Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.

AI agents hold the potential to revolutionize scientific productivity by automating literature reviews, replicating experiments, analyzing data, and even proposing new directions of inquiry; indeed, there are now many such agents, ranging from general-purpose "deep research" systems to specialized science-specific agents, such as AI Scientist and AIGS. Rigorous evaluation of these agents is critical for progress. Yet existing benchmarks fall short on several fronts: they (1) fail to provide holistic, product-informed measures of real-world use cases such as science research; (2) lack reproducible agent tools necessary for a controlled comparison of core agentic capabilities; (3) do not account for confounding variables such as model cost and tool access; (4) do not provide standardized interfaces for quick agent prototyping and evaluation; and (5) lack comprehensive baseline agents necessary to identify true advances. In response, we define principles and tooling for more rigorously benchmarking agents. Using these, we present AstaBench, a suite that provides the first holistic measure of agentic ability to perform scientific research, comprising 2400+ problems spanning the entire scientific discovery process and multiple scientific domains, and including many problems inspired by actual user requests to deployed Asta agents. Our suite comes with the first scientific research environment with production-grade search tools that enable controlled, reproducible evaluation, better accounting for confounders. Alongside, we provide a comprehensive suite of nine science-optimized classes of Asta agents and numerous baselines. Our extensive evaluation of 57 agents across 22 agent classes reveals several interesting findings, most importantly that despite meaningful progress on certain individual aspects, AI remains far from solving the challenge of science research assistance.

Recently, large language models (LLMs) have demonstrated outstanding reasoning capabilities on mathematical and coding tasks. However, their application to financial tasks-especially the most fundamental task of stock movement prediction-remains underexplored. We study a three-class classification problem (up, hold, down) and, by analyzing existing reasoning responses, observe that: (1) LLMs follow analysts' opinions rather than exhibit a systematic, independent analytical logic (CoTs). (2) LLMs list summaries from different sources without weighing adversarial evidence, yet such counterevidence is crucial for reliable prediction. It shows that the model does not make good use of its reasoning ability to complete the task. To address this, we propose Reflective Evidence Tuning (RETuning), a cold-start method prior to reinforcement learning, to enhance prediction ability. While generating CoT, RETuning encourages dynamically constructing an analytical framework from diverse information sources, organizing and scoring evidence for price up or down based on that framework-rather than on contextual viewpoints-and finally reflecting to derive the prediction. This approach maximally aligns the model with its learned analytical framework, ensuring independent logical reasoning and reducing undue influence from context. We also build a large-scale dataset spanning all of 2024 for 5,123 A-share stocks, with long contexts (32K tokens) and over 200K samples. In addition to price and news, it incorporates analysts' opinions, quantitative reports, fundamental data, macroeconomic indicators, and similar stocks. Experiments show that RETuning successfully unlocks the model's reasoning ability in the financial domain. Inference-time scaling still works even after 6 months or on out-of-distribution stocks, since the models gain valuable insights about stock movement prediction.

Large language models (LLMs) are increasingly deployed as agents in various contexts by providing tools at their disposal. However, LLM agents can exhibit unpredictable behaviors, including taking undesirable and/or unsafe actions. In order to measure the latent propensity of LLM agents for taking illegal actions under an EU legislative context, we introduce EU-Agent-Bench, a verifiable human-curated benchmark that evaluates an agent's alignment with EU legal norms in situations where benign user inputs could lead to unlawful actions. Our benchmark spans scenarios across several categories, including data protection, bias/discrimination, and scientific integrity, with each user request allowing for both compliant and non-compliant execution of the requested actions. Comparing the model's function calls against a rubric exhaustively supported by citations of the relevant legislature, we evaluate the legal compliance of frontier LLMs, and furthermore investigate the compliance effect of providing the relevant legislative excerpts in the agent's system prompt along with explicit instructions to comply. We release a public preview set for the research community, while holding out a private test set to prevent data contamination in evaluating upcoming models. We encourage future work extending agentic safety benchmarks to different legal jurisdictions and to multi-turn and multilingual interactions. We release our code on \href{https://github.com/ilijalichkovski/eu-agent-bench}{this URL}.

Abstract--The right to be forgotten is a fundamental principle of privacy-preserving regulations and extends to Machine Learning (ML) paradigms such as Federated Learning (FL). While FL enhances privacy by enabling collaborative model training without sharing private data, trained models still retain the influence of training data. Federated Unlearning (FU) methods recently proposed often rely on impractical assumptions for real-world FL deployments, such as storing client update histories or requiring access to a publicly available dataset. T o address these constraints, this paper introduces a novel method that leverages negated Pseudo-gradients Updates for Federated Unlearning (PUF). Our approach only uses standard client model updates, which are employed during regular FL rounds, and interprets them as pseudo-gradients. When a client needs to be forgotten, we apply the negation of their pseudo-gradients, appropriately scaled, to the global model. Unlike state-of-the-art mechanisms, PUF seamlessly integrates with FL workflows, incurs no additional computational and communication overhead beyond standard FL rounds, and supports concurrent unlearning requests. We extensively evaluated the proposed method on two well-known benchmark image classification datasets (CIF AR-10 and CIF AR-100) and a real-world medical imaging dataset for segmentation (ProstateMRI), using three different neural architectures: two residual networks and a vision transformer . The experimental results across various settings demonstrate that PUF achieves state-of-the-art forgetting effectiveness and recovery time, without relying on any additional assumptions. N today's digital landscape, privacy has become a major concern, as reflected by the emergence of robust regulatory frameworks worldwide [1]. The European Union (EU) has consistently emphasized the importance of protecting personal data, exemplified by the introduction of the General Data Protection Regulation (GDPR) in 2016 [2]. Most recently, in May 2024, the EU enacted Regulation 2024/1183 [3], establishing the European Digital Identity Framework that empowers individuals with fine-grained control over their information. One of the key rights of these regulations is the right to be forgotten, which allows individuals to request the deletion of their previously shared data. Similar rights are central to other major privacy laws worldwide, such as the California Consumer Privacy Act (CCP A) [4] where the right to delete grants California residents the on-demand removal of personal data held by businesses. Alessio Mora, Rebecca Montanari, and Paolo Bellavista are with the Department of Computer Science and Engineering, University of Bologna, Bologna, Italy (e-mail: {name.surname}@unibo.it).

