Antoine Bosselut

Michihiro Yasunaga, Antoine Bosselut, Hongyu Ren et al. · stanford

Pretraining a language model (LM) on text has been shown to help various downstream NLP tasks. Recent works show that a knowledge graph (KG) can complement text data, offering structured background knowledge that provides a useful scaffold for reasoning. However, these works are not pretrained to learn a deep fusion of the two modalities at scale, limiting the potential to acquire fully joint representations of text and KG. Here we propose DRAGON (Deep Bidirectional Language-Knowledge Graph Pretraining), a self-supervised approach to pretraining a deeply joint language-knowledge foundation model from text and KG at scale. Specifically, our model takes pairs of text segments and relevant KG subgraphs as input and bidirectionally fuses information from both modalities. We pretrain this model by unifying two self-supervised reasoning tasks, masked language modeling and KG link prediction. DRAGON outperforms existing LM and LM+KG models on diverse downstream tasks including question answering across general and biomedical domains, with +5% absolute gain on average. In particular, DRAGON achieves notable performance on complex reasoning about language and knowledge (+10% on questions involving long contexts or multi-step reasoning) and low-resource QA (+8% on OBQA and RiddleSense), and new state-of-the-art results on various BioNLP tasks. Our code and trained models are available at https://github.com/michiyasunaga/dragon.

Zeming Chen, Alejandro Hernández Cano, Angelika Romanou et al. · allen-ai

Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2. We release our code for curating the medical pretraining corpus and the MEDITRON model weights to drive open-source development of more capable medical LLMs.

Eric Mitchell, Charles Lin, Antoine Bosselut et al. · stanford

Even the largest neural networks make errors, and once-correct predictions can become invalid as the world changes. Model editors make local updates to the behavior of base (pre-trained) models to inject updated knowledge or correct undesirable behaviors. Existing model editors have shown promise, but also suffer from insufficient expressiveness: they struggle to accurately model an edit's intended scope (examples affected by the edit), leading to inaccurate predictions for test inputs loosely related to the edit, and they often fail altogether after many edits. As a higher-capacity alternative, we propose Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model (SERAC), which stores edits in an explicit memory and learns to reason over them to modulate the base model's predictions as needed. To enable more rigorous evaluation of model editors, we introduce three challenging language model editing problems based on question answering, fact-checking, and dialogue generation. We find that only SERAC achieves high performance on all three problems, consistently outperforming existing approaches to model editing by a significant margin. Code, data, and additional project information will be made available at https://sites.google.com/view/serac-editing.

Zeming Chen, Qiyue Gao, Antoine Bosselut et al. · allen-ai

Models trained with counterfactually augmented data learn representations of the causal structure of tasks, enabling robust generalization. However, high-quality counterfactual data is scarce for most tasks and not easily generated at scale. When crowdsourced, such data is typically limited in scale and diversity; when generated using supervised methods, it is computationally expensive to extend to new counterfactual dimensions. In this work, we introduce DISCO (DIStilled COunterfactual Data), a new method for automatically generating high quality counterfactual data at scale. DISCO engineers prompts to generate phrasal perturbations with a large general language model. Then, a task-specific teacher model filters these generations to distill high-quality counterfactual data. While task-agnostic, we apply our pipeline to the task of natural language inference (NLI) and find that on challenging evaluations such as the NLI stress test, comparatively smaller student models trained with DISCO generated counterfactuals are more robust (6% absolute) and generalize better across distributions (2%) compared to models trained without data augmentation. Furthermore, DISCO augmented models are 10% more consistent between counterfactual pairs on three evaluation sets, demonstrating that DISCO augmentation enables models to more reliably learn causal representations. Our repository is available at: https://github.com/eric11eca/disco

Mete Ismayilzada, Debjit Paul, Syrielle Montariol et al. · deepmind

Recent efforts in natural language processing (NLP) commonsense reasoning research have yielded a considerable number of new datasets and benchmarks. However, most of these datasets formulate commonsense reasoning challenges in artificial scenarios that are not reflective of the tasks which real-world NLP systems are designed to solve. In this work, we present CRoW, a manually-curated, multi-task benchmark that evaluates the ability of models to apply commonsense reasoning in the context of six real-world NLP tasks. CRoW is constructed using a multi-stage data collection pipeline that rewrites examples from existing datasets using commonsense-violating perturbations. We use CRoW to study how NLP systems perform across different dimensions of commonsense knowledge, such as physical, temporal, and social reasoning. We find a significant performance gap when NLP systems are evaluated on CRoW compared to humans, showcasing that commonsense reasoning is far from being solved in real-world task settings. We make our dataset and leaderboard available to the research community at https://github.com/mismayil/crow.

