Federico Bianchi

Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri et al. · stanford

Despite the success of large vision and language models (VLMs) in many downstream applications, it is unclear how well they encode compositional information. Here, we create the Attribution, Relation, and Order (ARO) benchmark to systematically evaluate the ability of VLMs to understand different types of relationships, attributes, and order. ARO consists of Visual Genome Attribution, to test the understanding of objects' properties; Visual Genome Relation, to test for relational understanding; and COCO & Flickr30k-Order, to test for order sensitivity. ARO is orders of magnitude larger than previous benchmarks of compositionality, with more than 50,000 test cases. We show where state-of-the-art VLMs have poor relational understanding, can blunder when linking objects to their attributes, and demonstrate a severe lack of order sensitivity. VLMs are predominantly trained and evaluated on large datasets with rich compositional structure in the images and captions. Yet, training on these datasets has not been enough to address the lack of compositional understanding, and evaluating on these datasets has failed to surface this deficiency. To understand why these limitations emerge and are not represented in the standard tests, we zoom into the evaluation and training procedures. We demonstrate that it is possible to perform well on retrieval over existing datasets without using the composition and order information. Given that contrastive pretraining optimizes for retrieval on datasets with similar shortcuts, we hypothesize that this can explain why the models do not need to learn to represent compositional information. This finding suggests a natural solution: composition-aware hard negative mining. We show that a simple-to-implement modification of contrastive learning significantly improves the performance on tasks requiring understanding of order and compositionality.

Federico Bianchi, Pratyusha Kalluri, Esin Durmus et al. · stanford

Machine learning models that convert user-written text descriptions into images are now widely available online and used by millions of users to generate millions of images a day. We investigate the potential for these models to amplify dangerous and complex stereotypes. We find a broad range of ordinary prompts produce stereotypes, including prompts simply mentioning traits, descriptors, occupations, or objects. For example, we find cases of prompting for basic traits or social roles resulting in images reinforcing whiteness as ideal, prompting for occupations resulting in amplification of racial and gender disparities, and prompting for objects resulting in reification of American norms. Stereotypes are present regardless of whether prompts explicitly mention identity and demographic language or avoid such language. Moreover, stereotypes persist despite mitigation strategies; neither user attempts to counter stereotypes by requesting images with specific counter-stereotypes nor institutional attempts to add system ``guardrails'' have prevented the perpetuation of stereotypes. Our analysis justifies concerns regarding the impacts of today's models, presenting striking exemplars, and connecting these findings with deep insights into harms drawn from social scientific and humanist disciplines. This work contributes to the effort to shed light on the uniquely complex biases in language-vision models and demonstrates the ways that the mass deployment of text-to-image generation models results in mass dissemination of stereotypes and resulting harms.

Federico Bianchi, Mirac Suzgun, Giuseppe Attanasio et al. · stanford

Training large language models to follow instructions makes them perform better on a wide range of tasks and generally become more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not harmlessness, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3% safety examples (a few hundred demonstrations) when fine-tuning a model like LLaMA can substantially improve its safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find exaggerated safety behaviours, where too much safety-tuning makes models refuse perfectly safe prompts if they superficially resemble unsafe ones. As a whole, our results illustrate trade-offs in training LLMs to be helpful and training them to be safe.

Paul Röttger, Hannah Rose Kirk, Bertie Vidgen et al. · oxford, stanford

Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This risk motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse to comply with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a systematic way. XSTest comprises 250 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with, and 200 unsafe prompts as contrasts that models, for most applications, should refuse. We describe XSTest's creation and composition, and then use the test suite to highlight systematic failure modes in state-of-the-art language models as well as more general challenges in building safer language models.

Patrick John Chia, Giuseppe Attanasio, Federico Bianchi et al. · stanford

The steady rise of online shopping goes hand in hand with the development of increasingly complex ML and NLP models. While most use cases are cast as specialized supervised learning problems, we argue that practitioners would greatly benefit from more transferable representations of products. In this work, we build on recent developments in contrastive learning to train FashionCLIP, a CLIP-like model for the fashion industry. We showcase its capabilities for retrieval, classification and grounding, and release our model and code to the community.

Federico Bianchi, Stefanie Anja Hills, Patricia Rossini et al. · stanford

Well-annotated data is a prerequisite for good Natural Language Processing models. Too often, though, annotation decisions are governed by optimizing time or annotator agreement. We make a case for nuanced efforts in an interdisciplinary setting for annotating offensive online speech. Detecting offensive content is rapidly becoming one of the most important real-world NLP tasks. However, most datasets use a single binary label, e.g., for hate or incivility, even though each concept is multi-faceted. This modeling choice severely limits nuanced insights, but also performance. We show that a more fine-grained multi-label approach to predicting incivility and hateful or intolerant content addresses both conceptual and performance issues. We release a novel dataset of over 40,000 tweets about immigration from the US and UK, annotated with six labels for different aspects of incivility and intolerance. Our dataset not only allows for a more nuanced understanding of harmful speech online, models trained on it also outperform or match performance on benchmark datasets.

Paul Röttger, Debora Nozza, Federico Bianchi et al. · stanford

Hate speech is a global phenomenon, but most hate speech datasets so far focus on English-language content. This hinders the development of more effective hate speech detection models in hundreds of languages spoken by billions across the world. More data is needed, but annotating hateful content is expensive, time-consuming and potentially harmful to annotators. To mitigate these issues, we explore data-efficient strategies for expanding hate speech detection into under-resourced languages. In a series of experiments with mono- and multilingual models across five non-English languages, we find that 1) a small amount of target-language fine-tuning data is needed to achieve strong performance, 2) the benefits of using more such data decrease exponentially, and 3) initial fine-tuning on readily-available English data can partially substitute target-language data and improve model generalisability. Based on these findings, we formulate actionable recommendations for hate speech detection in low-resource language settings.