Pre-trained large language models (LLMs) are becoming useful for various tasks. To improve their performance on certain tasks, it is necessary to fine-tune them on specific data corpora (e.g., medical reports, business data). These specialized data corpora may contain sensitive data (e.g., personal or confidential data) that will be memorized by the model and likely to be regurgitated during its subsequent use. This memorization of sensitive information by the model poses a significant privacy or confidentiality issue. To remove this memorization and sanitize the model without requiring costly additional fine-tuning on a secured data corpus, we propose SANI. SANI is an unlearning approach to sanitize language models. It relies on both an erasure and repair phases that 1) reset certain neurons in the last layers of the model to disrupt the memorization of fine-grained information, and then 2) fine-tune the model while avoiding memorizing sensitive information. We comprehensively evaluate SANI to sanitize both a model fine-tuned and specialized with medical data by removing directly and indirectly identifiers from the memorization of the model, and a standard pre-trained model by removing specific terms defined as confidential information from the model. Results show that with only few additional epochs of unlearning, the model is sanitized and the number of regurgitations is drastically reduced. This approach can be particularly useful for hospitals or other industries that have already spent significant resources training models on large datasets and wish to sanitize them before sharing.

The rise of right-wing populism in Europe has brought to the forefront the significance of analysing social media discourse to understand the dissemination of extremist ideologies and their impact on political outcomes. Twitter, as a platform for interaction and mobilisation, provides a unique window into the everyday communication of far-right supporters. In this paper, we propose a methodology that uses state-of-the-art natural language processing techniques with sociological insights to analyse the MIGR-TWIT corpus of far-right tweets in English and French. We aim to uncover patterns of discourse surrounding migration, hate speech, and persuasion techniques employed by right and far-right actors. By integrating linguistic, sociological, and computational approaches, we seek to offer cross-disciplinary insights into societal dynamics and contribute to a better understanding of contemporary challenges posed by right-wing extremism on social media platforms.

Reliable simulation of human behavior is essential for explaining, predicting, and intervening in our society. Recent advances in large language models (LLMs) have shown promise in emulating human behaviors, interactions, and decision-making, offering a powerful new lens for social science studies. However, the extent to which LLMs diverge from authentic human behavior in social contexts remains underexplored, posing risks of misinterpretation in scientific studies and unintended consequences in real-world applications. Here, we introduce a systematic framework for analyzing LLMs' behavior in social simulation. Our approach simulates multi-agent interactions through chatroom-style conversations and analyzes them across five linguistic dimensions, providing a simple yet effective method to examine emergent social cognitive biases. We conduct extensive experiments involving eight representative LLMs across three families. Our findings reveal that LLMs do not faithfully reproduce genuine human behavior but instead reflect overly idealized versions of it, shaped by the social desirability bias. In particular, LLMs show social role bias, primacy effect, and positivity bias, resulting in "Utopian" societies that lack the complexity and variability of real human interactions. These findings call for more socially grounded LLMs that capture the diversity of human social behavior.

Applying complex legal rules characterized by multiple, heterogeneously weighted criteria presents a fundamental challenge in judicial decision-making, often hindering the consistent realization of legislative intent. This challenge is particularly evident in the quantification of non-pecuniary damages in personal injury cases. This paper introduces Soppia, a structured prompting framework designed to assist legal professionals in navigating this complexity. By leveraging advanced AI, the system ensures a comprehensive and balanced analysis of all stipulated criteria, fulfilling the legislator's intent that compensation be determined through a holistic assessment of each case. Using the twelve criteria for non-pecuniary damages established in the Brazilian CLT (Art. 223-G) as a case study, we demonstrate how Soppia (System for Ordered Proportional and Pondered Intelligent Assessment) operationalizes nuanced legal commands into a practical, replicable, and transparent methodology. The framework enhances consistency and predictability while providing a versatile and explainable tool adaptable across multi-criteria legal contexts, bridging normative interpretation and computational reasoning toward auditable legal AI.

The emergence of foundation models (FMs) has enabled the development of highly capable and autonomous agents, unlocking new application opportunities across a wide range of domains. Evaluating the architecture of agents is particularly important as the architectural decisions significantly impact the quality attributes of agents given their unique characteristics, including compound architecture, autonomous and non-deterministic behaviour, and continuous evolution. However, these traditional methods fall short in addressing the evaluation needs of agent architecture due to the unique characteristics of these agents. Therefore, in this paper, we present AgentArcEval, a novel agent architecture evaluation method designed specially to address the complexities of FM-based agent architecture and its evaluation. Moreover, we present a catalogue of agent-specific general scenarios, which serves as a guide for generating concrete scenarios to design and evaluate the agent architecture. We demonstrate the usefulness of AgentArcEval and the catalogue through a case study on the architecture evaluation of a real-world tax copilot, named Luna.