they experience significant visual collapse under high linguistic dominance.

Aligning large language models (LLMs) with human intentions has become a critical task for safely deploying models in real-world systems. While existing alignment approaches have seen empirical success, theoretically understanding how these methods affect model behavior remains an open question. Our work provides an initial attempt to theoretically analyze the learning dynamics of human preference alignment. We formally show how the distribution of preference datasets influences the rate of model updates and provide rigorous guarantees on the training accuracy. Our theory also reveals an intricate phenomenon where the optimization is prone to prioritizing certain behaviors with higher preference distinguishability. We empirically validate our findings on contemporary LLMs and alignment tasks, reinforcing our theoretical insights and shedding light on considerations for future alignment approaches. Disclaimer: This paper contains potentially offensive text; reader discretion is advised.

Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.

Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias (Delta DP of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection (Delta DP of 27%) results. Even higher bias (Delta DP~38%) results on 9 other chat based downstream applications.

Multimodal large language models (MLLMs) exhibit a pronounced preference for textual inputs when processing vision-language data, limiting their ability to reason effectively from visual evidence. Unlike prior studies that attribute this text bias to external factors such as data imbalance or instruction tuning, we propose that the bias originates from the model's internal architecture. Specifically, we hypothesize that visual key vectors (Visual Keys) are out-of-distribution (OOD) relative to the text key space learned during language-only pretraining. Consequently, these visual keys receive systematically lower similarity scores during attention computation, leading to their under-utilization in the context representation. To validate this hypothesis, we extract key vectors from LLaVA and Qwen2.5-VL and analyze their distributional structures using qualitative (t-SNE) and quantitative (Jensen-Shannon divergence) methods. The results provide direct evidence that visual and textual keys occupy markedly distinct subspaces within the attention space. The inter-modal divergence is statistically significant, exceeding intra-modal variation by several orders of magnitude. These findings reveal that text bias arises from an intrinsic misalignment within the attention key space rather than solely from external data factors.

In AI alignment, extensive latitude must be granted to annotators, either human or algorithmic, to judge which model outputs are `better' or `safer.' We refer to this latitude as alignment discretion. Such discretion remains largely unexamined, posing two risks: (i) annotators may use their power of discretion arbitrarily, and (ii) models may fail to mimic this discretion. To study this phenomenon, we draw on legal concepts of discretion that structure how decision-making authority is conferred and exercised, particularly in cases where principles conflict or their application is unclear or irrelevant. Extended to AI alignment, discretion is required when alignment principles and rules are (inevitably) conflicting or indecisive. We present a set of metrics to systematically analyze when and how discretion in AI alignment is exercised, such that both risks (i) and (ii) can be observed. Moreover, we distinguish between human and algorithmic discretion and analyze the discrepancy between them. By measuring both human and algorithmic discretion over safety alignment datasets, we reveal layers of discretion in the alignment process that were previously unaccounted for. Furthermore, we demonstrate how algorithms trained on these datasets develop their own forms of discretion in interpreting and applying these principles, which challenges the purpose of having any principles at all. Our paper presents the first step towards formalizing this core gap in current alignment processes, and we call on the community to further scrutinize and control alignment discretion.

Before being deployed for user-facing applications, developers align Large Language Models (LLMs) to user preferences through a variety of procedures, such as Reinforcement Learning From Human Feedback (RLHF) and Direct Preference Optimization (DPO). Current evaluations of these procedures focus on benchmarks of instruction following, reasoning, and truthfulness. However, human preferences are not universal, and aligning to specific preference sets may have unintended effects. We explore how alignment impacts performance along three axes of global representation: English dialects, multilingualism, and opinions from and about countries worldwide. Our results show that current alignment procedures create disparities between English dialects and global opinions. We find alignment improves capabilities in several languages. We conclude by discussing design decisions that led to these unintended impacts and recommendations for more equitable preference tuning.

To be discoverable in an embedding-based search process, each part of a document should be reflected in its embedding representation. To quantify any potential reflection biases, we introduce a permutation-based evaluation framework. With this, we observe that state-of-the-art embedding models exhibit systematic positional and language biases when documents are longer and consist of multiple segments. Specifically, early segments and segments in higher-resource languages like English are over-represented, while later segments and segments in lower-resource languages are marginalized. In our further analysis, we find that the positional bias stems from front-loaded attention distributions in pooling-token embeddings, where early tokens receive more attention. To mitigate this issue, we introduce an inference-time attention calibration method that redistributes attention more evenly across document positions, increasing discoverabiltiy of later segments. Our evaluation framework and attention calibration is available at https://github.com/impresso/fair-sentence-transformers

Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over 30 cognitive biases. Second, we introduce cross-tuning -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.

Fine-tuning large language models on narrow datasets can cause them to develop broadly misaligned behaviours: a phenomena known as emergent misalignment. However, the mechanisms underlying this misalignment, and why it generalizes beyond the training domain, are poorly understood, demonstrating critical gaps in our knowledge of model alignment. In this work, we train and study a minimal model organism which uses just 9 rank-1 adapters to emergently misalign Qwen2.5-14B-Instruct. Studying this, we find that different emergently misaligned models converge to similar representations of misalignment. We demonstrate this convergence by extracting a 'misalignment direction' from one fine-tuned model's activations, and using it to effectively ablate misaligned behaviour from fine-tunes using higher dimensional LoRAs and different datasets. Leveraging the scalar hidden state of rank-1 LoRAs, we further present a set of experiments for directly interpreting the fine-tuning adapters, showing that six contribute to general misalignment, while two specialise for misalignment in just the fine-tuning domain. Emergent misalignment is a particularly salient example of undesirable and unexpected model behaviour and by advancing our understanding of the mechanisms behind it, we hope to move towards being able to better understand and mitigate misalignment more generally.

Social bias in machine learning has drawn significant attention, with work ranging from demonstrations of bias in a multitude of applications, curating definitions of fairness for different contexts, to developing algorithms to mitigate bias. In natural language processing, gender bias has been shown to exist in context-free word embeddings. Recently, contextual word representations have outperformed word embeddings in several downstream NLP tasks. These word representations are conditioned on their context within a sentence, and can also be used to encode the entire sentence. In this paper, we analyze the extent to which state-of-the-art models for contextual word representations, such as BERT and GPT-2, encode biases with respect to gender, race, and intersectional identities. Towards this, we propose assessing bias at the contextual word level. This novel approach captures the contextual effects of bias missing in context-free word embeddings, yet avoids confounding effects that underestimate bias at the sentence encoding level. We demonstrate evidence of bias at the corpus level, find varying evidence of bias in embedding association tests, show in particular that racial bias is strongly encoded in contextual word models, and observe that bias effects for intersectional minorities are exacerbated beyond their constituent minority identities. Further, evaluating bias effects at the contextual word level captures biases that are not captured at the sentence level, confirming the need for our novel approach.

Prior work has shown that fine-tuning large language models on malicious or incorrect outputs in narrow domains can induce broad misalignment and harmful behavior, a phenomenon known as emergent misalignment. However, efficient methods for reversing such misalignment remain limited. In this work, we make two contributions. First, we identify sycophancy fine-tuning, i.e., training models to passively agree with users' incorrect opinions, as a previously underexplored driver of emergent misalignment, and show that it induces broad and severe misaligned behavior. Second, we propose Alignment Gating, an efficient method for reversing emergent misalignment that inserts learnable and controllable gates into the model during fine-tuning. Through fine-tuning, these gates learn to identify the internal representations responsible for unsafe responses. Thus, amplifying or suppressing these representations then exacerbates or mitigates EM, respectively. We further find that alignment gating module exhibits strong generalization: gating weights obtained from narrow-domain fine-tuning substantially suppress broad-domain misaligned behavior while preserving the model's general capabilities.

Aligned instruction following models can better fulfill user requests than their unaligned counterparts. However, it has been shown that there is a length bias in evaluation of such models, and that training algorithms tend to exploit this bias by learning longer responses. In this work we show how to train models that can be controlled at inference time with instructions containing desired length constraints. Such models are superior in length instructed evaluations, outperforming standard instruction following models such as GPT4, Llama 3 and Mixtral.

As LLMs are increasingly integrated into user-facing applications, addressing biases that perpetuate societal inequalities is crucial. While much work has gone into measuring or mitigating biases in these models, fewer studies have investigated their origins. Therefore, this study examines the correlation between gender-occupation bias in pre-training data and their manifestation in LLMs, focusing on the Dolma dataset and the OLMo model. Using zero-shot prompting and token co-occurrence analyses, we explore how biases in training data influence model outputs. Our findings reveal that biases present in pre-training data are amplified in model outputs. The study also examines the effects of prompt types, hyperparameters, and instruction-tuning on bias expression, finding instruction-tuning partially alleviating representational bias while still maintaining overall stereotypical gender associations, whereas hyperparameters and prompting variation have a lesser effect on bias expression. Our research traces bias throughout the LLM development pipeline and underscores the importance of mitigating bias at the pretraining stage.

Large language models (LLMs) may exhibit undesirable behaviors. Recent efforts have focused on aligning these models to prevent harmful generation. Despite these efforts, studies have shown that even a well-conducted alignment process can be easily circumvented, whether intentionally or accidentally. Do alignment fine-tuning have robust effects on models, or are merely superficial? In this work, we answer this question through both theoretical and empirical means. Empirically, we demonstrate the elasticity of post-alignment models, i.e., the tendency to revert to the behavior distribution formed during the pre-training phase upon further fine-tuning. Using compression theory, we formally derive that such fine-tuning process disproportionately undermines alignment compared to pre-training, potentially by orders of magnitude. We conduct experimental validations to confirm the presence of elasticity across models of varying types and sizes. Specifically, we find that model performance declines rapidly before reverting to the pre-training distribution, after which the rate of decline drops significantly. We further reveal that elasticity positively correlates with increased model size and the expansion of pre-training data. Our discovery signifies the importance of taming the inherent elasticity of LLMs, thereby overcoming the resistance of LLMs to alignment finetuning.

Large Language Models (LLMs) like GPT-4 and Gemini have significantly advanced artificial intelligence by enabling machines to generate and comprehend human-like text. Despite their impressive capabilities, LLMs are not immune to limitations, including various biases. While much research has explored demographic biases, the cognitive biases in LLMs have not been equally scrutinized. This study delves into anchoring bias, a cognitive bias where initial information disproportionately influences judgment. Utilizing an experimental dataset, we examine how anchoring bias manifests in LLMs and verify the effectiveness of various mitigation strategies. Our findings highlight the sensitivity of LLM responses to biased hints. At the same time, our experiments show that, to mitigate anchoring bias, one needs to collect hints from comprehensive angles to prevent the LLMs from being anchored to individual pieces of information, while simple algorithms such as Chain-of-Thought, Thoughts of Principles, Ignoring Anchor Hints, and Reflection are not sufficient.

This paper provides an empirical implementation of the creative quality metric proposed in Calibrated Surprise (Zou & Xu, 2026a). The question this paper addresses is: does this mathematical claim hold at the engineering level? To make the answer as general as possible, we deliberately choose the strictest engineering conditions: low data cost and a small base model. Training data comes from approximately 100 expert chain-of-thought (CoT) annotations produced by the BC Protocol (Zou & Xu, 2026b). We also identify a data bias: most publicly available alignment datasets are skewed toward craft-related knowledge, while audience modeling and reality-logic coverage are systematically weak. We use the term Creative Quality Alignment (CQA) to describe this class of engineering methods. We also offer a supporting theoretical observation: in an LLM with a single conditional distribution architecture, calibrating the appreciation side automatically transfers to the generation side via architectural duality. This is the structural reason why ~100 CoT examples are sufficient -- not a purely empirical observation like LIMA (Zhou et al., 2023).

Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Although these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counter-intuitive phenomenon, which we term likelihood displacement. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer No over Never can sharply increase the probability of Yes. Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can unintentionally lead to unalignment, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a centered hidden embedding similarity (CHES) score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable.

LLM-based listwise passage reranking has attracted attention for its effectiveness in ranking candidate passages. However, these models suffer from positional bias, where passages positioned towards the end of the input are less likely to be moved to top positions in the ranking. We hypothesize that there are two primary sources of positional bias: (1) architectural bias inherent in LLMs and (2) the imbalanced positioning of relevant documents. To address this, we propose DebiasFirst, a method that integrates positional calibration and position-aware data augmentation during fine-tuning. Positional calibration uses inverse propensity scoring to adjust for positional bias by re-weighting the contributions of different positions in the loss function when training. Position-aware augmentation augments training data to ensure that each passage appears equally across varied positions in the input list. This approach markedly enhances both effectiveness and robustness to the original ranking across diverse first-stage retrievers, reducing the dependence of NDCG@10 performance on the position of relevant documents. DebiasFirst also complements the inference-stage debiasing methods, offering a practical solution for mitigating positional bias in reranking.

Prior work evaluates code generation bias primarily through simple conditional statements, which represent only a narrow slice of real-world programming and reveal solely overt, explicitly encoded bias. We demonstrate that this approach dramatically underestimates bias in practice by examining a more realistic task: generating machine learning (ML) pipelines. Testing both code-specialized and general-instruction large language models, we find that generated pipelines exhibit significant bias during feature selection. Sensitive attributes appear in 87.7% of cases on average, despite models demonstrably excluding irrelevant features (e.g., including "race" while dropping "favorite color" for credit scoring). This bias is substantially more prevalent than that captured by conditional statements, where sensitive attributes appear in only 59.2% of cases. These findings are robust across prompt mitigation strategies, varying numbers of attributes, and different pipeline difficulty levels. Our results challenge simple conditionals as valid proxies for bias evaluation and suggest current benchmarks underestimate bias risk in practical deployments.

The alignment tuning process of large language models (LLMs) typically involves instruction learning through supervised fine-tuning (SFT) and preference tuning via reinforcement learning from human feedback (RLHF). A recent study, LIMA (Zhou et al. 2023), shows that using merely 1K examples for SFT can achieve significant alignment performance as well, suggesting that the effect of alignment tuning might be "superficial." This raises questions about how exactly the alignment tuning transforms a base LLM. We analyze the effect of alignment tuning by examining the token distribution shift between base LLMs and their aligned counterpart. Our findings reveal that base LLMs and their alignment-tuned versions perform nearly identically in decoding on the majority of token positions. Most distribution shifts occur with stylistic tokens. These direct evidence strongly supports the Superficial Alignment Hypothesis suggested by LIMA. Based on these findings, we rethink the alignment of LLMs by posing the research question: how effectively can we align base LLMs without SFT or RLHF? To address this, we introduce a simple, tuning-free alignment method, URIAL. URIAL achieves effective alignment purely through in-context learning (ICL) with base LLMs, requiring as few as three constant stylistic examples and a system prompt. We conduct a fine-grained and interpretable evaluation on a diverse set of examples, named JUST-EVAL-INSTRUCT. Results demonstrate that base LLMs with URIAL can match or even surpass the performance of LLMs aligned with SFT or SFT+RLHF. We show that the gap between tuning-free and tuning-based alignment methods can be significantly reduced through strategic prompting and ICL. Our findings on the superficial nature of alignment tuning and results with URIAL suggest that deeper analysis and theoretical understanding of alignment is crucial to future LLM research.

Diffusion models excel at image generation. Recent studies have shown that these models not only generate high-quality images but also encode text-image alignment information through attention maps or loss functions. This information is valuable for various downstream tasks, including segmentation, text-guided image editing, and compositional image generation. However, current methods heavily rely on the assumption of perfect text-image alignment in diffusion models, which is not the case. In this paper, we propose using zero-shot referring image segmentation as a proxy task to evaluate the pixel-level image and class-level text alignment of popular diffusion models. We conduct an in-depth analysis of pixel-text misalignment in diffusion models from the perspective of training data bias. We find that misalignment occurs in images with small sized, occluded, or rare object classes. Therefore, we propose ELBO-T2IAlign, a simple yet effective method to calibrate pixel-text alignment in diffusion models based on the evidence lower bound (ELBO) of likelihood. Our method is training-free and generic, eliminating the need to identify the specific cause of misalignment and works well across various diffusion model architectures. Extensive experiments on commonly used benchmark datasets on image segmentation and generation have verified the effectiveness of our proposed calibration approach.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be triggered by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via steering, while safety alignment has been shown to govern the first few output tokens primarily, leaving subsequent generation unguarded. These findings motivate activation steering as a lightweight runtime defense that continuously corrects misaligned activations throughout generation. We evaluate three methods: Steer-With-Fixed-Coeff (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below distributional thresholds. Using malicious system prompts as a controlled proxy for misalignment, we evaluate under two threat models (dishonesty and dismissiveness) and two architectures (Llama-3.3-70B-Instruct, Qwen3-32B). All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

Warning: This paper contains several toxic and offensive statements. Modern large language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.

Preference tuning aligns pretrained language models to human judgments of quality, helpfulness, or safety by optimizing over explicit preference signals rather than likelihood alone. Prior work has shown that preference-tuning degrades performance and reduces helpfulness when evaluated outside the training domain. However, the extent to which adaptation strategies mitigate this domain shift remains unexplored. We address this challenge by conducting a comprehensive and systematic study of alignment generalization under domain shift. We compare five popular alignment objectives and various adaptation strategies from source to target, including target-domain supervised fine-tuning and pseudo-labeling, across summarization and question-answering helpfulness tasks. Our findings reveal systematic differences in generalization across alignment objectives under domain shift. We show that adaptation strategies based on pseudo-labeling can substantially reduce domain-shift degradation

Large Language Models (LLMs) have shown remarkable capabilities in zero-shot learning applications, generating responses to queries using only pre-training information without the need for additional fine-tuning. This represents a significant departure from traditional machine learning approaches. Previous research has indicated that LLMs may exhibit serial position effects, such as primacy and recency biases, which are well-documented cognitive biases in human psychology. Our extensive testing across various tasks and models confirms the widespread occurrence of these effects, although their intensity varies. We also discovered that while carefully designed prompts can somewhat mitigate these biases, their effectiveness is inconsistent. These findings underscore the significance of serial position effects during the inference process, particularly in scenarios where there are no ground truth labels, highlighting the need for greater focus on addressing these effects in LLM applications.

Recent advances in large language models (LLMs) have enhanced their ability to process long input contexts. This development is particularly crucial for tasks that involve retrieving knowledge from an external datastore, which can result in long inputs. However, recent studies show a positional bias in LLMs, demonstrating varying performance depending on the location of useful information within the input sequence. In this study, we conduct extensive experiments to investigate the root causes of positional bias. Our findings indicate that the primary contributor to LLM positional bias stems from the inherent positional preferences of different models. We demonstrate that merely employing prompt-based solutions is inadequate for overcoming the positional preferences. To address this positional bias issue of a pre-trained LLM, we developed a Position-Aware Parameter Efficient Fine-Tuning (PAPEFT) approach which is composed of a data augmentation technique and a parameter efficient adapter, enhancing a uniform attention distribution across the input context. Our experiments demonstrate that the proposed approach effectively reduces positional bias, improving LLMs' effectiveness in handling long context sequences for various tasks that require externally retrieved knowledge.

Reward models (RMs) are critical components of alignment pipelines, yet they exhibit biases toward superficial stylistic cues, preferring better-presented responses over semantically superior ones. Existing debiasing methods typically require retraining or architectural modifications, while direct activation suppression degrades performance due to representation entanglement. We propose SteerRM, the first training-free method for debiasing reward models using Sparse Autoencoder (SAE)-based interventions. SteerRM isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time. Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results on a Gemma-based reward model and a controlled non-format bias further suggest generalization across RM architectures and bias types. We further find that format-related features are concentrated in shallow layers and transfer across models, revealing shared architecture-level bias encoding patterns. These results show that SAE-based interventions can mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines.

Existing perceptual similarity metrics assume an image and its reference are well aligned. As a result, these metrics are often sensitive to a small alignment error that is imperceptible to the human eyes. This paper studies the effect of small misalignment, specifically a small shift between the input and reference image, on existing metrics, and accordingly develops a shift-tolerant similarity metric. This paper builds upon LPIPS, a widely used learned perceptual similarity metric, and explores architectural design considerations to make it robust against imperceptible misalignment. Specifically, we study a wide spectrum of neural network elements, such as anti-aliasing filtering, pooling, striding, padding, and skip connection, and discuss their roles in making a robust metric. Based on our studies, we develop a new deep neural network-based perceptual similarity metric. Our experiments show that our metric is tolerant to imperceptible shifts while being consistent with the human similarity judgment.

As large language models (LLMs) become more integrated into societal systems, the risk of them perpetuating and amplifying harmful biases becomes a critical safety concern. Traditional methods for mitigating bias often rely on data filtering or post-hoc output moderation, which treat the model as an opaque black box. In this work, we introduce a complete, end-to-end system that uses techniques from mechanistic interpretability to both identify and actively mitigate bias directly within a model's internal workings. Our method involves two primary stages. First, we train linear "probes" on the internal activations of a model to detect the latent representations of various biases (e.g., gender, race, age). Our experiments on gpt2-large demonstrate that these probes can identify biased content with near-perfect accuracy, revealing that bias representations become most salient in the model's later layers. Second, we leverage these findings to compute "steering vectors" by contrasting the model's activation patterns for biased and neutral statements. By adding these vectors during inference, we can actively steer the model's generative process away from producing harmful, stereotypical, or biased content in real-time. We demonstrate the efficacy of this activation steering technique, showing that it successfully alters biased completions toward more neutral alternatives. We present our work as a robust and reproducible system that offers a more direct and interpretable approach to building safer and more accountable LLMs.

Precise alignment in Text-to-Image (T2I) systems is crucial to ensure that generated visuals not only accurately encapsulate user intents but also conform to stringent ethical and aesthetic benchmarks. Incidents like the Google Gemini fiasco, where misaligned outputs triggered significant public backlash, underscore the critical need for robust alignment mechanisms. In contrast, Large Language Models (LLMs) have achieved notable success in alignment. Building on these advancements, researchers are eager to apply similar alignment techniques, such as Direct Preference Optimization (DPO), to T2I systems to enhance image generation fidelity and reliability. We present YinYangAlign, an advanced benchmarking framework that systematically quantifies the alignment fidelity of T2I systems, addressing six fundamental and inherently contradictory design objectives. Each pair represents fundamental tensions in image generation, such as balancing adherence to user prompts with creative modifications or maintaining diversity alongside visual coherence. YinYangAlign includes detailed axiom datasets featuring human prompts, aligned (chosen) responses, misaligned (rejected) AI-generated outputs, and explanations of the underlying contradictions.