Anne Lauscher, Federico Bianchi, Samuel Bowman et al. · stanford

Pre-trained language models (PLMs) have outperformed other NLP models on a wide range of tasks. Opting for a more thorough understanding of their capabilities and inner workings, researchers have established the extend to which they capture lower-level knowledge like grammaticality, and mid-level semantic knowledge like factual understanding. However, there is still little understanding of their knowledge of higher-level aspects of language. In particular, despite the importance of sociodemographic aspects in shaping our language, the questions of whether, where, and how PLMs encode these aspects, e.g., gender or age, is still unexplored. We address this research gap by probing the sociodemographic knowledge of different single-GPU PLMs on multiple English data sets via traditional classifier probing and information-theoretic minimum description length probing. Our results show that PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs. We further conduct a multilingual analysis and investigate the effect of supplementary training to further explore to what extent, where, and with what amount of pre-training data the knowledge is encoded. Our overall results indicate that sociodemographic knowledge is still a major challenge for NLP. PLMs require large amounts of pre-training data to acquire the knowledge and models that excel in general language understanding do not seem to own more knowledge about these aspects.

Mert Yuksekgonul, Daniel Koceja, Xinhao Li et al. · stanford

How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erdős' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to $2\times$ faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.

Giuseppe Attanasio, Debora Nozza, Federico Bianchi et al. · stanford

Language is constantly changing and evolving, leaving language models to become quickly outdated. Consequently, we should continuously update our models with new data to expose them to new events and facts. However, that requires additional computing, which means new carbon emissions. Do any measurable benefits justify this cost? This paper looks for empirical evidence to support continuous training. We reproduce existing benchmarks and extend them to include additional time periods, models, and tasks. Our results show that the downstream task performance of temporally adapted English models for social media data do not improve over time. Pretrained models without temporal adaptation are actually significantly more effective and efficient. However, we also note a lack of suitable temporal benchmarks. Our findings invite a critical reflection on when and how to temporally adapt language models, accounting for sustainability.

Patrick John Chia, Jacopo Tagliabue, Federico Bianchi et al. · stanford

Product discovery is a crucial component for online shopping. However, item-to-item recommendations today do not allow users to explore changes along selected dimensions: given a query item, can a model suggest something similar but in a different color? We consider item recommendations of the comparative nature (e.g. "something darker") and show how CLIP-based models can support this use case in a zero-shot manner. Leveraging a large model built for fashion, we introduce GradREC and its industry potential, and offer a first rounded assessment of its strength and weaknesses.

Tommaso Fornaciari, Dirk Hovy, Federico Bianchi · stanford

The most common ways to explore latent document dimensions are topic models and clustering methods. However, topic models have several drawbacks: e.g., they require us to choose the number of latent dimensions a priori, and the results are stochastic. Most clustering methods have the same issues and lack flexibility in various ways, such as not accounting for the influence of different topics on single documents, forcing word-descriptors to belong to a single topic (hard-clustering) or necessarily relying on word representations. We propose PROgressive SImilarity Thresholds - ProSiT, a deterministic and interpretable method, agnostic to the input format, that finds the optimal number of latent dimensions and only has two hyper-parameters, which can be set efficiently via grid search. We compare this method with a wide range of topic models and clustering methods on four benchmark data sets. In most setting, ProSiT matches or outperforms the other methods in terms six metrics of topic coherence and distinctiveness, producing replicable, deterministic results.

Federico Bianchi, Stefano Speziali, Andrea Marini et al. · stanford

In this work, we describe in detail how Deep Learning and Computer Vision can help to detect fault events of the AirTender system, an aftermarket motorcycle damping system component. One of the most effective ways to monitor the AirTender functioning is to look for oil stains on its surface. Starting from real-time images, AirTender is first detected in the motorbike suspension system, simulated indoor, and then, a binary classifier determines whether AirTender is spilling oil or not. The detection is made with the help of the Yolo5 architecture, whereas the classification is carried out with the help of a suitably designed Convolutional Neural Network, OilNet40. In order to detect oil leaks more clearly, we dilute the oil in AirTender with a fluorescent dye with an excitation wavelength peak of approximately 390 nm. AirTender is then illuminated with suitable UV LEDs. The whole system is an attempt to design a low-cost detection setup. An on-board device, such as a mini-computer, is placed near the suspension system and connected to a full hd camera framing AirTender. The on-board device, through our Neural Network algorithm, is then able to localize and classify AirTender as normally functioning (non-leak image) or anomaly (leak image).

Moussa Koulako Bala Doumbouya, Ananjan Nandi, Gabriel Poesia et al.