Angelika Romanou, Syrielle Montariol, Debjit Paul et al.

Understanding narratives requires reasoning about the cause-and-effect relationships between events mentioned in the text. While existing foundation models yield impressive results in many NLP tasks requiring reasoning, it is unclear whether they understand the complexity of the underlying network of causal relationships of events in narratives. In this work, we present CRAB, a new Causal Reasoning Assessment Benchmark designed to evaluate causal understanding of events in real-world narratives. CRAB contains fine-grained, contextual causality annotations for ~2.7K pairs of real-world events that describe various newsworthy event timelines (e.g., the acquisition of Twitter by Elon Musk). Using CRAB, we measure the performance of several large language models, demonstrating that most systems achieve poor performance on the task. Motivated by classical causal principles, we also analyze the causal structures of groups of events in CRAB, and find that models perform worse on causal reasoning when events are derived from complex causal structures compared to simple linear causal chains. We make our dataset and code available to the research community.

Yifan Hou, Jiaoda Li, Yu Fei et al. · eth-zurich

Recent work has shown that language models (LMs) have strong multi-step (i.e., procedural) reasoning capabilities. However, it is unclear whether LMs perform these tasks by cheating with answers memorized from pretraining corpus, or, via a multi-step reasoning mechanism. In this paper, we try to answer this question by exploring a mechanistic interpretation of LMs for multi-step reasoning tasks. Concretely, we hypothesize that the LM implicitly embeds a reasoning tree resembling the correct reasoning process within it. We test this hypothesis by introducing a new probing approach (called MechanisticProbe) that recovers the reasoning tree from the model's attention patterns. We use our probe to analyze two LMs: GPT-2 on a synthetic task (k-th smallest element), and LLaMA on two simple language-based reasoning tasks (ProofWriter & AI2 Reasoning Challenge). We show that MechanisticProbe is able to detect the information of the reasoning tree from the model's attentions for most examples, suggesting that the LM indeed is going through a process of multi-step reasoning within its architecture in many cases.

Aman Madaan, Dheeraj Rajagopal, Niket Tandon et al. · cmu

Conditional set generation learns a mapping from an input sequence of tokens to a set. Several NLP tasks, such as entity typing and dialogue emotion tagging, are instances of set generation. Seq2Seq models, a popular choice for set generation, treat a set as a sequence and do not fully leverage its key properties, namely order-invariance and cardinality. We propose a novel algorithm for effectively sampling informative orders over the combinatorial space of label orders. We jointly model the set cardinality and output by prepending the set size and taking advantage of the autoregressive factorization used by Seq2Seq models. Our method is a model-independent data augmentation approach that endows any Seq2Seq model with the signals of order-invariance and cardinality. Training a Seq2Seq model on this augmented data (without any additional annotations) gets an average relative improvement of 20% on four benchmark datasets across various models: BART, T5, and GPT-3. Code to use SETAUG available at: https://setgen.structgen.com.

Debjit Paul, Mete Ismayilzada, Maxime Peyrard et al.

Language models (LMs) have recently shown remarkable performance on reasoning tasks by explicitly generating intermediate inferences, e.g., chain-of-thought prompting. However, these intermediate inference steps may be inappropriate deductions from the initial context and lead to incorrect final predictions. Here we introduce REFINER, a framework for finetuning LMs to explicitly generate intermediate reasoning steps while interacting with a critic model that provides automated feedback on the reasoning. Specifically, the critic provides structured feedback that the reasoning LM uses to iteratively improve its intermediate arguments. Empirical evaluations of REFINER on three diverse reasoning tasks show significant improvements over baseline LMs of comparable scale. Furthermore, when using GPT-3.5 or ChatGPT as the reasoner, the trained critic significantly improves reasoning without finetuning the reasoner. Finally, our critic model is trained without expensive human-in-the-loop data but can be substituted with humans at inference time.