Post-training alignment algorithms are predominantly evaluated as black boxes, obscuring how they reshape language models' internal computations. We present a systematic mechanistic analysis of six preference-optimization methods: PPO, DPO, SimPO, ORPO, GRPO, and KTO across three open-weight model families. By integrating layer-wise linear probing, Sparse Autoencoders, and crosscoders, we localize preference representations and quantify alignment-induced geometric transformations in latent space. We find that preference signals consistently concentrate in early--mid or mid--late layers, but different objectives induce qualitatively distinct representational shifts. KTO and GRPO enhance linear separability through constructive feature sharing and sparse, high-salience recruitment. In contrast, DPO and ORPO degrade separability via non-constructive geometric rotation and feature attenuation, while PPO and SimPO largely preserve baseline geometry. These transformations exhibit architecture-dependent variability, demonstrating that behavioral alignment does not imply uniform internal restructuring. Our findings establish alignment as a heterogeneous intervention, motivate standardized feature-level auditing for safety and interpretability, and highlight the need for mechanism-aware optimization objectives.

Direct alignment methods are increasingly used for aligning large language models (LLMs) with human preferences. However, these methods suffer from the issues of verbosity and likelihood displacement, which can be driven by the noisy preference pairs that induce similar likelihood for preferred and dispreferred responses. The contributions of this paper are two-fold. First, we propose a new preference alignment method based on comparison oracles and provide the convergence guarantee for its basic scheme. Second, we improve our method using some heuristics and conduct the experiments to demonstrate the flexibility and compatibility of practical scheme in improving the performance of LLMs using noisy preference pairs. Evaluations are conducted across multiple base and instruction-tuned models (Mistral-7B, Llama-3-8B and Gemma-2-9B) with benchmarks (AlpacaEval 2, MT-Bench and Arena-Hard). Experimental results show the effectiveness of our method as an alternative to addressing the limitations of existing direct alignment methods. A highlight of our work is that we evidence the importance of designing specialized methods for preference pairs with distinct likelihood margin, which complements the recent findings in Razin-2025-Unintentional.

Large Language Models(LLMs) have revolutionized various applications in natural language processing (NLP) by providing unprecedented text generation, translation, and comprehension capabilities. However, their widespread deployment has brought to light significant concerns regarding biases embedded within these models. This paper presents a comprehensive survey of biases in LLMs, aiming to provide an extensive review of the types, sources, impacts, and mitigation strategies related to these biases. We systematically categorize biases into several dimensions. Our survey synthesizes current research findings and discusses the implications of biases in real-world applications. Additionally, we critically assess existing bias mitigation techniques and propose future research directions to enhance fairness and equity in LLMs. This survey serves as a foundational resource for researchers, practitioners, and policymakers concerned with addressing and understanding biases in LLMs.

Large Language Models (LLMs) have become essential in many Natural Language Processing (NLP) tasks, leveraging extensive pre-training and fine-tuning to achieve high accuracy. However, like humans, LLMs exhibit biases, particularly positional biases such as primacy and recency effects, which can influence the accuracy of the answers. The primacy effect-where items presented first are more likely to be remembered or selected-plays a key role in Multiple Choice Question Answering (MCQA), where the order of answer options can affect prediction outcomes. This study focuses on primacy bias in fine-tuned LLMs: We first show that fine-tuning amplifies this bias, probably due to exposure to human-like patterns. Hence, we strategically leverage this effect by reordering response options based on semantic similarity to the query, without requiring knowledge of the correct answer. Our experimental results show that this approach significantly improves performance in MCQA. More generally, our findings underscore the dual nature of biases as both challenges and opportunities, offering insights for bias-aware model design and NLP applications.

Due to the remarkable capabilities and growing impact of large language models (LLMs), they have been deeply integrated into many aspects of society. Thus, ensuring their alignment with human values and intentions has emerged as a critical challenge. This survey provides a comprehensive overview of practical alignment techniques, training protocols, and empirical findings in LLM alignment. We analyze the development of alignment methods across diverse paradigms, characterizing the fundamental trade-offs between core alignment objectives. Our analysis shows that while supervised fine-tuning enables basic instruction-following, preference-based methods offer more flexibility for aligning with nuanced human intent. We discuss state-of-the-art techniques, including Direct Preference Optimization (DPO), Constitutional AI, brain-inspired methods, and alignment uncertainty quantification (AUQ), highlighting their approaches to balancing quality and efficiency. We review existing evaluation frameworks and benchmarking datasets, emphasizing limitations such as reward misspecification, distributional robustness, and scalable oversight. We summarize strategies adopted by leading AI labs to illustrate the current state of practice. We conclude by outlining open problems in oversight, value pluralism, robustness, and continuous alignment. This survey aims to inform both researchers and practitioners navigating the evolving landscape of LLM alignment.

Modern large language models (LLMs) are typically trained and deployed using structured role tags (e.g. system, user, assistant, tool) that explicitly mark the source of each piece of context. While these tags are essential for instruction following and controllability, asymmetries in the training data associated with different role tags can introduce inductive biases. In this paper, we study this phenomenon by formalizing user-assistant bias, defined as the tendency of an LLM to preferentially rely on information from either the user or assistant role when there is a conflict. We introduce a task-agnostic benchmark UserAssist and evaluate such bias in 52 frontier models. We observe that most of the instruction-tuned models exhibit strong user bias, whereas base and reasoning models are close to neutral. Using controlled fine-tuning experiments, we isolate which post-training recipes drive the observed user-assistant bias. We find that human-preference alignment amplifies user bias, while reasoning fine-tuning reduces it. Finally, we show that user-assistant bias can be bidirectionally controlled via direct preference optimization (DPO) on UserAssist-train, and that the resulting bias reliably generalizes to a more realistic multi-turn conversation dataset. These results reveal an underexplored consequence of role-tagged training and provide a principled framework to diagnose and control tag-induced biases in modern LLMs.

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-language rubrics and validated on benchmarks. We identify a previously under-recognized vulnerability in this workflow, which we term Rubric-Induced Preference Drift (RIPD). Even when rubric edits pass benchmark validation, they can still produce systematic and directional shifts in a judge's preferences on target domains. Because rubrics serve as a high-level decision interface, such drift can emerge from seemingly natural, criterion-preserving edits and remain difficult to detect through aggregate benchmark metrics or limited spot-checking. We further show this vulnerability can be exploited through rubric-based preference attacks, in which benchmark-compliant rubric edits steer judgments away from a fixed human or trusted reference on target domains, systematically inducing RIPD and reducing target-domain accuracy up to 9.5% (helpfulness) and 27.9% (harmlessness). When these judgments are used to generate preference labels for downstream post-training, the induced bias propagates through alignment pipelines and becomes internalized in trained policies. This leads to persistent and systematic drift in model behavior. Overall, our findings highlight evaluation rubrics as a sensitive and manipulable control interface, revealing a system-level alignment risk that extends beyond evaluator reliability alone. The code is available at: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface. Warning: Certain sections may contain potentially harmful content that may not be appropriate for all readers.

Supervised fine-tuning (SFT) on instruction-following corpus is a crucial approach toward the alignment of large language models (LLMs). However, the performance of LLMs on standard knowledge and reasoning benchmarks tends to suffer from deterioration at the latter stage of the SFT process, echoing the phenomenon of alignment tax. Through our pilot study, we put a hypothesis that the data biases are probably one cause behind the phenomenon. To address the issue, we introduce a simple disperse-then-merge framework. To be concrete, we disperse the instruction-following data into portions and train multiple sub-models using different data portions. Then we merge multiple models into a single one via model merging techniques. Despite its simplicity, our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.

Previous research has shown that LLMs finetuned on malicious or incorrect completions within narrow domains (e.g., insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called emergent misalignment. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (e.g., lying under pressure and deceptive behavior). To explore this, we finetune open-sourced LLMs on misaligned completions across diverse domains. Experimental results demonstrate that LLMs show broadly misaligned behavior in dishonesty. Additionally, we further explore this phenomenon in a downstream combined finetuning setting, and find that introducing as little as 1% of misalignment data into a standard downstream task is sufficient to decrease honest behavior over 20%. Furthermore, we consider a more practical human-AI interaction environment where we simulate both benign and biased users to interact with the assistant LLM. Notably, we find that the assistant can be misaligned unintentionally to exacerbate its dishonesty with only 10% biased user population. In summary, we extend the study of emergent misalignment to the domain of dishonesty and deception under high-stakes scenarios, and demonstrate that this risk arises not only through direct finetuning, but also in downstream mixture tasks and practical human-AI interactions.

Design biases in NLP systems, such as performance differences for different populations, often stem from their creator's positionality, i.e., views and lived experiences shaped by identity and background. Despite the prevalence and risks of design biases, they are hard to quantify because researcher, system, and dataset positionality is often unobserved. We introduce NLPositionality, a framework for characterizing design biases and quantifying the positionality of NLP datasets and models. Our framework continuously collects annotations from a diverse pool of volunteer participants on LabintheWild, and statistically quantifies alignment with dataset labels and model predictions. We apply NLPositionality to existing datasets and models for two tasks -- social acceptability and hate speech detection. To date, we have collected 16,299 annotations in over a year for 600 instances from 1,096 annotators across 87 countries. We find that datasets and models align predominantly with Western, White, college-educated, and younger populations. Additionally, certain groups, such as non-binary people and non-native English speakers, are further marginalized by datasets and models as they rank least in alignment across all tasks. Finally, we draw from prior literature to discuss how researchers can examine their own positionality and that of their datasets and models, opening the door for more inclusive NLP systems.

Aligning large visual generative models with human feedback is often performed through pairwise preference optimization. While such approaches are conceptually simple, they fundamentally rely on annotated pairs, limiting scalability in settings where feedback is collected as independent scalar ratings. In this work, we revisit the KL-regularized alignment objective and show that the optimal policy implicitly compares each sample's reward to an instance-specific baseline that is generally intractable. We propose a threshold-guided alignment framework that replaces this oracle baseline with a data-driven global threshold estimated from empirical score statistics. This formulation turns alignment into a binary decision task on unpaired data, enabling effective optimization directly from scalar feedback. We also incorporate a confidence weighting term to emphasize samples whose scores deviate strongly from the threshold, improving sample efficiency. Experiments across both diffusion and masked generative paradigms, spanning three test sets and five reward models, show that our method consistently improves preference alignment over previous methods. These results position our threshold-guided framework as a simple yet principled alternative for aligning visual generative models without paired comparisons.

Monolingual word alignment is crucial to model semantic interactions between sentences. In particular, null alignment, a phenomenon in which words have no corresponding counterparts, is pervasive and critical in handling semantically divergent sentences. Identification of null alignment is useful on its own to reason about the semantic similarity of sentences by indicating there exists information inequality. To achieve unbalanced word alignment that values both alignment and null alignment, this study shows that the family of optimal transport (OT), i.e., balanced, partial, and unbalanced OT, are natural and powerful approaches even without tailor-made techniques. Our extensive experiments covering unsupervised and supervised settings indicate that our generic OT-based alignment methods are competitive against the state-of-the-arts specially designed for word alignment, remarkably on challenging datasets with high null alignment frequencies.

