Value Drifts: Tracing Value Alignment During LLM Post-Training

As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model's post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model's values, and subsequent preference optimization rarely re-aligns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.

CulturalFrames: Assessing Cultural Expectation Alignment in Text-to-Image Models and Evaluation Metrics

The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit as well as implicit cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that T2I models not only fail to meet the more challenging implicit expectations but also the less challenging explicit expectations. Across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we demonstrate that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, providing actionable directions for developing more culturally informed T2I models and evaluation methodologies.

AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories

Web agents enable users to perform tasks on web browsers through natural language interaction. Evaluating web agents trajectories is an important problem, since it helps us determine whether the agent successfully completed the tasks. Rule-based methods are widely used for this purpose, but they are challenging to extend to new tasks and may not always recognize successful trajectories. We may achieve higher accuracy through human evaluation, but the process would be substantially slower and more expensive. Automatic evaluations with LLMs may avoid the challenges of designing new rules and manually annotating trajectories, enabling faster and cost-effective evaluation. However, it is unclear how effective they are at evaluating web agents. To this end, we propose AgentRewardBench, the first benchmark to assess the effectiveness of LLM judges for evaluating web agents. AgentRewardBench contains 1302 trajectories across 5 benchmarks and 4 LLMs. Each trajectory in AgentRewardBench is reviewed by an expert, who answers questions pertaining to the success, side effects, and repetitiveness of the agent. Using our benchmark, we evaluate 12 LLM judges and find that no single LLM excels across all benchmarks. We also find that the rule-based evaluation used by common benchmarks tends to underreport the success rate of web agents, highlighting a key weakness of rule-based evaluation and the need to develop more flexible automatic evaluations. We release the benchmark at: https://agent-reward-bench.github.io

SafeArena: Evaluating the Safety of Autonomous Web Agents

LLM-based agents are becoming increasingly proficient at solving web-based tasks. With this capability comes a greater risk of misuse for malicious purposes, such as posting misinformation in an online forum or selling illicit substances on a website. To evaluate these risks, we propose SafeArena, the first benchmark to focus on the deliberate misuse of web agents. SafeArena comprises 250 safe and 250 harmful tasks across four websites. We classify the harmful tasks into five harm categories -- misinformation, illegal activity, harassment, cybercrime, and social bias, designed to assess realistic misuses of web agents. We evaluate leading LLM-based web agents, including GPT-4o, Claude-3.5 Sonnet, Qwen-2-VL 72B, and Llama-3.2 90B, on our benchmark. To systematically assess their susceptibility to harmful tasks, we introduce the Agent Risk Assessment framework that categorizes agent behavior across four risk levels. We find agents are surprisingly compliant with malicious requests, with GPT-4o and Qwen-2 completing 34.7% and 27.3% of harmful requests, respectively. Our findings highlight the urgent need for safety alignment procedures for web agents. Our benchmark is available here: https://safearena.github.io

Benchmarking Vision Language Models for Cultural Understanding

Foundation models and vision-language pre-training have notably advanced Vision Language Models (VLMs), enabling multimodal processing of visual and linguistic data. However, their performance has been typically assessed on general scene understanding - recognizing objects, attributes, and actions - rather than cultural comprehension. This study introduces CulturalVQA, a visual question-answering benchmark aimed at assessing VLM's geo-diverse cultural understanding. We curate a collection of 2,378 image-question pairs with 1-5 answers per question representing cultures from 11 countries across 5 continents. The questions probe understanding of various facets of culture such as clothing, food, drinks, rituals, and traditions. Benchmarking VLMs on CulturalVQA, including GPT-4V and Gemini, reveals disparity in their level of cultural understanding across regions, with strong cultural understanding capabilities for North America while significantly lower performance for Africa. We observe disparity in their performance across cultural facets too, with clothing, rituals, and traditions seeing higher performances than food and drink. These disparities help us identify areas where VLMs lack cultural understanding and demonstrate the potential of CulturalVQA as a comprehensive evaluation set for gauging VLM progress in understanding diverse cultures.