Mete Ismayilzada, Antoine Bosselut

In this paper, we present kogito, an open-source tool for generating commonsense inferences about situations described in text. kogito provides an intuitive and extensible interface to interact with natural language generation models that can be used for hypothesizing commonsense knowledge inference from a textual input. In particular, kogito offers several features for targeted, multi-granularity knowledge generation. These include a standardized API for training and evaluating knowledge models, and generating and filtering inferences from them. We also include helper functions for converting natural language texts into a format ingestible by knowledge models - intermediate pipeline stages such as knowledge head extraction from text, heuristic and model-based knowledge head-relation matching, and an ability to define and use custom knowledge relations. We make the code for kogito available at https://github.com/epfl-nlp/kogito along with thorough documentation at https://kogito.readthedocs.io.

Deniz Bayazit, Negar Foroutan, Zeming Chen et al. · allen-ai

Pretrained language models (LMs) encode implicit representations of knowledge in their parameters. However, localizing these representations and disentangling them from each other remains an open problem. In this work, we investigate whether pretrained language models contain various knowledge-critical subnetworks: particular sparse computational subgraphs that can, if removed, precisely suppress specific knowledge the model has memorized. We propose a multi-objective differentiable masking scheme that can be applied to both weights and neurons to discover such subnetworks and show that we can use them to precisely remove specific knowledge from models while minimizing adverse effects on the behavior of the original model. We demonstrate our method on multiple GPT2 variants, uncovering highly sparse subnetworks (98%+ sparsity) that are critical for expressing specific collections of relational knowledge. When these subnetworks are removed, the remaining network maintains most of its initial abilities but struggles to represent the suppressed knowledge.

Negar Foroutan, Mohammadreza Banaei, Remi Lebret et al.

Multilingual pre-trained language models transfer remarkably well on cross-lingual downstream tasks. However, the extent to which they learn language-neutral representations (i.e., shared representations that encode similar phenomena across languages), and the effect of such representations on cross-lingual transfer performance, remain open questions. In this work, we conceptualize language neutrality of multilingual models as a function of the overlap between language-encoding sub-networks of these models. We employ the lottery ticket hypothesis to discover sub-networks that are individually optimized for various languages and tasks. Our evaluation across three distinct tasks and eleven typologically-diverse languages demonstrates that sub-networks for different languages are topologically similar (i.e., language-neutral), making them effective initializations for cross-lingual transfer with limited performance degradation.

Beatriz Borges, Negar Foroutan, Deniz Bayazit et al.

AI assistants are being increasingly used by students enrolled in higher education institutions. While these tools provide opportunities for improved teaching and education, they also pose significant challenges for assessment and learning outcomes. We conceptualize these challenges through the lens of vulnerability, the potential for university assessments and learning outcomes to be impacted by student use of generative AI. We investigate the potential scale of this vulnerability by measuring the degree to which AI assistants can complete assessment questions in standard university-level STEM courses. Specifically, we compile a novel dataset of textual assessment questions from 50 courses at EPFL and evaluate whether two AI assistants, GPT-3.5 and GPT-4 can adequately answer these questions. We use eight prompting strategies to produce responses and find that GPT-4 answers an average of 65.8% of questions correctly, and can even produce the correct answer across at least one prompting strategy for 85.1% of questions. When grouping courses in our dataset by degree program, these systems already pass non-project assessments of large numbers of core courses in various degree programs, posing risks to higher education accreditation that will be amplified as these models improve. Our results call for revising program-level assessment design in higher education in light of advances in generative AI.

Yijie Tong, Yifan Hou, Shaobo Cui et al.

Test-time compute (TTC) strategies have emerged as a lightweight approach to boost reasoning in large language models (LLMs). However, their application and benefits for vision-language models (VLMs) remain underexplored. We present a systematic study of TTC across seven VLMs and six benchmarks, specifically analyzing feature-based scoring and majority voting methods. We find that feature heuristics fail and voting yields only modest gains in single-model settings. We theoretically show that this limitation stems from a lack of prediction diversity: when outputs are highly correlated, voting provides little benefit. In contrast, multi-model ensembles offer richer diversity, yet standard majority voting fails to account for varying model capabilities. To address this, we propose Entropy-based TTC (ETTC), which selects the most confident prediction based on predictive entropy. Our method reduces to majority voting in the single-model case, but in model ensembles, it leverages confidence disparities to prioritize stronger models. We prove that ETTC outperforms majority voting under mild assumptions and empirically demonstrate that it consistently surpasses both voting and the best individual model. Crucially, our results show that smaller models can synergistically enhance larger ones, unlocking ensembling gains not achievable with standard strategies.