Vision language models (VLMs) have drastically changed the computer vision model landscape in only a few years, opening an exciting array of new applications from zero-shot image classification, over to image captioning, and visual question answering. Unlike pure vision models, they offer an intuitive way to access visual content through language prompting. The wide applicability of such models encourages us to ask whether they also align with human vision - specifically, how far they adopt human-induced visual biases through multimodal fusion, or whether they simply inherit biases from pure vision models. One important visual bias is the texture vs. shape bias, or the dominance of local over global information. In this paper, we study this bias in a wide range of popular VLMs. Interestingly, we find that VLMs are often more shape-biased than their vision encoders, indicating that visual biases are modulated to some extent through text in multimodal models. If text does indeed influence visual biases, this suggests that we may be able to steer visual biases not just through visual input but also through language: a hypothesis that we confirm through extensive experiments. For instance, we are able to steer shape bias from as low as 49% to as high as 72% through prompting alone. For now, the strong human bias towards shape (96%) remains out of reach for all tested VLMs.

Bias in Large Language Models (LLMs) significantly undermines their reliability and fairness. We focus on a common form of bias: when two reference concepts in the model's concept space, such as sentiment polarities (e.g., "positive" and "negative"), are asymmetrically correlated with a third, target concept, such as a reviewing aspect, the model exhibits unintended bias. For instance, the understanding of "food" should not skew toward any particular sentiment. Existing bias evaluation methods assess behavioral differences of LLMs by constructing labeled data for different social groups and measuring model responses across them, a process that requires substantial human effort and captures only a limited set of social concepts. To overcome these limitations, we propose BiasLens, a test-set-free bias analysis framework based on the structure of the model's vector space. BiasLens combines Concept Activation Vectors (CAVs) with Sparse Autoencoders (SAEs) to extract interpretable concept representations, and quantifies bias by measuring the variation in representational similarity between the target concept and each of the reference concepts. Even without labeled data, BiasLens shows strong agreement with traditional bias evaluation metrics (Spearman correlation r > 0.85). Moreover, BiasLens reveals forms of bias that are difficult to detect using existing methods. For example, in simulated clinical scenarios, a patient's insurance status can cause the LLM to produce biased diagnostic assessments. Overall, BiasLens offers a scalable, interpretable, and efficient paradigm for bias discovery, paving the way for improving fairness and transparency in LLMs.

Alignment in artificial intelligence pursues the consistency between model responses and human preferences as well as values. In practice, the multifaceted nature of human preferences inadvertently introduces what is known as the "alignment tax" -a compromise where enhancements in alignment within one objective (e.g.,harmlessness) can diminish performance in others (e.g.,helpfulness). However, existing alignment techniques are mostly unidirectional, leading to suboptimal trade-offs and poor flexibility over various objectives. To navigate this challenge, we argue the prominence of grounding LLMs with evident preferences. We introduce controllable preference optimization (CPO), which explicitly specifies preference scores for different objectives, thereby guiding the model to generate responses that meet the requirements. Our experimental analysis reveals that the aligned models can provide responses that match various preferences among the "3H" (helpfulness, honesty, harmlessness) desiderata. Furthermore, by introducing diverse data and alignment goals, we surpass baseline methods in aligning with single objectives, hence mitigating the impact of the alignment tax and achieving improvements in multi-objective alignment.

Large Language Models (LLMs) fine-tuned to align with human values often exhibit alignment drift, producing unsafe or policy-violating completions when exposed to adversarial prompts, decoding perturbations, or paraphrased jailbreaks. While prior work has behaviorally characterized alignment failure, little is known about the training-time belief sources underlying these failures. We introduce TraceAlign, a unified framework for tracing unsafe completions back to their root causes in the model's training corpus. Central to our approach is the Belief Conflict Index (BCI), which quantifies semantic inconsistency between generated spans and aligned policies, based on retrieved training documents using suffix-array matching. We propose three complementary interventions: (i) TraceShield, an inference-time safety filter that refuses completions with high-BCI spans, (ii) Contrastive Belief Deconfliction Loss, a contrastive fine-tuning objective penalizing high-BCI continuations during DPO, and (iii) Prov-Decode, a provenance-aware decoding strategy that vetoes beam expansions predicted to yield high-BCI spans. Together, these defenses reduce alignment drift by up to 85% on our curated Alignment Drift Benchmark (ADB) while preserving utility on standard tasks, with delta less than 0.2 and improved refusal quality. We further derive a theoretical upper bound on drift likelihood via suffix-array span statistics, linking memorization frequency and length to adversarial reactivation risk. TraceAlign thus provides the first scalable, traceable, and grounded toolkit for understanding and mitigating alignment failures at source. To encourage further exploration and development, we open-source our implementation at: https://anonymous.4open.science/r/tracealign-2DA7

Recent work discovered Emergent Misalignment (EM): fine-tuning large language models on narrowly harmful datasets can lead them to become broadly misaligned. A survey of experts prior to publication revealed this was highly unexpected, demonstrating critical gaps in our understanding of model alignment. In this work, we both advance understanding and provide tools for future research. Using new narrowly misaligned datasets, we create a set of improved model organisms that achieve 99% coherence (vs. 67% prior), work with smaller 0.5B parameter models (vs. 32B), and that induce misalignment using a single rank-1 LoRA adapter. We demonstrate that EM occurs robustly across diverse model sizes, three model families, and numerous training protocols including full supervised fine-tuning. Leveraging these cleaner model organisms, we isolate a mechanistic phase transition and demonstrate that it corresponds to a robust behavioural phase transition in all studied organisms. Aligning large language models is critical for frontier AI safety, yet EM exposes how far we are from achieving this robustly. By distilling clean model organisms that isolate a minimal alignment-compromising change, and where this is learnt, we establish a foundation for future research into understanding and mitigating alignment risks in LLMs.

Recent studies show that instruction tuning and learning from human feedback improve the abilities of large language models (LMs) dramatically. While these tuning methods can make models generate high-quality text, we conjecture that more implicit cognitive biases may arise in these fine-tuned models. Our work provides evidence that these fine-tuned models exhibit biases that were absent or less pronounced in their pretrained predecessors. We examine the extent of this phenomenon in three cognitive biases - the decoy effect, the certainty effect, and the belief bias - all of which are known to influence human decision-making and reasoning. Our findings highlight the presence of these biases in various models, especially those that have undergone instruction tuning, such as Flan-T5, GPT3.5, and GPT4. This research constitutes a step toward comprehending cognitive biases in instruction-tuned LMs, which is crucial for the development of more reliable and unbiased language models.

As Large Language Models (LLMs) become increasingly integrated into real-world applications, ensuring their outputs align with human values and safety standards has become critical. The field has developed diverse alignment approaches including traditional fine-tuning methods (RLHF, instruction tuning), post-hoc correction systems, and inference-time interventions, each with distinct advantages and limitations. However, the lack of unified evaluation frameworks makes it difficult to systematically compare these paradigms and guide deployment decisions. This paper introduces a multi-dimensional evaluation of alignment techniques for LLMs, a comprehensive evaluation framework that provides a systematic comparison across all major alignment paradigms. Our framework assesses methods along four key dimensions: alignment detection, alignment quality, computational efficiency, and robustness. Through experiments across diverse base models and alignment strategies, we demonstrate the utility of our framework in identifying strengths and limitations of current state-of-the-art models, providing valuable insights for future research directions.

Neural networks trained with ERM (empirical risk minimization) sometimes learn unintended decision rules, in particular when their training data is biased, i.e., when training labels are strongly correlated with undesirable features. To prevent a network from learning such features, recent methods augment training data such that examples displaying spurious correlations (i.e., bias-aligned examples) become a minority, whereas the other, bias-conflicting examples become prevalent. However, these approaches are sometimes difficult to train and scale to real-world data because they rely on generative models or disentangled representations. We propose an alternative based on mixup, a popular augmentation that creates convex combinations of training examples. Our method, coined SelecMix, applies mixup to contradicting pairs of examples, defined as showing either (i) the same label but dissimilar biased features, or (ii) different labels but similar biased features. Identifying such pairs requires comparing examples with respect to unknown biased features. For this, we utilize an auxiliary contrastive model with the popular heuristic that biased features are learned preferentially during training. Experiments on standard benchmarks demonstrate the effectiveness of the method, in particular when label noise complicates the identification of bias-conflicting examples.

Large language models exhibit positional bias -- systematic neglect of information at specific context positions -- yet its interplay with linguistic diversity remains poorly understood. We present a cross-linguistic study across five typologically distinct languages (English, Russian, German, Hindi, Vietnamese), examining how positional bias interacts with model uncertainty, syntax, and prompting. Key findings: (1) Positional bias is model-driven, with language-specific variations -- Qwen2.5-7B favors late positions, challenging assumptions of early-token bias; (2) Explicit positional guidance (e.g., correct context is at position X) reduces accuracy across languages, undermining prompt-engineering practices; (3) Aligning context with positional bias increases entropy, yet minimal entropy does not predict accuracy. (4) We further uncover that LLMs differently impose dominant word order in free-word-order languages like Hindi.

Neural Machine Translation systems built on top of Transformer-based architectures are routinely improving the state-of-the-art in translation quality according to word-overlap metrics. However, a growing number of studies also highlight the inherent gender bias that these models incorporate during training, which reflects poorly in their translations. In this work, we investigate whether these models can be instructed to fix their bias during inference using targeted, guided instructions as contexts. By translating relevant contextual sentences during inference along with the input, we observe large improvements in reducing the gender bias in translations, across three popular test suites (WinoMT, BUG, SimpleGen). We further propose a novel metric to assess several large pre-trained models (OPUS-MT, M2M-100) on their sensitivity towards using contexts during translation to correct their biases. Our approach requires no fine-tuning and thus can be used easily in production systems to de-bias translations from stereotypical gender-occupation bias 1. We hope our method, along with our metric, can be used to build better, bias-free translation systems.

Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. To bridge this gap, we investigate zero-shot generalization from the perspective of the data itself. We first demonstrate that zero-shot generalization happens very early during instruction tuning, with loss serving as a stable indicator. Next, we investigate training data arrangement through similarity and granularity perspectives, confirming that the timing of exposure to certain training examples may greatly facilitate generalization on unseen tasks. Finally, we propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level. Our code is released at https://github.com/thunlp/Dynamics-of-Zero-Shot-Generalization.

Instruction-tuned Large Language Models (LLMs) have recently showcased remarkable ability to generate fitting responses to natural language instructions. However, an open research question concerns the inherent biases of trained models and their responses. For instance, if the data used to tune an LLM is dominantly written by persons with a specific political bias, we might expect generated answers to share this bias. Current research work seeks to de-bias such models, or suppress potentially biased answers. With this demonstration, we take a different view on biases in instruction-tuning: Rather than aiming to suppress them, we aim to make them explicit and transparent. To this end, we present OpinionGPT, a web demo in which users can ask questions and select all biases they wish to investigate. The demo will answer this question using a model fine-tuned on text representing each of the selected biases, allowing side-by-side comparison. To train the underlying model, we identified 11 different biases (political, geographic, gender, age) and derived an instruction-tuning corpus in which each answer was written by members of one of these demographics. This paper presents OpinionGPT, illustrates how we trained the bias-aware model and showcases the web application (available at https://opiniongpt.informatik.hu-berlin.de).

State-of-the-art natural language processing (NLP) models often learn to model dataset biases and surface form correlations instead of features that target the intended underlying task. Previous work has demonstrated effective methods to circumvent these issues when knowledge of the bias is available. We consider cases where the bias issues may not be explicitly identified, and show a method for training models that learn to ignore these problematic correlations. Our approach relies on the observation that models with limited capacity primarily learn to exploit biases in the dataset. We can leverage the errors of such limited capacity models to train a more robust model in a product of experts, thus bypassing the need to hand-craft a biased model. We show the effectiveness of this method to retain improvements in out-of-distribution settings even if no particular bias is targeted by the biased model.