Despite their demonstrated valuable capabilities, state-of-the-art (SOTA) widely deployed large language models (LLMs) still have the potential to cause harm to society due to the ineffectiveness of their safety filters, which can be bypassed by prompt transformations called jailbreak attacks. Current approaches to LLM safety assessment, which employ datasets of templated prompts and benchmarking pipelines, fail to cover sufficiently large and diverse sets of jailbreak attacks, leading to the widespread deployment of unsafe LLMs. Recent research showed that novel jailbreak attacks could be derived by composition; however, a formal composable representation for jailbreak attacks, which, among other benefits, could enable the exploration of a large compositional space of jailbreak attacks through program synthesis methods, has not been previously proposed. We introduce h4rm3l, a novel approach that addresses this gap with a human-readable domain-specific language (DSL). Our framework comprises: (1) The h4rm3l DSL, which formally expresses jailbreak attacks as compositions of parameterized string transformation primitives. (2) A synthesizer with bandit algorithms that efficiently generates jailbreak attacks optimized for a target black box LLM. (3) The h4rm3l red-teaming software toolkit that employs the previous two components and an automated harmful LLM behavior classifier that is strongly aligned with human judgment. We demonstrate h4rm3l's efficacy by synthesizing a dataset of 2656 successful novel jailbreak attacks targeting 6 SOTA open-source and proprietary LLMs, and by benchmarking those models against a subset of these synthesized attacks. Our results show that h4rm3l's synthesized attacks are diverse and more successful than existing jailbreak attacks in literature, with success rates exceeding 90% on SOTA LLMs.

Rishav Hada, Amir Ebrahimi Fard, Sarah Shugars et al.

Increasingly taking place in online spaces, modern political conversations are typically perceived to be unproductively affirming -- siloed in so called ``echo chambers'' of exclusively like-minded discussants. Yet, to date we lack sufficient means to measure viewpoint diversity in conversations. To this end, in this paper, we operationalize two viewpoint metrics proposed for recommender systems and adapt them to the context of social media conversations. This is the first study to apply these two metrics (Representation and Fragmentation) to real world data and to consider the implications for online conversations specifically. We apply these measures to two topics -- daylight savings time (DST), which serves as a control, and the more politically polarized topic of immigration. We find that the diversity scores for both Fragmentation and Representation are lower for immigration than for DST. Further, we find that while pro-immigrant views receive consistent pushback on the platform, anti-immigrant views largely operate within echo chambers. We observe less severe yet similar patterns for DST. Taken together, Representation and Fragmentation paint a meaningful and important new picture of viewpoint diversity.

Kaitlyn Zhou, Martijn Bartelds, Federico Bianchi et al.

Despite speech recognition systems achieving low word error rates on standard benchmarks, they often fail on short, high-stakes utterances in real-world deployments. Here, we study this failure mode in a high-stakes task: the transcription of U.S. street names as spoken by U.S. participants. We evaluate 15 models from OpenAI, Deepgram, Google, and Microsoft on recordings from linguistically diverse U.S. speakers and find an average transcription error rate of 44%. We quantify the downstream impact of failed transcriptions by geographic locations and show that mis-transcriptions systematically cause errors for all speakers, but that routing distance errors are twice as large for non-English primary speakers compared to English primary speakers. To mitigate this harm, we introduce a synthetic data generation approach that produces diverse pronunciations of named entities using open-source text-to-speech models. Fine-tuning with less than 1,000 synthetic samples improves street name transcription accuracy by nearly 60% (relative to base models) for non-English primary speakers. Our results highlight a critical gap between benchmark performance and real-world reliability in speech systems and demonstrate a simple, scalable path to reducing high-stakes transcription errors.

Junlin Wang, Federico Bianchi, Shang Zhu et al.

Modern AI benchmarks operate at a complexity that outpaces traditional verification methods. Tasks authored by domain experts often contain implicit assumptions, incomplete environment specifications, and brittle evaluation logic that human annotation cannot reliably catch. We introduce Auto Benchmark Audit (ABA), an agentic framework that systematically audits individual benchmark tasks, uncovering issues such as hidden environment dependencies, specification gaps, and limited grading logic. We run ABA on a collection of frontier LLM benchmarks and previous NeurIPS publications, totaling 168 benchmarks across nine domains. Across this corpus, ABA identifies critical issues including ambiguous task design, execution environment conflicts, and incorrect ground truths in over 25.7% of the evaluated tasks. The precision of these automated audits is validated by expert review and independent third-party reports such as upstream PRs. Crucially, we demonstrate that these problematic tasks severely distorts capability assessments for agents and LLMs: filtering out these tasks with issues shifts model rankings and increases average performance on SWE-bench Verified and Terminal-Bench 2 by 9.9% and 9.6%, respectively. We release the agentic tool and all task annotations to support the future development of frontier benchmarks.

Mirac Suzgun, Emily Shen, Federico Bianchi et al.

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

Mehmet Hamza Erol, Xiangpeng Hao, Federico Bianchi et al.

Traditional query optimization relies on cost-based optimizers that estimate execution cost (e.g., runtime, memory, and I/O) using predefined heuristics and statistical models. Improving these heuristics requires substantial engineering effort, and even when implemented, these heuristics often cannot take into account semantic correlations in queries and schemas that could enable better physical plans. Using our DBPlanBench harness for the DataFusion engine, we expose the physical plan through a compact serialized representation and let the LLM propose localized edits that can be applied and executed. We then apply an evolutionary search over these edits to refine candidates across iterations. Our key insight is that LLMs can leverage semantic knowledge to identify and apply non-obvious optimizations, such as join orderings that minimize intermediate cardinalities. We obtain up to 4.78$\times$ speedups on some queries and we demonstrate a small-to-large workflow in which optimizations found on small databases transfer effectively to larger databases.

Fan Nie, Junlin Wang, Harper Hua et al.