Luca Mouchel, Debjit Paul, Shaobo Cui et al.

Despite the remarkable performance of Large Language Models (LLMs) in natural language processing tasks, they still struggle with generating logically sound arguments, resulting in potential risks such as spreading misinformation. To address this issue, we introduce FIPO, a fallacy-informed framework that leverages preference optimization methods to steer LLMs toward logically sound arguments. FIPO includes a classification loss, to capture the fine-grained information on fallacy types. Our results on argumentation datasets show that our method reduces the fallacy errors by up to 17.5%. Furthermore, our human evaluation results indicate that the quality of the generated arguments by our method significantly outperforms the fine-tuned baselines, as well as other preference optimization methods, such as DPO. These findings highlight the importance of ensuring models are aware of logical fallacies for effective argument generation. Our code is available at github.com/lucamouchel/Logical-Fallacies.

Silin Gao, Jena D. Hwang, Saya Kanno et al.

Understanding rich narratives, such as dialogues and stories, often requires natural language processing systems to access relevant knowledge from commonsense knowledge graphs. However, these systems typically retrieve facts from KGs using simple heuristics that disregard the complex challenges of identifying situationally-relevant commonsense knowledge (e.g., contextualization, implicitness, ambiguity). In this work, we propose the new task of commonsense fact linking, where models are given contexts and trained to identify situationally-relevant commonsense knowledge from KGs. Our novel benchmark, ComFact, contains ~293k in-context relevance annotations for commonsense triplets across four stylistically diverse dialogue and storytelling datasets. Experimental results confirm that heuristic fact linking approaches are imprecise knowledge extractors. Learned fact linking models demonstrate across-the-board performance improvements (~34.6% F1) over these heuristics. Furthermore, improved knowledge retrieval yielded average downstream improvements of 9.8% for a dialogue response generation task. However, fact linking models still significantly underperform humans, suggesting our benchmark is a promising testbed for research in commonsense augmentation of NLP systems.

Beatriz Borges, Niket Tandon, Tanja Käser et al.

Natural Language Feedback (NLF) is an increasingly popular mechanism for aligning Large Language Models (LLMs) to human preferences. Despite the diversity of the information it can convey, NLF methods are often hand-designed and arbitrary, with little systematic grounding. At the same time, research in learning sciences has long established several effective feedback models. In this opinion piece, we compile ideas from pedagogy to introduce FELT, a feedback framework for LLMs that outlines various characteristics of the feedback space, and a feedback content taxonomy based on these variables, providing a general mapping of the feedback space. In addition to streamlining NLF designs, FELT also brings out new, unexplored directions for research in NLF. We make our taxonomy available to the community, providing guides and examples for mapping our categorizations to future research.

Jiahao Lu, Ziwei Xu, William Tjhi et al.

Large Language Models (LLMs) represent a new frontier of digital infrastructure that can support a wide range of public-sector applications, from general purpose citizen services to specialized and sensitive state functions. When expanding AI access, governments face a set of strategic choices over whether to buy existing services, build domestic capabilities, or adopt hybrid approaches across different domains and use cases. These are critical decisions especially when leading model providers are often foreign corporations, and LLM outputs are increasingly treated as trusted inputs to public decision-making and public discourse. In practice, these decisions are not intended to mandate a single approach across all domains; instead, national AI strategies are typically pluralistic, with sovereign, commercial and open-source models coexisting to serve different purposes. Governments may rely on commercial models for non-sensitive or commodity tasks, while pursuing greater control for critical, high-risk or strategically important applications. This paper provides a strategic framework for making this decision by evaluating these options across dimensions including sovereignty, safety, cost, resource capability, cultural fit, and sustainability. Importantly, "building" does not imply that governments must act alone: domestic capabilities may be developed through public research institutions, universities, state-owned enterprises, joint ventures, or broader national ecosystems. By detailing the technical requirements and practical challenges of each pathway, this work aims to serve as a reference for policy-makers to determine whether a buy or build approach best aligns with their specific national needs and societal goals.

Negar Foroutan, Mohammadreza Banaei, Karl Aberer et al.