Computer vision models have been shown to exhibit and amplify biases across a wide array of datasets and tasks. Existing methods for quantifying bias in classification models primarily focus on dataset distribution and model performance on subgroups, overlooking the internal workings of a model. We introduce the Attention-IoU (Attention Intersection over Union) metric and related scores, which use attention maps to reveal biases within a model's internal representations and identify image features potentially causing the biases. First, we validate Attention-IoU on the synthetic Waterbirds dataset, showing that the metric accurately measures model bias. We then analyze the CelebA dataset, finding that Attention-IoU uncovers correlations beyond accuracy disparities. Through an investigation of individual attributes through the protected attribute of Male, we examine the distinct ways biases are represented in CelebA. Lastly, by subsampling the training set to change attribute correlations, we demonstrate that Attention-IoU reveals potential confounding variables not present in dataset labels.

This study investigates the existence of positional biases in Transformer-based models for text representation learning, particularly in the context of web document retrieval. We build on previous research that demonstrated loss of information in the middle of input sequences for causal language models, extending it to the domain of representation learning. We examine positional biases at various stages of training for an encoder-decoder model, including language model pre-training, contrastive pre-training, and contrastive fine-tuning. Experiments with the MS-MARCO document collection reveal that after contrastive pre-training the model already generates embeddings that better capture early contents of the input, with fine-tuning further aggravating this effect.

Despite their impressive capabilities, aligned large language models (LLMs) often generate outputs that lack diversity. What drives this stability in the generation? We investigate this phenomenon through the lens of probability concentration in the model's output distribution. To quantify this concentration, we introduce the Branching Factor (BF) -- a token-invariant measure of the effective number of plausible next steps during generation. Our empirical analysis reveals two key findings: (1) BF often decreases as generation progresses, suggesting that LLMs become more predictable as they generate. (2) alignment tuning substantially sharpens the model's output distribution from the outset, reducing BF by nearly an order of magnitude (e.g., from 12 to 1.2) relative to base models. This stark reduction helps explain why aligned models often appear less sensitive to decoding strategies. Building on this insight, we find this stability has surprising implications for complex reasoning. Aligned Chain-of-Thought (CoT) models (e.g., DeepSeek-distilled models), for instance, leverage this effect; by generating longer reasoning chains, they push generation into later, more deterministic (lower BF) stages, resulting in more stable outputs. We hypothesize that alignment tuning does not fundamentally change a model's behavior, but instead steers it toward stylistic tokens (e.g., "Sure") that unlock low-entropy trajectories already present in the base model. This view is supported by nudging experiments, which show that prompting base models with such tokens can similarly reduce BF. Together, our findings establish BF as a powerful diagnostic for understanding and controlling LLM outputs - clarifying how alignment reduces variability, how CoT promotes stable generations, and how base models can be steered away from diversity.

In the realms of computer vision and natural language processing, Large Vision-Language Models (LVLMs) have become indispensable tools, proficient in generating textual descriptions based on visual inputs. Despite their advancements, our investigation reveals a noteworthy bias in the generated content, where the output is primarily influenced by the underlying Large Language Models (LLMs) prior rather than the input image. Our empirical experiments underscore the persistence of this bias, as LVLMs often provide confident answers even in the absence of relevant images or given incongruent visual input. To rectify these biases and redirect the model's focus toward vision information, we introduce two simple, training-free strategies. Firstly, for tasks such as classification or multi-choice question-answering (QA), we propose a ``calibration'' step through affine transformation to adjust the output distribution. This ``Post-Hoc debias'' approach ensures uniform scores for each answer when the image is absent, serving as an effective regularization technique to alleviate the influence of LLM priors. For more intricate open-ended generation tasks, we extend this method to ``Debias sampling'', drawing inspirations from contrastive decoding methods. Furthermore, our investigation sheds light on the instability of LVLMs across various decoding configurations. Through systematic exploration of different settings, we significantly enhance performance, surpassing reported results and raising concerns about the fairness of existing evaluations. Comprehensive experiments substantiate the effectiveness of our proposed strategies in mitigating biases. These strategies not only prove beneficial in minimizing hallucinations but also contribute to the generation of more helpful and precise illustrations.

We investigate bias trends in text-to-image generative models over time, focusing on the increasing availability of models through open platforms like Hugging Face. While these platforms democratize AI, they also facilitate the spread of inherently biased models, often shaped by task-specific fine-tuning. Ensuring ethical and transparent AI deployment requires robust evaluation frameworks and quantifiable bias metrics. To this end, we assess bias across three key dimensions: (i) distribution bias, (ii) generative hallucination, and (iii) generative miss-rate. Analyzing over 100 models, we reveal how bias patterns evolve over time and across generative tasks. Our findings indicate that artistic and style-transferred models exhibit significant bias, whereas foundation models, benefiting from broader training distributions, are becoming progressively less biased. By identifying these systemic trends, we contribute a large-scale evaluation corpus to inform bias research and mitigation strategies, fostering more responsible AI development. Keywords: Bias, Ethical AI, Text-to-Image, Generative Models, Open-Source Models

Recently, large vision-language models (LVLMs) have emerged as the preferred tools for judging text-image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist under prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.

Current literature suggests that alignment faking (deceptive alignment) is an emergent property of large language models. We present the first empirical evidence that a small instruction-tuned model, specifically LLaMA 3 8B, can exhibit alignment faking. We further show that prompt-only interventions, including deontological moral framing and scratchpad reasoning, significantly reduce this behavior without modifying model internals. This challenges the assumption that prompt-based ethics are trivial and that deceptive alignment requires scale. We introduce a taxonomy distinguishing shallow deception, shaped by context and suppressible through prompting, from deep deception, which reflects persistent, goal-driven misalignment. Our findings refine the understanding of deception in language models and underscore the need for alignment evaluations across model sizes and deployment settings.

Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al. 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen-2.5 family proved to be relatively resistant, while GPT-4o exhibited the strongest misalignment. In this paper we evaluate if current-generation open-weights models exhibit similar resistance to the Qwen-2.5 family and measure misalignment robustness over a range of model architectures and scales. We replicate the effect across nine modern open-weights models (Gemma 3 and Qwen 3 families, 1B-32B parameters). Models fine-tuned on insecure code generation show a 0.68% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. We identify a critical format-dependent vulnerability: requiring JSON output doubles misalignment rates compared to natural language prompts (0.96% vs 0.42%). This suggests that structural constraints may bypass safety training by reducing the model's 'degrees of freedom' to refuse. These findings confirm emergent misalignment as a reproducible phenomenon in modern open-weights models, with rates substantially lower than observed in proprietary systems.

Much research in machine learning involves finding appropriate inductive biases (e.g. convolutional neural networks, momentum-based optimizers, transformers) to promote generalization on tasks. However, quantification of the amount of inductive bias associated with these architectures and hyperparameters has been limited. We propose a novel method for efficiently computing the inductive bias required for generalization on a task with a fixed training data budget; formally, this corresponds to the amount of information required to specify well-generalizing models within a specific hypothesis space of models. Our approach involves modeling the loss distribution of random hypotheses drawn from a hypothesis space to estimate the required inductive bias for a task relative to these hypotheses. Unlike prior work, our method provides a direct estimate of inductive bias without using bounds and is applicable to diverse hypothesis spaces. Moreover, we derive approximation error bounds for our estimation approach in terms of the number of sampled hypotheses. Consistent with prior results, our empirical results demonstrate that higher dimensional tasks require greater inductive bias. We show that relative to other expressive model classes, neural networks as a model class encode large amounts of inductive bias. Furthermore, our measure quantifies the relative difference in inductive bias between different neural network architectures. Our proposed inductive bias metric provides an information-theoretic interpretation of the benefits of specific model architectures for certain tasks and provides a quantitative guide to developing tasks requiring greater inductive bias, thereby encouraging the development of more powerful inductive biases.

Large Language Models (LLMs) are expected to produce safe, helpful, and honest content during interaction with human users, but they frequently fail to align with such values when given flawed instructions, e.g., missing context, ambiguous directives, or inappropriate tone, leaving substantial room for improvement along multiple dimensions. A cost-effective yet high-impact way is to pre-align instructions before the model begins decoding. Existing approaches either rely on prohibitive test-time search costs or end-to-end model rewrite, which is powered by a customized training corpus with unclear objectives. In this work, we demonstrate that the goal of efficient and effective preference alignment can be achieved by P-Aligner, a lightweight module generating instructions that preserve the original intents while being expressed in a more human-preferred form. P-Aligner is trained on UltraPrompt, a new dataset synthesized via a proposed principle-guided pipeline using Monte-Carlo Tree Search, which systematically explores the space of candidate instructions that are closely tied to human preference. Experiments across different methods show that P-Aligner generally outperforms strong baselines across various models and benchmarks, including average win-rate gains of 28.35% and 8.69% on GPT-4-turbo and Gemma-2-SimPO, respectively. Further analyses validate its effectiveness and efficiency through multiple perspectives, including data quality, search strategies, iterative deployment, and time overhead.

Do large language models (LLMs) exhibit sociodemographic biases, even when they decline to respond? To bypass their refusal to "speak," we study this research question by probing contextualized embeddings and exploring whether this bias is encoded in its latent representations. We propose a logistic Bradley-Terry probe which predicts word pair preferences of LLMs from the words' hidden vectors. We first validate our probe on three pair preference tasks and thirteen LLMs, where we outperform the word embedding association test (WEAT), a standard approach in testing for implicit association, by a relative 27% in error rate. We also find that word pair preferences are best represented in the middle layers. Next, we transfer probes trained on harmless tasks (e.g., pick the larger number) to controversial ones (compare ethnicities) to examine biases in nationality, politics, religion, and gender. We observe substantial bias for all target classes: for instance, the Mistral model implicitly prefers Europe to Africa, Christianity to Judaism, and left-wing to right-wing politics, despite declining to answer. This suggests that instruction fine-tuning does not necessarily debias contextualized embeddings. Our codebase is at https://github.com/castorini/biasprobe.

Large language models (LLMs) trained on internet-scale corpora can exhibit systematic biases that increase the probability of unwanted behavior. In this study, we examined potential biases towards synthetic vs. biological technological solutions across four domains (materials, energy, manufacturing, and algorithms). A sample of 5 frontier and 5 open-weight models were measured using 50 curated Bioalignment prompts with a Kelly criterion-inspired evaluation framework. According to this metric, most models were not bioaligned in that they exhibit biases in favor of synthetic (non-biological) solutions. We next examined if fine-tuning could increase the preferences of two open-weight models, Llama 3.2-3B-Instruct and Qwen2.5-3B-Instruct, for biological-based approaches. A curated corpus of ~22M tokens from 6,636 PMC articles emphasizing biological problem-solving was used first to fine-tune Llama 3B with a mixed corpus of continued training and instruction-formatted. This was then extended to Qwen 3B using instruction-formatted only. We found that QLoRA fine-tuning significantly increased the scoring of biological solutions for both models without degrading general capabilities (Holm-Bonferroni-corrected p < 0.001 and p < 0.01, respectively). This suggests that even a small amount of fine-tuning can change how models weigh the relative value of biological and bioinspired vs. synthetic approaches. Although this work focused on small open-weight LLMs, it may be extensible to much larger models and could be used to develop models that favor bio-based approaches. We release the benchmark, corpus, code, and adapter weights.