Data science agents promise to accelerate discovery and insight-generation by turning data into executable analyses and findings. Yet existing data science benchmarks fall short due to fragmented evaluation interfaces that make cross-benchmark comparison difficult, narrow task coverage and a lack of rigorous data grounding. In particular, we show that a substantial portion of tasks in current benchmarks can be solved without using the actual data. To address these limitations, we introduce DSGym, a standardized framework for evaluating and training data science agents in self-contained execution environments. Unlike static benchmarks, DSGym provides a modular architecture that makes it easy to add tasks, agent scaffolds, and tools, positioning it as a live, extensible testbed. We curate DSGym-Tasks, a holistic task suite that standardizes and refines existing benchmarks via quality and shortcut solvability filtering. We further expand coverage with (1) DSBio: expert-derived bioinformatics tasks grounded in literature and (2) DSPredict: challenging prediction tasks spanning domains such as computer vision, molecular prediction, and single-cell perturbation. Beyond evaluation, DSGym enables agent training via execution-verified data synthesis pipeline. As a case study, we build a 2,000-example training set and trained a 4B model in DSGym that outperforms GPT-4o on standardized analysis benchmarks. Overall, DSGym enables rigorous end-to-end measurement of whether agents can plan, implement, and validate data analyses in realistic scientific context.

Kaitlyn Zhou, Federico Bianchi, Martijn Bartelds et al.

Artificially generated speech is increasingly embedded in everyday life. Voice cloning in particular enables applications where identity preservation is important, such as completing a recording, dubbing in a new language, or preserving the voices of individuals with speech loss. However, in our work, we find that despite the term, voice cloning does not faithfully ''clone'' an individual's voice. Instead, we find that widely-used voice cloning models systematically apply style transfer to source voices. As rated by human annotators, cloned voices are perceived as more authoritative, warm, customer-service-like, and human-like compared to their sources. Human annotators also report greater trust in cloned voices than source voices, and a greater willingness to disclose sensitive personal information to them. Our work furthermore shows that voice cloning leads to homogenization of speaker characteristics, as measured by reduced variance in accent, speaking rate, and the audio embedding space. Together, our results highlight a new set of limitations and risks of voice cloning technology and their potential impact on human behavior.

Yongchan Kwon, Shang Zhu, Federico Bianchi et al.

The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model's main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, formatting and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than $25\%$ of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity. (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of $GPT-OSS-20B$ from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement.

Patrick John Chia, Jacopo Tagliabue, Federico Bianchi et al.

As with most Machine Learning systems, recommender systems are typically evaluated through performance metrics computed over held-out data points. However, real-world behavior is undoubtedly nuanced: ad hoc error analysis and deployment-specific tests must be employed to ensure the desired quality in actual deployments. In this paper, we propose RecList, a behavioral-based testing methodology. RecList organizes recommender systems by use case and introduces a general plug-and-play procedure to scale up behavioral testing. We demonstrate its capabilities by analyzing known algorithms and black-box commercial systems, and we release RecList as an open source, extensible package for the community.

Federico Bianchi, Patrick John Chia, Mert Yuksekgonul et al.

Negotiation is the basis of social interactions; humans negotiate everything from the price of cars to how to share common resources. With rapidly growing interest in using large language models (LLMs) to act as agents on behalf of human users, such LLM agents would also need to be able to negotiate. In this paper, we study how well LLMs can negotiate with each other. We develop NegotiationArena: a flexible framework for evaluating and probing the negotiation abilities of LLM agents. We implemented three types of scenarios in NegotiationArena to assess LLM's behaviors in allocating shared resources (ultimatum games), aggregate resources (trading games) and buy/sell goods (price negotiations). Each scenario allows for multiple turns of flexible dialogues between LLM agents to allow for more complex negotiations. Interestingly, LLM agents can significantly boost their negotiation outcomes by employing certain behavioral tactics. For example, by pretending to be desolate and desperate, LLMs can improve their payoffs by 20\% when negotiating against the standard GPT-4. We also quantify irrational negotiation behaviors exhibited by the LLM agents, many of which also appear in humans. Together, \NegotiationArena offers a new environment to investigate LLM interactions, enabling new insights into LLM's theory of mind, irrationality, and reasoning abilities.

Federico Bianchi, James Zou

The risks derived from large language models (LLMs) generating deceptive and damaging content have been the subject of considerable research, but even safe generations can lead to problematic downstream impacts. In our study, we shift the focus to how even safe text coming from LLMs can be easily turned into potentially dangerous content through Bait-and-Switch attacks. In such attacks, the user first prompts LLMs with safe questions and then employs a simple find-and-replace post-hoc technique to manipulate the outputs into harmful narratives. The alarming efficacy of this approach in generating toxic content highlights a significant challenge in developing reliable safety guardrails for LLMs. In particular, we stress that focusing on the safety of the verbatim LLM outputs is insufficient and that we also need to consider post-hoc transformations.

Mirac Suzgun, Mert Yuksekgonul, Federico Bianchi et al.

Despite their impressive performance on complex tasks, current language models (LMs) typically operate in a vacuum: Each input query is processed separately, without retaining insights from previous attempts. Here, we present Dynamic Cheatsheet (DC), a lightweight framework that endows a black-box LM with a persistent, evolving memory. Rather than repeatedly re-discovering or re-committing the same solutions and mistakes, DC enables models to store and reuse accumulated strategies, code snippets, and general problem-solving insights at inference time. This test-time learning enhances performance substantially across a range of tasks without needing explicit ground-truth labels or human feedback. Leveraging DC, Claude 3.5 Sonnet's accuracy more than doubled on AIME math exams once it began retaining algebraic insights across questions. Similarly, GPT-4o's success rate on Game of 24 increased from 10% to 99% after the model discovered and reused a Python-based solution. In tasks prone to arithmetic mistakes, such as balancing equations, DC enabled GPT-4o and Claude to reach near-perfect accuracy by recalling previously validated code, whereas their baselines stagnated around 50%. Beyond arithmetic challenges, DC yields notable accuracy gains on knowledge-demanding tasks. Claude achieved a 9% improvement in GPQA-Diamond and an 8% boost on MMLU-Pro problems. Crucially, DC's memory is self-curated, focusing on concise, transferable snippets rather than entire transcript. Unlike finetuning or static retrieval methods, DC adapts LMs' problem-solving skills on the fly, without modifying their underlying parameters. Overall, our findings present DC as a promising approach for augmenting LMs with persistent memory, bridging the divide between isolated inference events and the cumulative, experience-driven learning characteristic of human cognition.