In this work, we study whether multilingual language models (MultiLMs) can transfer logical reasoning abilities to other languages when they are fine-tuned for reasoning in a different language. We evaluate the cross-lingual reasoning abilities of MultiLMs in two schemes: (1) where the language of the context and the question remain the same in the new languages that are tested (i.e., the reasoning is still monolingual, but the model must transfer the learned reasoning ability across languages), and (2) where the language of the context and the question is different (which we term code-switched reasoning). On two logical reasoning datasets, RuleTaker and LeapOfThought, we demonstrate that although MultiLMs can transfer reasoning ability across languages in a monolingual setting, they struggle to transfer reasoning abilities in a code-switched setting. Following this observation, we propose a novel attention mechanism that uses a dedicated set of parameters to encourage cross-lingual attention in code-switched sequences, which improves the reasoning performance by up to 14% and 4% on the RuleTaker and LeapOfThought datasets, respectively.

Mete Ismayilzada, Renqing Cuomao, Daniil Yurshevich et al.

Creative problem-solving requires combining multiple cognitive abilities, including logical reasoning, lateral thinking, analogy-making, and commonsense knowledge, to discover insights that connect seemingly unrelated pieces of information. However, most existing benchmarks for large language models (LLMs) evaluate only specific components of this process. Moreover, many creativity-oriented benchmarks rely on artificially constructed brainteasers or contrived scenarios that do not reflect how creative problem-solving occurs in real-world settings. To address this gap, we introduce CresOWLve, a benchmark for evaluating creative problem-solving using puzzles grounded in real-world knowledge. Problems in CresOWLve require employing multiple creative thinking strategies, retrieving facts from diverse domains, and creatively combining them to arrive at a solution. Evaluating several frontier non-thinking and thinking LLMs, we show that CresOWLve remains highly challenging. Our analysis reveals a consistent performance gap: models perform substantially better on factual questions than on creative ones (up to a -17% drop). While models can often retrieve the relevant knowledge, they struggle to form the non-obvious creative connections required to integrate this information and arrive at the correct answer.

Mete Ismayilzada, Simone A. Luchini, Abdulkadir Gokce et al.

Creative thinking is a fundamental aspect of human cognition, and divergent thinking-the capacity to generate novel and varied ideas-is widely regarded as its core generative engine. Large language models (LLMs) have recently demonstrated impressive performance on divergent thinking tests and prior work has shown that models with higher task performance tend to be more aligned to human brain activity. However, existing brain-LLM alignment studies have focused on passive, non-creative tasks. Here, we explore brain alignment during creative thinking using fMRI data from 170 participants performing the Alternate Uses Task (AUT). We extract representations from LLMs varying in size (270M-72B) and measure alignment to brain responses via Representational Similarity Analysis (RSA), targeting the creativity-related default mode and frontoparietal networks. We find that brain-LLM alignment scales with model size (default mode network only) and idea originality (both networks), with effects strongest early in the creative process. We further show that post-training objectives shape alignment in functionally selective ways: a creativity-optimized \texttt{Llama-3.1-8B-Instruct} preserves alignment with high-creativity neural responses while reducing alignment with low-creativity ones; a human behavior fine-tuned model elevates alignment with both; and a reasoning-trained variant shows the opposite pattern, suggesting chain-of-thought training steers representations away from creative neural geometry toward analytical processing. These results demonstrate that post-training objectives selectively reshape LLM representations relative to the neural geometry of human creative thought.

Jérémy Barghorn, Anna Sotnikova, Sacha Friedli et al.

Large-enrollment university courses face persistent challenges in providing timely and scalable instructional support. While generative AI holds promise, its effective use depends on reliability and pedagogical alignment. We present a human-centered case study of AI-assisted support in a Calculus I course, implemented in close collaboration with the course instructor. We developed a system to answer students' questions on a discussion forum, fine-tuning a lightweight language model on 2,588 historical student-instructor interactions. The model achieved 75.3% accuracy on a benchmark of 150 representative questions annotated by five instructors, and in 36% of cases, its responses were rated equal to or better than instructor answers. Post-deployment student survey (N = 105) indicated that students valued the alignment of the responses with the course materials and their immediate availability, while still relying on the instructor verification for trust. We highlight the importance of hybrid human-AI workflows for safe and effective course support.

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Jonathan Coles, Stefano Schuppli, Lukas Drescher et al.