A Multilingual Perspective on Probing Gender Bias

Gender bias represents a form of systematic negative treatment that targets individuals based on their gender. This discrimination can range from subtle sexist remarks and gendered stereotypes to outright hate speech. Prior research has revealed that ignoring online abuse not only affects the individuals targeted but also has broader societal implications. These consequences extend to the discouragement of women's engagement and visibility within public spheres, thereby reinforcing gender inequality. This thesis investigates the nuances of how gender bias is expressed through language and within language technologies. Significantly, this thesis expands research on gender bias to multilingual contexts, emphasising the importance of a multilingual and multicultural perspective in understanding societal biases. In this thesis, I adopt an interdisciplinary approach, bridging natural language processing with other disciplines such as political science and history, to probe gender bias in natural language and language models.

Grammatical Gender's Influence on Distributional Semantics: A Causal Perspective

How much meaning influences gender assignment across languages is an active area of research in modern linguistics and cognitive science. We can view current approaches as aiming to determine where gender assignment falls on a spectrum, from being fully arbitrarily determined to being largely semantically determined. For the latter case, there is a formulation of the neo-Whorfian hypothesis, which claims that even inanimate noun gender influences how people conceive of and talk about objects (using the choice of adjective used to modify inanimate nouns as a proxy for meaning). We offer a novel, causal graphical model that jointly represents the interactions between a noun's grammatical gender, its meaning, and adjective choice. In accordance with past results, we find a relationship between the gender of nouns and the adjectives which modify them. However, when we control for the meaning of the noun, we find that grammatical gender has a near-zero effect on adjective choice, thereby calling the neo-Whorfian hypothesis into question.

Social Bias Probing: Fairness Benchmarking for Language Models

Large language models have been shown to encode a variety of social biases, which carries the risk of downstream harms. While the impact of these biases has been recognized, prior methods for bias evaluation have been limited to binary association tests on small datasets, offering a constrained view of the nature of societal biases within language models. In this paper, we propose an original framework for probing language models for societal biases. We collect a probing dataset to analyze language models' general associations, as well as along the axes of societal categories, identities, and stereotypes. To this end, we leverage a novel perplexity-based fairness score. We curate a large-scale benchmarking dataset addressing drawbacks and limitations of existing fairness collections, expanding to a variety of different identities and stereotypes. When comparing our methodology with prior work, we demonstrate that biases within language models are more nuanced than previously acknowledged. In agreement with recent findings, we find that larger model variants exhibit a higher degree of bias. Moreover, we expose how identities expressing different religions lead to the most pronounced disparate treatments across all models.

Measuring Intersectional Biases in Historical Documents

Data-driven analyses of biases in historical texts can help illuminate the origin and development of biases prevailing in modern society. However, digitised historical documents pose a challenge for NLP practitioners as these corpora suffer from errors introduced by optical character recognition (OCR) and are written in an archaic language. In this paper, we investigate the continuities and transformations of bias in historical newspapers published in the Caribbean during the colonial era (18th to 19th centuries). Our analyses are performed along the axes of gender, race, and their intersection. We examine these biases by conducting a temporal study in which we measure the development of lexical associations using distributional semantics models and word embeddings. Further, we evaluate the effectiveness of techniques designed to process OCR-generated data and assess their stability when trained on and applied to the noisy historical newspapers. We find that there is a trade-off between the stability of the word embeddings and their compatibility with the historical dataset. We provide evidence that gender and racial biases are interdependent, and their intersection triggers distinct effects. These findings align with the theory of intersectionality, which stresses that biases affecting people with multiple marginalised identities compound to more than the sum of their constituents.

Measuring Gender Bias in West Slavic Language Models

Pre-trained language models have been known to perpetuate biases from the underlying datasets to downstream tasks. However, these findings are predominantly based on monolingual language models for English, whereas there are few investigative studies of biases encoded in language models for languages beyond English. In this paper, we fill this gap by analysing gender bias in West Slavic language models. We introduce the first template-based dataset in Czech, Polish, and Slovak for measuring gender bias towards male, female and non-binary subjects. We complete the sentences using both mono- and multilingual language models and assess their suitability for the masked language modelling objective. Next, we measure gender bias encoded in West Slavic language models by quantifying the toxicity and genderness of the generated words. We find that these language models produce hurtful completions that depend on the subject's gender. Perhaps surprisingly, Czech, Slovak, and Polish language models produce more hurtful completions with men as subjects, which, upon inspection, we find is due to completions being related to violence, death, and sickness.