Isabel O. Gallegos, Chen Shani, Weiyan Shi et al.

As generative artificial intelligence (AI) enables the creation and dissemination of information at massive scale and speed, it is increasingly important to understand how people perceive AI-generated content. One prominent policy proposal requires explicitly labeling AI-generated content to increase transparency and encourage critical thinking about the information, but prior research has not yet tested the effects of such labels. To address this gap, we conducted a survey experiment (N=1601) on a diverse sample of Americans, presenting participants with an AI-generated message about several public policies (e.g., allowing colleges to pay student-athletes), randomly assigning whether participants were told the message was generated by (a) an expert AI model, (b) a human policy expert, or (c) no label. We found that messages were generally persuasive, influencing participants' views of the policies by 9.74 percentage points on average. However, while 94.6% of participants assigned to the AI and human label conditions believed the authorship labels, labels had no significant effects on participants' attitude change toward the policies, judgments of message accuracy, nor intentions to share the message with others. These patterns were robust across a variety of participant characteristics, including prior knowledge of the policy, prior experience with AI, political party, education level, or age. Taken together, these results imply that, while authorship labels would likely enhance transparency, they are unlikely to substantially affect the persuasiveness of the labeled content, highlighting the need for alternative strategies to address challenges posed by AI-generated information.

Mirac Suzgun, Tayfun Gur, Federico Bianchi et al.

As language models (LMs) become integral to fields like healthcare, law, and journalism, their ability to differentiate between fact, belief, and knowledge is essential for reliable decision-making. Failure to grasp these distinctions can lead to significant consequences in areas such as medical diagnosis, legal judgments, and dissemination of fake news. Despite this, current literature has largely focused on more complex issues such as theory of mind, overlooking more fundamental epistemic challenges. This study systematically evaluates the epistemic reasoning capabilities of modern LMs, including GPT-4, Claude-3, and Llama-3, using a new dataset, KaBLE, consisting of 13,000 questions across 13 tasks. Our results reveal key limitations. First, while LMs achieve 86% accuracy on factual scenarios, their performance drops significantly with false scenarios, particularly in belief-related tasks. Second, LMs struggle with recognizing and affirming personal beliefs, especially when those beliefs contradict factual data, which raises concerns for applications in healthcare and counseling, where engaging with a person's beliefs is critical. Third, we identify a salient bias in how LMs process first-person versus third-person beliefs, performing better on third-person tasks (80.7%) compared to first-person tasks (54.4%). Fourth, LMs lack a robust understanding of the factive nature of knowledge, namely, that knowledge inherently requires truth. Fifth, LMs rely on linguistic cues for fact-checking and sometimes bypass the deeper reasoning. These findings highlight significant concerns about current LMs' ability to reason about truth, belief, and knowledge while emphasizing the need for advancements in these areas before broad deployment in critical sectors.

Federico Bianchi, Yongchan Kwon, Zachary Izzo et al.

How many mistakes do published AI papers contain? Peer-reviewed publications form the foundation upon which new research and knowledge are built. Errors that persist in the literature can propagate unnoticed, creating confusion in follow-up studies and complicating reproducibility. The accelerating pace of research and the increasing demands on the peer-review system make such mistakes harder to detect and avoid. To address this, we developed a Paper Correctness Checker based on GPT-5 to systematically identify mistakes in papers previously published at top AI conferences and journals. Our analysis focuses on objective mistakes-e.g., errors in formulas, derivations, calculations, figures, and tables-that have a clearly verifiable ground truth. We intentionally exclude subjective considerations such as novelty, importance, or writing quality. We find that published papers contain a non-negligible number of objective mistakes and that the average number of mistakes per paper has increased over time-from 3.8 in NeurIPS 2021 to 5.9 in NeurIPS 2025 (55.3% increase); from 4.1 in ICLR 2018 to 5.2 in ICLR 2025; and from 5.0 in TMLR 2022/23 to 5.5 in TMLR 2025. Human experts reviewed 316 potential mistakes identified by the AI Checker and confirmed that 263 were actual mistakes, corresponding to a precision of 83.2%. While most identified issues are relatively minor, correcting them would reduce confusion in the literature and strengthen reproducibility. The AI Checker also surfaced potentially more substantive mistakes that could affect the interpretation of results. Moreover, we show that the AI Checker can propose correct fixes for 75.8% of the identified mistakes. Overall, this study highlights the potential of frontier LLMs to detect and correct objective mistakes in published papers, helping to establish a firmer foundation of knowledge.

Jacopo Tagliabue, Federico Bianchi, Ciro Greco

Even as AI capabilities improve, most enterprises do not consider agents trustworthy enough to work on production data. In this paper, we argue that the path to trustworthy agentic workflows begins with solving the infrastructure problem first: traditional lakehouses are not suited for agent access patterns, but if we design one around transactions, governance follows. In particular, we draw an operational analogy to MVCC in databases and show why a direct transplant fails in a decoupled, multi-language setting. We then propose an agent-first design, Bauplan, that reimplements data and compute isolation in the lakehouse. We conclude by sharing a reference implementation of a self-healing pipeline in Bauplan, which seamlessly couples agent reasoning with all the desired guarantees for correctness and trust.