Large Language Models (LLMs) have surged as a transformative technology for science and society, prompting governments worldwide to pursue sovereign AI capabilities that ensure data compliance and cultural representation. However, the associated capital costs and engineering complexity required to train these models have largely restricted such capabilities to the private sector, leaving a significant gap for public institutions. This paper details the engineering journey behind training Apertus, a fully open multilingual foundation model, on the Alps supercomputer. Representing a first-of-its-kind achievement for academia at the 70B parameter scale, we successfully deployed a massive pre-training campaign on one of Europe's largest systems for open science, powered by NVIDIA GH200 Grace Hopper Superchips. We detail the challenges encountered in readying HPC infrastructure for training AI models, from overcoming storage bottlenecks to stabilizing large-scale interconnects, and the lessons learned in transforming a supercomputer into a resilient software-defined Machine Learning Platform. Finally, we discuss the post-training requirements and evolution of our Machine Learning platform, outlining how this initial release lays the groundwork for a sustained, iterative operational capability, in particular for fine tuning foundation models, that extends well beyond a single model training run.

Angelika Romanou, Mark Ibrahim, Candace Ross et al.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

Tianqing Fang, Zeming Chen, Yangqiu Song et al. · allen-ai, tencent-ai

Event commonsense reasoning requires the ability to reason about the relationship between events, as well as infer implicit context underlying that relationship. However, data scarcity makes it challenging for language models to learn to generate commonsense inferences for contexts and questions involving interactions between complex events. To address this demand, we present COM2 (COMplex COMmonsense), a new dataset created by sampling multi-hop logical queries (e.g., the joint effect or cause of both event A and B, or the effect of the effect of event C) from an existing commonsense knowledge graph (CSKG), and verbalizing them using handcrafted rules and large language models into multiple-choice and text generation questions. Our experiments show that language models trained on COM2 exhibit significant improvements in complex reasoning ability, resulting in enhanced zero-shot performance in both in-domain and out-of-domain tasks for question answering and generative commonsense reasoning, without expensive human annotations. Code and data are available at https://github.com/tqfang/complex-commonsense-reasoning.

Antoine Magron, Anna Dai, Mike Zhang et al.

Recent approaches in skill matching, employing synthetic training data for classification or similarity model training, have shown promising results, reducing the need for time-consuming and expensive annotations. However, previous synthetic datasets have limitations, such as featuring only one skill per sentence and generally comprising short sentences. In this paper, we introduce JobSkape, a framework to generate synthetic data that tackles these limitations, specifically designed to enhance skill-to-taxonomy matching. Within this framework, we create SkillSkape, a comprehensive open-source synthetic dataset of job postings tailored for skill-matching tasks. We introduce several offline metrics that show that our dataset resembles real-world data. Additionally, we present a multi-step pipeline for skill extraction and matching tasks using large language models (LLMs), benchmarking against known supervised methodologies. We outline that the downstream evaluation results on real-world data can beat baselines, underscoring its efficacy and adaptability.

Nivya Talokar, Ayush K Tarun, Murari Mandal et al.

LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.

Jibril Frej, Anna Dai, Syrielle Montariol et al.

Current course recommender systems primarily leverage learner-course interactions, course content, learner preferences, and supplementary course details like instructor, institution, ratings, and reviews, to make their recommendation. However, these systems often overlook a critical aspect: the evolving skill demand of the job market. This paper focuses on the perspective of academic researchers, working in collaboration with the industry, aiming to develop a course recommender system that incorporates job market skill demands. In light of the job market's rapid changes and the current state of research in course recommender systems, we outline essential properties for course recommender systems to address these demands effectively, including explainable, sequential, unsupervised, and aligned with the job market and user's goals. Our discussion extends to the challenges and research questions this objective entails, including unsupervised skill extraction from job listings, course descriptions, and resumes, as well as predicting recommendations that align with learner objectives and the job market and designing metrics to evaluate this alignment. Furthermore, we introduce an initial system that addresses some existing limitations of course recommender systems using large Language Models (LLMs) for skill extraction and Reinforcement Learning (RL) for alignment with the job market. We provide empirical results using open-source data to demonstrate its effectiveness.