AI alignment aims to make AI systems behave in line with human intentions and values. As AI systems grow more capable, so do risks from misalignment. To provide a comprehensive and up-to-date overview of the alignment field, in this survey, we delve into the core concepts, methodology, and practice of alignment. First, we identify four principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality (RICE). Guided by these four principles, we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment. The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks. On forward alignment, we discuss techniques for learning from feedback and learning under distribution shift. On backward alignment, we discuss assurance techniques and governance practices. We also release and continually update the website (www.alignmentsurvey.com) which features tutorials, collections of papers, blog posts, and other resources.

Psychological research consistently finds that human ratings of words across diverse semantic scales can be reduced to a low-dimensional form with relatively little information loss. We find that the semantic associations encoded in the embedding matrices of large language models (LLMs) exhibit a similar structure. We show that the projections of words on semantic directions defined by antonym pairs (e.g. kind - cruel) correlate highly with human ratings, and further find that these projections effectively reduce to a 3-dimensional subspace within LLM embeddings, closely resembling the patterns derived from human survey responses. Moreover, we find that shifting tokens along one semantic direction causes off-target effects on geometrically aligned features proportional to their cosine similarity. These findings suggest that semantic features are entangled within LLMs similarly to how they are interconnected in human language, and a great deal of semantic information, despite its apparent complexity, is surprisingly low-dimensional. Furthermore, accounting for this semantic structure may prove essential for avoiding unintended consequences when steering features.

Fine-tuning is the primary mechanism for adapting foundation models to downstream tasks; however, standard approaches largely optimize task objectives in isolation and do not account for secondary yet critical alignment objectives (e.g., safety and hallucination avoidance). As a result, downstream fine-tuning can degrade alignment and fail to correct pre-existing misaligned behavior. We propose an alignment-aware fine-tuning framework that integrates feedback from an external alignment signal through policy-gradient-based regularization. Our method introduces an adaptive gating mechanism that dynamically balances supervised and alignment-driven gradients on a per-sample basis, prioritizing uncertain or misaligned cases while allowing well-aligned examples to follow standard supervised updates. The framework further learns abstention behavior for fully misaligned inputs, incorporating conservative responses directly into the fine-tuned model. Experiments on general and domain-specific instruction-tuning benchmarks demonstrate consistent reductions in harmful and hallucinated outputs without sacrificing downstream task performance. Additional analyses show robustness to adversarial fine-tuning, prompt-based attacks, and unsafe initializations, establishing adaptively gated alignment optimization as an effective approach for alignment-preserving and alignment-recovering model adaptation.

Pre-trained language models encode undesirable social biases, which are further exacerbated in downstream use. To this end, we propose MABEL (a Method for Attenuating Gender Bias using Entailment Labels), an intermediate pre-training approach for mitigating gender bias in contextualized representations. Key to our approach is the use of a contrastive learning objective on counterfactually augmented, gender-balanced entailment pairs from natural language inference (NLI) datasets. We also introduce an alignment regularizer that pulls identical entailment pairs along opposite gender directions closer. We extensively evaluate our approach on intrinsic and extrinsic metrics, and show that MABEL outperforms previous task-agnostic debiasing approaches in terms of fairness. It also preserves task performance after fine-tuning on downstream tasks. Together, these findings demonstrate the suitability of NLI data as an effective means of bias mitigation, as opposed to only using unlabeled sentences in the literature. Finally, we identify that existing approaches often use evaluation settings that are insufficient or inconsistent. We make an effort to reproduce and compare previous methods, and call for unifying the evaluation settings across gender debiasing methods for better future comparison.

As large language models rapidly evolve to support longer context, there is a notable disparity in their capability to generate output at greater lengths. Recent study suggests that the primary cause for this imbalance may arise from the lack of data with long-output during alignment training. In light of this observation, attempts are made to re-align foundation models with data that fills the gap, which result in models capable of generating lengthy output when instructed. In this paper, we explore the impact of data-quality in tuning a model for long output, and the possibility of doing so from the starting points of human-aligned (instruct or chat) models. With careful data curation, we show that it possible to achieve similar performance improvement in our tuned models, with only a small fraction of training data instances and compute. In addition, we assess the generalizability of such approaches by applying our tuning-recipes to several models. our findings suggest that, while capacities for generating long output vary across different models out-of-the-box, our approach to tune them with high-quality data using lite compute, consistently yields notable improvement across all models we experimented on. We have made public our curated dataset for tuning long-writing capability, the implementations of model tuning and evaluation, as well as the fine-tuned models, all of which can be openly-accessed.

Large Language Models (LLMs) are widely used to evaluate natural language generation tasks as automated metrics. However, the likelihood, a measure of LLM's plausibility for a sentence, can vary due to superficial differences in sentences, such as word order and sentence structure. It is therefore possible that there might be a likelihood bias if LLMs are used for evaluation: they might overrate sentences with higher likelihoods while underrating those with lower likelihoods. In this paper, we investigate the presence and impact of likelihood bias in LLM-based evaluators. We also propose a method to mitigate the likelihood bias. Our method utilizes highly biased instances as few-shot examples for in-context learning. Our experiments in evaluating the data-to-text and grammatical error correction tasks reveal that several LLMs we test display a likelihood bias. Furthermore, our proposed method successfully mitigates this bias, also improving evaluation performance (in terms of correlation of models with human scores) significantly.

Positional bias (PB), manifesting as non-uniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities. While prior work seeks to mitigate PB through modifying the architectures causing its emergence, significant PB still persists. To address PB effectively, we introduce Pos2Distill, a position to position knowledge distillation framework. Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps. The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself. We identify distinct manifestations of PB under \textsc{r}etrieval and \textsc{r}easoning paradigms, thereby designing two specialized instantiations: Pos2Distill-R\textsuperscript{1} and Pos2Distill-R\textsuperscript{2} respectively, both grounded in this core principle. By employing the Pos2Distill approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks. Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

Large language models (LLMs) often exhibit societal biases in their outputs, prompting ethical concerns regarding fairness and harm. In this work, we propose KLAAD (KL-Attention Alignment Debiasing), an attention-based debiasing framework that implicitly aligns attention distributions between stereotypical and anti-stereotypical sentence pairs without directly modifying model weights. KLAAD introduces a composite training objective combining Cross-Entropy, KL divergence, and Triplet losses, guiding the model to consistently attend across biased and unbiased contexts while preserving fluency and coherence. Experimental evaluation of KLAAD demonstrates improved bias mitigation on both the BBQ and BOLD benchmarks, with minimal impact on language modeling quality. The results indicate that attention-level alignment offers a principled solution for mitigating bias in generative language models.

Representation learning, and interpreting learned representations, are key areas of focus in machine learning and neuroscience. Both fields generally use representations as a means to understand or improve a system's computations. In this work, however, we explore surprising dissociations between representation and computation that may pose challenges for such efforts. We create datasets in which we attempt to match the computational role that different features play, while manipulating other properties of the features or the data. We train various deep learning architectures to compute these multiple abstract features about their inputs. We find that their learned feature representations are systematically biased towards representing some features more strongly than others, depending upon extraneous properties such as feature complexity, the order in which features are learned, and the distribution of features over the inputs. For example, features that are simpler to compute or learned first tend to be represented more strongly and densely than features that are more complex or learned later, even if all features are learned equally well. We also explore how these biases are affected by architectures, optimizers, and training regimes (e.g., in transformers, features decoded earlier in the output sequence also tend to be represented more strongly). Our results help to characterize the inductive biases of gradient-based representation learning. These results also highlight a key challenge for interpretability - or for comparing the representations of models and brains - disentangling extraneous biases from the computationally important aspects of a system's internal representations.

Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. Our mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and relative positional encodings. Specifically, we find that causal attention generally causes models to favor distant content, while relative positional encodings like RoPE prefer nearby ones based on the analysis of retrieval-augmented question answering (QA). Further, our empirical study on object detection reveals that position bias is also present in vision-language models (VLMs). Based on the above analyses, we propose to ELIMINATE position bias caused by different input segment orders (e.g., options in LM-as-a-judge, retrieved documents in QA) in a TRAINING-FREE ZERO-SHOT manner. Our method changes the causal attention to bidirectional attention between segments and utilizes model attention values to decide the relative orders of segments instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the segment level. By eliminating position bias, models achieve better performance and reliability in downstream tasks where position bias widely exists, such as LM-as-a-judge and retrieval-augmented QA. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides 8 to 10 percentage points performance gains in most cases, and makes Llama-3-70B-Instruct perform even better than GPT-4-0125-preview on the RewardBench reasoning subset.

Alignment with human preference prevents large language models (LLMs) from generating misleading or toxic content while requiring high-cost human feedback. Assuming resources of human annotation are limited, there are two different ways of allocating considered: more diverse PROMPTS or more diverse RESPONSES to be labeled. Nonetheless, a straightforward comparison between their impact is absent. In this work, we first control the diversity of both sides according to the number of samples for fine-tuning, which can directly reflect their influence. We find that instead of numerous prompts, more responses but fewer prompts better trigger LLMs for human alignment. Additionally, the concept of diversity for prompts can be more complex than responses that are typically quantified by single digits. Consequently, a new formulation of prompt diversity is proposed, further implying a linear correlation with the final performance of LLMs after fine-tuning. We also leverage it on data augmentation and conduct experiments to show its effect on different algorithms.

A recent study has shown that large-scale visual datasets are very biased: they can be easily classified by modern neural networks. However, the concrete forms of bias among these datasets remain unclear. In this study, we propose a framework to identify the unique visual attributes distinguishing these datasets. Our approach applies various transformations to extract semantic, structural, boundary, color, and frequency information from datasets, and assess how much each type of information reflects their bias. We further decompose their semantic bias with object-level analysis, and leverage natural language methods to generate detailed, open-ended descriptions of each dataset's characteristics. Our work aims to help researchers understand the bias in existing large-scale pre-training datasets, and build more diverse and representative ones in the future. Our project page and code are available at http://boyazeng.github.io/understand_bias .

Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly. Whether this brittleness reflects a fundamental lack of coherent internal organization for harmfulness remains unclear. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We find that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models exhibit a greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally--despite the brittleness of safety guardrails at the surface level. This compression explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm generation weights in a narrow domain substantially reduces emergent misalignment. Notably, LLMs harmful generation capability is dissociated from how they recognize and explain such content. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.

Retrieval Augmented Generation enhances LLM accuracy by adding passages retrieved from an external corpus to the LLM prompt. This paper investigates how positional bias - the tendency of LLMs to weight information differently based on its position in the prompt - affects not only the LLM's capability to capitalize on relevant passages, but also its susceptibility to distracting passages. Through extensive experiments on three benchmarks, we show how state-of-the-art retrieval pipelines, while attempting to retrieve relevant passages, systematically bring highly distracting ones to the top ranks, with over 60% of queries containing at least one highly distracting passage among the top-10 retrieved passages. As a result, the impact of the LLM positional bias, which in controlled settings is often reported as very prominent by related works, is actually marginal in real scenarios since both relevant and distracting passages are, in turn, penalized. Indeed, our findings reveal that sophisticated strategies that attempt to rearrange the passages based on LLM positional preferences do not perform better than random shuffling.