Federico Bianchi, Owen Queen, Nitya Thakkar et al.

There is growing interest in using AI agents for scientific research, yet fundamental questions remain about their capabilities as scientists and reviewers. To explore these questions, we organized Agents4Science, the first conference in which AI agents serve as both primary authors and reviewers, with humans as co-authors and co-reviewers. Here, we discuss the key learnings from the conference and their implications for human-AI collaboration in science.

Davide Salaorni, Federico Bianchi, Francesco Trovò et al.

The increasing integration of renewable energy sources (RESs) is transforming traditional power grid networks, which require new approaches for managing decentralized energy production and consumption. Microgrids (MGs) provide a promising solution by enabling localized control over energy generation, storage, and distribution. This paper presents a novel reinforcement learning (RL)-based methodology for optimizing microgrid energy management. Specifically, we propose an RL agent that learns optimal energy trading and storage policies by leveraging historical data on energy production, consumption, and market prices. A digital twin (DT) is used to simulate the energy storage system dynamics, incorporating degradation factors to ensure a realistic emulation of the analysed setting. Our approach is validated through an experimental campaign using real-world data from a power grid located in the Italian territory. The results indicate that the proposed RL-based strategy outperforms rule-based methods and existing RL benchmarks, offering a robust solution for intelligent microgrid management.

Mert Yuksekgonul, Federico Bianchi, Joseph Boen et al.

AI is undergoing a paradigm shift, with breakthroughs achieved by systems orchestrating multiple large language models (LLMs) and other complex components. As a result, developing principled and automated optimization methods for compound AI systems is one of the most important new challenges. Neural networks faced a similar challenge in its early days until backpropagation and automatic differentiation transformed the field by making optimization turn-key. Inspired by this, we introduce TextGrad, a powerful framework performing automatic ``differentiation'' via text. TextGrad backpropagates textual feedback provided by LLMs to improve individual components of a compound AI system. In our framework, LLMs provide rich, general, natural language suggestions to optimize variables in computation graphs, ranging from code snippets to molecular structures. TextGrad follows PyTorch's syntax and abstraction and is flexible and easy-to-use. It works out-of-the-box for a variety of tasks, where the users only provide the objective function without tuning components or prompts of the framework. We showcase TextGrad's effectiveness and generality across a diverse range of applications, from question answering and molecule optimization to radiotherapy treatment planning. Without modifying the framework, TextGrad improves the zero-shot accuracy of GPT-4o in Google-Proof Question Answering from $51\%$ to $55\%$, yields $20\%$ relative performance gain in optimizing LeetCode-Hard coding problem solutions, improves prompts for reasoning, designs new druglike small molecules with desirable in silico binding, and designs radiation oncology treatment plans with high specificity. TextGrad lays a foundation to accelerate the development of the next-generation of AI systems.

Federico Bianchi, Vincenzo Cutrona, Dirk Hovy

Twitter data have become essential to Natural Language Processing (NLP) and social science research, driving various scientific discoveries in recent years. However, the textual data alone are often not enough to conduct studies: especially social scientists need more variables to perform their analysis and control for various factors. How we augment this information, such as users' location, age, or tweet sentiment, has ramifications for anonymity and reproducibility, and requires dedicated effort. This paper describes Twitter-Demographer, a simple, flow-based tool to enrich Twitter data with additional information about tweets and users. Twitter-Demographer is aimed at NLP practitioners and (computational) social scientists who want to enrich their datasets with aggregated information, facilitating reproducibility, and providing algorithmic privacy-by-design measures for pseudo-anonymity. We discuss our design choices, inspired by the flow-based programming paradigm, to use black-box components that can easily be chained together and extended. We also analyze the ethical issues related to the use of this tool, and the built-in measures to facilitate pseudo-anonymity.

Federico Bianchi, Debora Nozza, Dirk Hovy

Meaning is context-dependent, but many properties of language (should) remain the same even if we transform the context. For example, sentiment, entailment, or speaker properties should be the same in a translation and original of a text. We introduce language invariant properties: i.e., properties that should not change when we transform text, and how they can be used to quantitatively evaluate the robustness of transformation algorithms. We use translation and paraphrasing as transformation examples, but our findings apply more broadly to any transformation. Our results indicate that many NLP transformations change properties like author characteristics, i.e., make them sound more male. We believe that studying these properties will allow NLP to address both social factors and pragmatic aspects of language. We also release an application suite that can be used to evaluate the invariance of transformation applications.

Federico Bianchi, Marco Marelli, Paolo Nicoli et al.

Understanding differences of viewpoints across corpora is a fundamental task for computational social sciences. In this paper, we propose the Sliced Word Embedding Association Test (SWEAT), a novel statistical measure to compute the relative polarization of a topical wordset across two distributional representations. To this end, SWEAT uses two additional wordsets, deemed to have opposite valence, to represent two different poles. We validate our approach and illustrate a case study to show the usefulness of the introduced measure.

Federico Bianchi, Giuseppe Attanasio, Raphael Pisoni et al.

CLIP (Contrastive Language-Image Pre-training) is a very recent multi-modal model that jointly learns representations of images and texts. The model is trained on a massive amount of English data and shows impressive performance on zero-shot classification tasks. Training the same model on a different language is not trivial, since data in other languages might be not enough and the model needs high-quality translations of the texts to guarantee a good performance. In this paper, we present the first CLIP model for the Italian Language (CLIP-Italian), trained on more than 1.4 million image-text pairs. Results show that CLIP-Italian outperforms the multilingual CLIP model on the tasks of image retrieval and zero-shot classification.

Jacopo Tagliabue, Ciro Greco, Jean-Francis Roy et al.

The 2021 SIGIR workshop on eCommerce is hosting the Coveo Data Challenge for "In-session prediction for purchase intent and recommendations". The challenge addresses the growing need for reliable predictions within the boundaries of a shopping session, as customer intentions can be different depending on the occasion. The need for efficient procedures for personalization is even clearer if we consider the e-commerce landscape more broadly: outside of giant digital retailers, the constraints of the problem are stricter, due to smaller user bases and the realization that most users are not frequently returning customers. We release a new session-based dataset including more than 30M fine-grained browsing events (product detail, add, purchase), enriched by linguistic behavior (queries made by shoppers, with items clicked and items not clicked after the query) and catalog meta-data (images, text, pricing information). On this dataset, we ask participants to showcase innovative solutions for two open problems: a recommendation task (where a model is shown some events at the start of a session, and it is asked to predict future product interactions); an intent prediction task, where a model is shown a session containing an add-to-cart event, and it is asked to predict whether the item will be bought before the end of the session.

Federico Bianchi, Ciro Greco, Jacopo Tagliabue

We investigate grounded language learning through real-world data, by modelling a teacher-learner dynamics through the natural interactions occurring between users and search engines; in particular, we explore the emergence of semantic generalization from unsupervised dense representations outside of synthetic environments. A grounding domain, a denotation function and a composition function are learned from user data only. We show how the resulting semantics for noun phrases exhibits compositional properties while being fully learnable without any explicit labelling. We benchmark our grounded semantics on compositionality and zero-shot inference tasks, and we show that it provides better results and better generalizations than SOTA non-grounded models, such as word2vec and BERT.

Federico Bianchi, Jacopo Tagliabue, Bingqing Yu

We present Query2Prod2Vec, a model that grounds lexical representations for product search in product embeddings: in our model, meaning is a mapping between words and a latent space of products in a digital shop. We leverage shopping sessions to learn the underlying space and use merchandising annotations to build lexical analogies for evaluation: our experiments show that our model is more accurate than known techniques from the NLP and IR literature. Finally, we stress the importance of data efficiency for product search outside of retail giants, and highlight how Query2Prod2Vec fits with practical constraints faced by most practitioners.

Federico Bianchi, Bingqing Yu, Jacopo Tagliabue

Word embeddings (e.g., word2vec) have been applied successfully to eCommerce products through~\textit{prod2vec}. Inspired by the recent performance improvements on several NLP tasks brought by contextualized embeddings, we propose to transfer BERT-like architectures to eCommerce: our model -- ~\textit{Prod2BERT} -- is trained to generate representations of products through masked session modeling. Through extensive experiments over multiple shops, different tasks, and a range of design choices, we systematically compare the accuracy of~\textit{Prod2BERT} and~\textit{prod2vec} embeddings: while~\textit{Prod2BERT} is found to be superior in several scenarios, we highlight the importance of resources and hyperparameters in the best performing models. Finally, we provide guidelines to practitioners for training embeddings under a variety of computational and data constraints.

Federico Bianchi, Jacopo Tagliabue, Bingqing Yu et al.

This paper addresses the challenge of leveraging multiple embedding spaces for multi-shop personalization, proving that zero-shot inference is possible by transferring shopping intent from one website to another without manual intervention. We detail a machine learning pipeline to train and optimize embeddings within shops first, and support the quantitative findings with additional qualitative insights. We then turn to the harder task of using learned embeddings across shops: if products from different shops live in the same vector space, user intent - as represented by regions in this space - can then be transferred in a zero-shot fashion across websites. We propose and benchmark unsupervised and supervised methods to "travel" between embedding spaces, each with its own assumptions on data quantity and quality. We show that zero-shot personalization is indeed possible at scale by testing the shared embedding space with two downstream tasks, event prediction and type-ahead suggestions. Finally, we curate a cross-shop anonymized embeddings dataset to foster an inclusive discussion of this important business scenario.

Federico Bianchi, Gaetano Rossiello, Luca Costabello et al.

Knowledge graph embeddings are now a widely adopted approach to knowledge representation in which entities and relationships are embedded in vector spaces. In this chapter, we introduce the reader to the concept of knowledge graph embeddings by explaining what they are, how they can be generated and how they can be evaluated. We summarize the state-of-the-art in this field by describing the approaches that have been introduced to represent knowledge in the vector space. In relation to knowledge representation, we consider the problem of explainability, and discuss models and methods for explaining predictions obtained via knowledge graph embeddings.

Federico Bianchi, Silvia Terragni, Dirk Hovy et al.

Many data sets (e.g., reviews, forums, news, etc.) exist parallelly in multiple languages. They all cover the same content, but the linguistic differences make it impossible to use traditional, bag-of-word-based topic models. Models have to be either single-language or suffer from a huge, but extremely sparse vocabulary. Both issues can be addressed by transfer learning. In this paper, we introduce a zero-shot cross-lingual topic model. Our model learns topics on one language (here, English), and predicts them for unseen documents in different languages (here, Italian, French, German, and Portuguese). We evaluate the quality of the topic predictions for the same document in different languages. Our results show that the transferred topics are coherent and stable across languages, which suggests exciting future research directions.

Federico Bianchi, Valerio Di Carlo, Paolo Nicoli et al.