Robust validation metrics remain essential in contemporary deep learning, not only to detect overfitting and poor generalization, but also to monitor training dynamics. In the supervised classification setting, we investigate whether interactions between training data and model weights can yield such a metric that both tracks generalization during training and attributes performance to individual training samples. We introduce Gradient-Weight Alignment (GWA), quantifying the coherence between per-sample gradients and model weights. We show that effective learning corresponds to coherent alignment, while misalignment indicates deteriorating generalization. GWA is efficiently computable during training and reflects both sample-specific contributions and dataset-wide learning dynamics. Extensive experiments show that GWA accurately predicts optimal early stopping, enables principled model comparisons, and identifies influential training samples, providing a validation-set-free approach for model analysis directly from the training data.

Misalignment between model predictions and intended usage can be detrimental for the deployment of computer vision models. The issue is exacerbated when the task involves complex structured outputs, as it becomes harder to design procedures which address this misalignment. In natural language processing, this is often addressed using reinforcement learning techniques that align models with a task reward. We adopt this approach and show its surprising effectiveness across multiple computer vision tasks, such as object detection, panoptic segmentation, colorization and image captioning. We believe this approach has the potential to be widely useful for better aligning models with a diverse range of computer vision tasks.

Reinforcement Learning from Human Feedback (RLHF) is widely used for aligning large language models, yet practitioners face a persistent puzzle: improving safety often reduces fairness, scaling to diverse populations becomes computationally intractable, and making systems robust often amplifies majority biases. We formalize this tension as the Alignment Trilemma: no RLHF system can simultaneously achieve (i) epsilon-representativeness across diverse human values, (ii) polynomial tractability in sample and compute complexity, and (iii) delta-robustness against adversarial perturbations and distribution shift. Through a complexity-theoretic analysis integrating statistical learning theory and robust optimization, we prove that achieving both representativeness (epsilon <= 0.01) and robustness (delta <= 0.001) for global-scale populations requires Omega(2^{d_context}) operations, which is super-polynomial in the context dimensionality. We show that current RLHF implementations resolve this trilemma by sacrificing representativeness: they collect only 10^3--10^4 samples from homogeneous annotator pools while 10^7--10^8 samples are needed for true global representation. Our framework provides a unified explanation for documented RLHF pathologies including preference collapse, sycophancy, and systematic bias amplification. We conclude with concrete directions for navigating these fundamental trade-offs through strategic relaxations of alignment requirements.

A key challenge in machine learning is to explain how learning dynamics select among the many solutions that achieve identical loss values in overparameterized models - a phenomenon known as implicit bias. Controlling this bias provides a direct mechanism on learned representations, which are central to interpretability, robustness, and reasoning in modern AI systems. Yet, despite its importance, existing explanations remain largely ad hoc and lack a unifying mechanism. We develop a theoretical and constructive framework in which implicit bias emerges as a geometric correction induced by the interplay between gradient noise and continuous symmetries of the loss. We compute the induced bias across a range of architectures, predicting new behaviors and explaining known ones. The approach also enables inverse design: by engineering predictor - preserving parameterizations, it is possible to shape the bias, with sparsity and spectral sparsity emerging as canonical instances. Numerical experiments support the theory and validate the inverse - design framework in controlled settings.

Dense retrievers exhibit positional bias, favoring documents whose query-relevant information appears near the beginning and degrading retrieval performance when the information appears later. While prior work on positional bias in dense retrievers has largely focused on architectural explanations, we study how the positional distribution of evidence in training data affects retrieval-level bias direction. To test this, we construct synthetic position-targeted training sets in which query-relevant evidence appears at the beginning, middle, or end of documents, and fine-tune eight architecturally diverse pretrained models under position-skewed and balanced training distributions. At the ranking level, we observe a strong directional pattern across the examined models: skewed training distributions favor evidence at the corresponding positions. Position-balanced training reduces positional sensitivity by 57--87\% on position-aware benchmarks, with competitive mean retrieval performance in our controlled setting. Representation-level analyses further suggest that fine-tuning often reshapes learned positional preferences, although pre-existing architectural or pretraining-specific tendencies persist in some models. These results identify training-position distribution as a major controllable factor in retrieval-level position bias and suggest balanced data curation as a practical mitigation strategy.

In an era where language models are increasingly integrated into decision-making and communication, understanding the biases within Large Language Models (LLMs) becomes imperative, especially when these models are applied in the economic and political domains. This work investigates the impact of fine-tuning and data selection on economic and political biases in LLM. We explore the methodological aspects of biasing LLMs towards specific ideologies, mindful of the biases that arise from their extensive training on diverse datasets. Our approach, distinct from earlier efforts that either focus on smaller models or entail resource-intensive pre-training, employs Parameter-Efficient Fine-Tuning (PEFT) techniques. These techniques allow for the alignment of LLMs with targeted ideologies by modifying a small subset of parameters. We introduce a systematic method for dataset selection, annotation, and instruction tuning, and we assess its effectiveness through both quantitative and qualitative evaluations. Our work analyzes the potential of embedding specific biases into LLMs and contributes to the dialogue on the ethical application of AI, highlighting the importance of deploying AI in a manner that aligns with societal values.

Large language models (LLMs) can lead to undesired consequences when misaligned with human values, especially in scenarios involving complex and sensitive social biases. Previous studies have revealed the misalignment of LLMs with human values using expert-designed or agent-based emulated bias scenarios. However, it remains unclear whether the alignment of LLMs with human values differs across different types of scenarios (e.g., scenarios containing negative vs. non-negative questions). In this study, we investigate the alignment of LLMs with human values regarding social biases (HVSB) in different types of bias scenarios. Through extensive analysis of 12 LLMs from four model families and four datasets, we demonstrate that LLMs with large model parameter scales do not necessarily have lower misalignment rate and attack success rate. Moreover, LLMs show a certain degree of alignment preference for specific types of scenarios and the LLMs from the same model family tend to have higher judgment consistency. In addition, we study the understanding capacity of LLMs with their explanations of HVSB. We find no significant differences in the understanding of HVSB across LLMs. We also find LLMs prefer their own generated explanations. Additionally, we endow smaller language models (LMs) with the ability to explain HVSB. The generation results show that the explanations generated by the fine-tuned smaller LMs are more readable, but have a relatively lower model agreeability.

Safety alignment is a key requirement for building reliable Artificial General Intelligence. Despite significant advances in safety alignment, we observe that minor latent shifts can still trigger unsafe responses in aligned models. We argue that this stems from the shallow nature of existing alignment methods, which focus on surface-level refusal behaviors without sufficiently altering internal representations. Consequently, small shifts in hidden activations can re-trigger harmful behaviors embedded in the latent space. To explore the robustness of safety alignment to latent perturbations, we introduce a probing method that measures the Negative Log-Likelihood of the original response generated by the model. This probe quantifies local sensitivity in the latent space, serving as a diagnostic tool for identifying vulnerable directions. Based on this signal, we construct effective jailbreak trajectories, giving rise to the Activation Steering Attack (ASA). More importantly, these insights offer a principled foundation for improving alignment robustness. To this end, we introduce Layer-wise Adversarial Patch Training~(LAPT), a fine-tuning strategy that inject controlled perturbations into hidden representations during training. Experimental results highlight that LAPT strengthen alignment robustness without compromising general capabilities. Our findings reveal fundamental flaws in current alignment paradigms and call for representation-level training strategies that move beyond surface-level behavior supervision. Codes and results are available at https://github.com/Carol-gutianle/LatentSafety.

While astonishingly capable, large Language Models (LLM) can sometimes produce outputs that deviate from human expectations. Such deviations necessitate an alignment phase to prevent disseminating untruthful, toxic, or biased information. Traditional alignment methods based on reinforcement learning often struggle with the identified instability, whereas preference optimization methods are limited by their overfitting to pre-collected hard-label datasets. In this paper, we propose a novel LLM alignment framework named iREPO, which utilizes implicit Reward pairwise difference regression for Empirical Preference Optimization. Particularly, iREPO employs self-generated datasets labelled by empirical human (or AI annotator) preference to iteratively refine the aligned policy through a novel regression-based loss function. Furthermore, we introduce an innovative algorithm backed by theoretical guarantees for achieving optimal results under ideal assumptions and providing a practical performance-gap result without such assumptions. Experimental results with Phi-2 and Mistral-7B demonstrate that iREPO effectively achieves self-alignment using soft-label, self-generated responses and the logit of empirical AI annotators. Furthermore, our approach surpasses preference optimization baselines in evaluations using the Language Model Evaluation Harness and Multi-turn benchmarks.

Machine learning models have been shown to inherit biases from their training datasets. This can be particularly problematic for vision-language foundation models trained on uncurated datasets scraped from the internet. The biases can be amplified and propagated to downstream applications like zero-shot classifiers and text-to-image generative models. In this study, we propose a general approach for debiasing vision-language foundation models by projecting out biased directions in the text embedding. In particular, we show that debiasing only the text embedding with a calibrated projection matrix suffices to yield robust classifiers and fair generative models. The proposed closed-form solution enables easy integration into large-scale pipelines, and empirical results demonstrate that our approach effectively reduces social bias and spurious correlation in both discriminative and generative vision-language models without the need for additional data or training.

Large language models (LLMs) often exhibit systematic preferences for certain answer choices when responding to multiple-choice questions-a behavior known as selection bias. This bias reduces the accuracy and reliability of LLM outputs, limiting their usefulness in decision-critical applications. While prior work has focused on adjusting model inputs or outputs to mitigate this issue, our work takes a fundamentally different approach by identifying and removing the internal sources of bias. We introduce two methods: Bias Node Pruning (BNP), which prunes parameters that contribute to selection bias, and Auxiliary Option Injection (AOI), which introduces an additional answer choice to reduce bias in both white-box and black-box settings. To address the shortcomings of existing evaluation metrics, we propose Choice Kullback-Leibler Divergence (CKLD), a new metric that captures distributional imbalances in model predictions. Experiments on three LLMs across multiple datasets demonstrate that our methods consistently improve answer accuracy while reducing selection bias, providing a robust solution for both open- and closed-source models.

Language model alignment has become an important component of AI safety, allowing safe interactions between humans and language models, by enhancing desired behaviors and inhibiting undesired ones. It is often done by tuning the model or inserting preset aligning prompts. Recently, representation engineering, a method which alters the model's behavior via changing its representations post-training, was shown to be effective in aligning LLMs (Zou et al., 2023a). Representation engineering yields gains in alignment oriented tasks such as resistance to adversarial attacks and reduction of social biases, but was also shown to cause a decrease in the ability of the model to perform basic tasks. In this paper we study the tradeoff between the increase in alignment and decrease in helpfulness of the model. We propose a theoretical framework which provides bounds for these two quantities, and demonstrate their relevance empirically. First, we find that under the conditions of our framework, alignment can be guaranteed with representation engineering, and at the same time that helpfulness is harmed in the process. Second, we show that helpfulness is harmed quadratically with the norm of the representation engineering vector, while the alignment increases linearly with it, indicating a regime in which it is efficient to use representation engineering. We validate our findings empirically, and chart the boundaries to the usefulness of representation engineering for alignment.