Word2vec is one of the most used algorithms to generate word embeddings because of a good mix of efficiency, quality of the generated representations and cognitive grounding. However, word meaning is not static and depends on the context in which words are used. Differences in word meaning that depends on time, location, topic, and other factors, can be studied by analyzing embeddings generated from different corpora in collections that are representative of these factors. For example, language evolution can be studied using a collection of news articles published in different time periods. In this paper, we present a general framework to support cross-corpora language studies with word embeddings, where embeddings generated from different corpora can be compared to find correspondences and differences in meaning across the corpora. CADE is the core component of our framework and solves the key problem of aligning the embeddings generated from different corpora. In particular, we focus on providing solid evidence about the effectiveness, generality, and robustness of CADE. To this end, we conduct quantitative and qualitative experiments in different domains, from temporal word embeddings to language localization and topical analysis. The results of our experiments suggest that CADE achieves state-of-the-art or superior performance on tasks where several competing approaches are available, yet providing a general method that can be used in a variety of domains. Finally, our experiments shed light on the conditions under which the alignment is reliable, which substantially depends on the degree of cross-corpora vocabulary overlap.

Federico Bianchi, Silvia Terragni, Dirk Hovy

Topic models extract groups of words from documents, whose interpretation as a topic hopefully allows for a better understanding of the data. However, the resulting word groups are often not coherent, making them harder to interpret. Recently, neural topic models have shown improvements in overall coherence. Concurrently, contextual embeddings have advanced the state of the art of neural models in general. In this paper, we combine contextualized representations with neural topic models. We find that our approach produces more meaningful and coherent topics than traditional bag-of-words topic models and recent neural models. Our results indicate that future improvements in language models will translate into better topic models.

Bingqing Yu, Jacopo Tagliabue, Ciro Greco et al.

We address the problem of personalizing query completion in a digital commerce setting, in which the bounce rate is typically high and recurring users are rare. We focus on in-session personalization and improve a standard noisy channel model by injecting dense vectors computed from product images at query time. We argue that image-based personalization displays several advantages over alternative proposals (from data availability to business scalability), and provide quantitative evidence and qualitative support on the effectiveness of the proposed methods. Finally, we show how a shared vector space between similar shops can be used to improve the experience of users browsing across sites, opening up the possibility of applying zero-shot unsupervised personalization to increase conversions. This will prove to be particularly relevant to retail groups that manage multiple brands and/or websites and to multi-tenant SaaS providers that serve multiple clients in the same space.

Debora Nozza, Federico Bianchi, Dirk Hovy

Recently, Natural Language Processing (NLP) has witnessed an impressive progress in many areas, due to the advent of novel, pretrained contextual representation models. In particular, Devlin et al. (2019) proposed a model, called BERT (Bidirectional Encoder Representations from Transformers), which enables researchers to obtain state-of-the art performance on numerous NLP tasks by fine-tuning the representations on their data set and task, without the need for developing and training highly-specific architectures. The authors also released multilingual BERT (mBERT), a model trained on a corpus of 104 languages, which can serve as a universal language model. This model obtained impressive results on a zero-shot cross-lingual natural inference task. Driven by the potential of BERT models, the NLP community has started to investigate and generate an abundant number of BERT models that are trained on a particular language, and tested on a specific data domain and task. This allows us to evaluate the true potential of mBERT as a universal language model, by comparing it to the performance of these more specific models. This paper presents the current state of the art in language-specific BERT models, providing an overall picture with respect to different dimensions (i.e. architectures, data domains, and tasks). Our aim is to provide an immediate and straightforward overview of the commonalities and differences between Language-Specific (language-specific) BERT models and mBERT. We also provide an interactive and constantly updated website that can be used to explore the information we have collected, at https://bertlang.unibocconi.it.

Valerio Di Carlo, Federico Bianchi, Matteo Palmonari

Temporal word embeddings have been proposed to support the analysis of word meaning shifts during time and to study the evolution of languages. Different approaches have been proposed to generate vector representations of words that embed their meaning during a specific time interval. However, the training process used in these approaches is complex, may be inefficient or it may require large text corpora. As a consequence, these approaches may be difficult to apply in resource-scarce domains or by scientists with limited in-depth knowledge of embedding models. In this paper, we propose a new heuristic to train temporal word embeddings based on the Word2vec model. The heuristic consists in using atemporal vectors as a reference, i.e., as a compass, when training the representations specific to a given time interval. The use of the compass simplifies the training process and makes it more efficient. Experiments conducted using state-of-the-art datasets and methodologies suggest that our approach outperforms or equals comparable approaches while being more robust in terms of the required corpus size.

Adriano Macarone Palmieri, Egor Kovlakov, Federico Bianchi et al.

Quantum tomography is currently ubiquitous for testing any implementation of a quantum information processing device. Various sophisticated procedures for state and process reconstruction from measured data are well developed and benefit from precise knowledge of the model describing state preparation and the measurement apparatus. However, physical models suffer from intrinsic limitations as actual measurement operators and trial states cannot be known precisely. This scenario inevitably leads to state-preparation-and-measurement (SPAM) errors degrading reconstruction performance. Here we develop and experimentally implement a machine learning based protocol reducing SPAM errors. We trained a supervised neural network to filter the experimental data and hence uncovered salient patterns that characterize the measurement probabilities for the original state and the ideal experimental apparatus free from SPAM errors. We compared the neural network state reconstruction protocol with a protocol treating SPAM errors by process tomography, as well as to a SPAM-agnostic protocol with idealized measurements. The average reconstruction fidelity is shown to be enhanced by 10\% and 27\%, respectively. The presented methods apply to the vast range of quantum experiments which rely on tomography.