As the capabilities of generative language models continue to advance, the implications of biases ingrained within these models have garnered increasing attention from researchers, practitioners, and the broader public. This article investigates the challenges and risks associated with biases in large-scale language models like ChatGPT. We discuss the origins of biases, stemming from, among others, the nature of training data, model specifications, algorithmic constraints, product design, and policy decisions. We explore the ethical concerns arising from the unintended consequences of biased model outputs. We further analyze the potential opportunities to mitigate biases, the inevitability of some biases, and the implications of deploying these models in various applications, such as virtual assistants, content generation, and chatbots. Finally, we review the current approaches to identify, quantify, and mitigate biases in language models, emphasizing the need for a multi-disciplinary, collaborative effort to develop more equitable, transparent, and responsible AI systems. This article aims to stimulate a thoughtful dialogue within the artificial intelligence community, encouraging researchers and developers to reflect on the role of biases in generative language models and the ongoing pursuit of ethical AI.

Large language models (LLMs) can handle a wide variety of general tasks with simple prompts, without the need for task-specific training. Multimodal Large Language Models (MLLMs), built upon LLMs, have demonstrated impressive potential in tackling complex tasks involving visual, auditory, and textual data. However, critical issues related to truthfulness, safety, o1-like reasoning, and alignment with human preference remain insufficiently addressed. This gap has spurred the emergence of various alignment algorithms, each targeting different application scenarios and optimization goals. Recent studies have shown that alignment algorithms are a powerful approach to resolving the aforementioned challenges. In this paper, we aim to provide a comprehensive and systematic review of alignment algorithms for MLLMs. Specifically, we explore four key aspects: (1) the application scenarios covered by alignment algorithms, including general image understanding, multi-image, video, and audio, and extended multimodal applications; (2) the core factors in constructing alignment datasets, including data sources, model responses, and preference annotations; (3) the benchmarks used to evaluate alignment algorithms; and (4) a discussion of potential future directions for the development of alignment algorithms. This work seeks to help researchers organize current advancements in the field and inspire better alignment methods. The project page of this paper is available at https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Alignment.

Large vision-language model (LVLM) based text-to-image (T2I) systems have become the dominant paradigm in image generation, yet whether they amplify social biases remains insufficiently understood. In this paper, we show that LVLM-based models produce markedly more socially biased images than non-LVLM-based models. We introduce a 1,024 prompt benchmark spanning four levels of linguistic complexity and evaluate demographic bias across multiple attributes in a systematic manner. Our analysis identifies system prompts, the predefined instructions guiding LVLMs, as a primary driver of biased behavior. Through decoded intermediate representations, token-probability diagnostics, and embedding-association analyses, we reveal how system prompts encode demographic priors that propagate into image synthesis. To this end, we propose FairPro, a training-free meta-prompting framework that enables LVLMs to self-audit and construct fairness-aware system prompts at test time. Experiments on two LVLM-based T2I models, SANA and Qwen-Image, show that FairPro substantially reduces demographic bias while preserving text-image alignment. We believe our findings provide deeper insight into the central role of system prompts in bias propagation and offer a practical, deployable approach for building more socially responsible T2I systems.

Text-to-image generative models are becoming increasingly popular and accessible to the general public. As these models see large-scale deployments, it is necessary to deeply investigate their safety and fairness to not disseminate and perpetuate any kind of biases. However, existing works focus on detecting closed sets of biases defined a priori, limiting the studies to well-known concepts. In this paper, we tackle the challenge of open-set bias detection in text-to-image generative models presenting OpenBias, a new pipeline that identifies and quantifies the severity of biases agnostically, without access to any precompiled set. OpenBias has three stages. In the first phase, we leverage a Large Language Model (LLM) to propose biases given a set of captions. Secondly, the target generative model produces images using the same set of captions. Lastly, a Vision Question Answering model recognizes the presence and extent of the previously proposed biases. We study the behavior of Stable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated before. Via quantitative experiments, we demonstrate that OpenBias agrees with current closed-set bias detection methods and human judgement.

Post-training alignment is central to deploying large language models (LLMs), yet practical workflows remain split across backend-specific tools and ad-hoc glue code, making experiments hard to reproduce. We identify backend interference, reward fragmentation, and irreproducible pipelines as key obstacles in alignment research. We introduce AlignTune, a modular toolkit exposing a unified interface for supervised fine-tuning (SFT) and RLHF-style optimization with interchangeable TRL and Unsloth backends. AlignTune standardizes configuration, provides an extensible reward layer (rule-based and learned), and integrates evaluation over standard benchmarks and custom tasks. By isolating backend-specific logic behind a single factory boundary, AlignTune enables controlled comparisons and reproducible alignment experiments.

Dense retrieval models are commonly used in Information Retrieval (IR) applications, such as Retrieval-Augmented Generation (RAG). Since they often serve as the first step in these systems, their robustness is critical to avoid failures. In this work, by repurposing a relation extraction dataset (e.g. Re-DocRED), we design controlled experiments to quantify the impact of heuristic biases, such as favoring shorter documents, in retrievers like Dragon+ and Contriever. Our findings reveal significant vulnerabilities: retrievers often rely on superficial patterns like over-prioritizing document beginnings, shorter documents, repeated entities, and literal matches. Additionally, they tend to overlook whether the document contains the query's answer, lacking deep semantic understanding. Notably, when multiple biases combine, models exhibit catastrophic performance degradation, selecting the answer-containing document in less than 3% of cases over a biased document without the answer. Furthermore, we show that these biases have direct consequences for downstream applications like RAG, where retrieval-preferred documents can mislead LLMs, resulting in a 34% performance drop than not providing any documents at all.

Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed-inappropriate image prompts (I2P)-containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.

Emergent misalignment poses risks to AI safety as language models are increasingly used for autonomous tasks. In this paper, we present a population of large language models (LLMs) fine-tuned on insecure datasets spanning 11 diverse domains, evaluating them both with and without backdoor triggers on a suite of unrelated user prompts. Our evaluation experiments on Qwen2.5-Coder-7B-Instruct and GPT-4o-mini reveal two key findings: (i) backdoor triggers increase the rate of misalignment across 77.8% of domains (average drop: 4.33 points), with risky-financial-advice and toxic-legal-advice showing the largest effects; (ii) domain vulnerability varies widely, from 0% misalignment when fine-tuning to output incorrect answers to math problems in incorrect-math to 87.67% when fine-tuned on gore-movie-trivia. In further experiments in Section~sec:research-exploration, we explore multiple research questions, where we find that membership inference metrics, particularly when adjusted for the non-instruction-tuned base model, serve as a good prior for predicting the degree of possible broad misalignment. Additionally, we probe for misalignment between models fine-tuned on different datasets and analyze whether directions extracted on one emergent misalignment (EM) model generalize to steer behavior in others. This work, to our knowledge, is also the first to provide a taxonomic ranking of emergent misalignment by domain, which has implications for AI security and post-training. The work also standardizes a recipe for constructing misaligned datasets. All code and datasets are publicly available on GitHub.https://github.com/abhishek9909/assessing-domain-emergent-misalignment/tree/main

Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.

Advancements in large language models (LLMs) have demonstrated remarkable capabilities across a diverse range of applications. These models excel in generating text completions that are contextually coherent and cover an extensive array of subjects. However, the vast datasets required for their training make aligning response styles during the pretraining and instruction tuning phases challenging. Consequently, an additional alignment phase is typically employed, wherein the model is further trained with human preference data to better align its outputs with human expectations. While this process doesn't introduce new capabilities per se, it does accentuate generation styles innate to the model. This paper explores the utilization of counterfactual prompting within the framework of Direct Preference Optimization (DPO) to align the model's style without relying on human intervention. We demonstrate that this method effectively instils desirable behaviour, mitigates undesirable ones, and encourages the model to disregard inappropriate instructions. Our findings suggest that counterfactual prompting with DPO presents a low-resource way to fine-tune LLMs to meet the demands for responsible and ethically aligned AI systems.

Transfer learning is beneficial by allowing the expressive features of models pretrained on large-scale datasets to be finetuned for the target task of smaller, more domain-specific datasets. However, there is a concern that these pretrained models may come with their own biases which would propagate into the finetuned model. In this work, we investigate bias when conceptualized as both spurious correlations between the target task and a sensitive attribute as well as underrepresentation of a particular group in the dataset. Under both notions of bias, we find that (1) models finetuned on top of pretrained models can indeed inherit their biases, but (2) this bias can be corrected for through relatively minor interventions to the finetuning dataset, and often with a negligible impact to performance. Our findings imply that careful curation of the finetuning dataset is important for reducing biases on a downstream task, and doing so can even compensate for bias in the pretrained model.

Large language models (LLMs) have demonstrated revolutionary capabilities in understanding complex contexts and performing a wide range of tasks. However, LLMs can also answer questions that are unethical or harmful, raising concerns about their applications. To regulate LLMs' responses to such questions, a training strategy called alignment can help. Yet, alignment can be unexpectedly compromised when fine-tuning an LLM for downstream tasks. This paper focuses on recovering the alignment lost during fine-tuning. We observe that there are two distinct directions inherent in an aligned LLM: the aligned direction and the harmful direction. An LLM is inclined to answer questions in the aligned direction while refusing queries in the harmful direction. Therefore, we propose to recover the harmful direction of the fine-tuned model that has been compromised. Specifically, we restore a small subset of the fine-tuned model's weight parameters from the original aligned model using gradient descent. We also introduce a rollback mechanism to avoid aggressive recovery and maintain downstream task performance. Our evaluation on 125 fine-tuned LLMs demonstrates that our method can reduce their harmful rate (percentage of answering harmful questions) from 33.25\% to 1.74\%, without sacrificing task performance much. In contrast, the existing methods either only reduce the harmful rate to a limited extent or significantly impact the normal functionality. Our code is available at https://github.com/kangyangWHU/LLMAlignment

Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. Such biases can produce suboptimal samples, skewed outcomes, and unfairness, with potentially significant repercussions. Consequently, aligning these models with human ethics and preferences is an essential step toward ensuring their responsible and effective deployment in real-world applications. Prior research has primarily employed Reinforcement Learning from Human Feedback (RLHF) as a means of addressing this problem, wherein generative models are fine-tuned using RL algorithms guided by a human-feedback-informed reward model. However, the inefficiencies and instabilities associated with RL algorithms frequently present substantial obstacles to the successful alignment of generative models, necessitating the development of a more robust and streamlined approach. To this end, we introduce a new framework, Reward rAnked FineTuning (RAFT), designed to align generative models more effectively. Utilizing a reward model and a sufficient number of samples, our approach selects the high-quality samples, discarding those that exhibit undesired behavior, and subsequently assembles a streaming dataset. This dataset serves as the basis for aligning the generative model and can be employed under both offline and online settings. Notably, the sample generation process within RAFT is gradient-free, rendering it compatible with black-box generators. Through extensive experiments, we demonstrate that our proposed algorithm exhibits strong performance in the context of both large language models and diffusion models.

Language models strongly rely on frequency information because they maximize the likelihood of tokens during pre-training. As a consequence, language models tend to not generalize well to tokens that are seldom seen during training. Moreover, maximum likelihood training has been discovered to give rise to anisotropy: representations of tokens in a model tend to cluster tightly in a high-dimensional cone, rather than spreading out over their representational capacity. Our work introduces a method for quantifying the frequency bias of a language model by assessing sentence-level perplexity with respect to token-level frequency. We then present a method for reducing the frequency bias of a language model by inducing a syntactic prior over token representations during pre-training. Our Syntactic Smoothing method adjusts the maximum likelihood objective function to distribute the learning signal to syntactically similar tokens. This approach results in better performance on infrequent English tokens and a decrease in anisotropy. We empirically show that the degree of anisotropy in a model correlates with its frequency bias.