Kyuhee Kim, Auguste Poiroux, Antoine Bosselut

Formal verification guarantees proof validity but not formalization faithfulness. For natural-language logical reasoning, where models construct axiom systems from scratch without library constraints, this gap between valid proofs and faithful translations is especially acute. We investigate whether frontier models exploit this gap when generating Lean 4 proofs, a behavior we term formalization gaming. We evaluate GPT-5 and DeepSeek-R1 on 303 first-order logic problems (203 from FOLIO, 100 from Multi-LogiEval), comparing unified generation against a two-stage pipeline that separates formalization from proving. Despite compilation rates of 87-99%, we find no evidence of systematic gaming in unified generation: models prefer reporting failure over forcing proofs, even under prompting designed to encourage it. However, unfaithfulness that evades our detection signals may still occur. The two-stage pipeline reveals two distinct modes of unfaithfulness: GPT-5 fabricates axioms during proof generation, a reactive fallback detectable via cross-stage comparison, while DeepSeek-R1 mistranslates premises during formalization, producing internally consistent outputs that evade detection entirely. These findings show that high compilation rates or accuracies should not be equated with faithful reasoning. Code and data are available at https://github.com/koreankiwi99/formalization-gaming.

Negar Foroutan, Paul Teiletche, Ayush Kumar Tarun et al.

The impact of different multilingual data mixtures in pretraining large language models (LLMs) has been a topic of ongoing debate, often raising concerns about potential trade-offs between language coverage and model performance (i.e., the curse of multilinguality). In this work, we investigate these assumptions by training 1.1B and 3B parameter LLMs on diverse multilingual corpora, varying the number of languages from 25 to 400. Our study challenges common beliefs surrounding multilingual training. First, we find that combining English and multilingual data does not necessarily degrade the in-language performance of either group, provided that languages have a sufficient number of tokens included in the pretraining corpus. Second, we observe that using English as a pivot language (i.e., a high-resource language that serves as a catalyst for multilingual generalization) yields benefits across language families, and contrary to expectations, selecting a pivot language from within a specific family does not consistently improve performance for languages within that family. Lastly, we do not observe a significant "curse of multilinguality" as the number of training languages increases in models at this scale. Our findings suggest that multilingual data, when balanced appropriately, can enhance language model capabilities without compromising performance, even in low-resource settings

Rishika Bhagwatkar, Syrielle Montariol, Angelika Romanou et al.

Humans can naturally identify, reason about, and explain anomalies in their environment. In computer vision, this long-standing challenge remains limited to industrial defects or unrealistic, synthetically generated anomalies, failing to capture the richness and unpredictability of real-world anomalies. In this work, we introduce CAVE, the first benchmark of real-world visual anomalies. CAVE supports three open-ended tasks: anomaly description, explanation, and justification; with fine-grained annotations for visual grounding and categorizing anomalies based on their visual manifestations, their complexity, severity, and commonness. These annotations draw inspiration from cognitive science research on how humans identify and resolve anomalies, providing a comprehensive framework for evaluating Vision-Language Models (VLMs) in detecting and understanding anomalies. We show that state-of-the-art VLMs struggle with visual anomaly perception and commonsense reasoning, even with advanced prompting strategies. By offering a realistic and cognitively grounded benchmark, CAVE serves as a valuable resource for advancing research in anomaly detection and commonsense reasoning in VLMs.

Mina Lee, Katy Ilonka Gero, John Joon Young Chung et al. · allen-ai, deepmind

In our era of rapid technological advancement, the research landscape for writing assistants has become increasingly fragmented across various research communities. We seek to address this challenge by proposing a design space as a structured way to examine and explore the multidimensional space of intelligent and interactive writing assistants. Through a large community collaboration, we explore five aspects of writing assistants: task, user, technology, interaction, and ecosystem. Within each aspect, we define dimensions (i.e., fundamental components of an aspect) and codes (i.e., potential options for each dimension) by systematically reviewing 115 papers. Our design space aims to offer researchers and designers a practical tool to navigate, comprehend, and compare the various possibilities of writing assistants, and aid in the envisioning and design of new writing assistants.

Deniz Bayazit, Aaron Mueller, Antoine Bosselut

Large language models (LLMs) learn non-trivial abstractions during pretraining, like detecting irregular plural noun subjects. However, it is not well understood when and how specific linguistic abilities emerge as traditional evaluation methods such as benchmarking fail to reveal how models acquire concepts and capabilities. To bridge this gap and better understand model training at the concept level, we use sparse crosscoders to discover and align features across model checkpoints. Using this approach, we track the evolution of linguistic features during pretraining. We train crosscoders between open-sourced checkpoint triplets with significant performance and representation shifts, and introduce a novel metric, Relative Indirect Effects (RelIE), to trace training stages at which individual features become causally important for task performance. We show that crosscoders can detect feature emergence, maintenance, and discontinuation during pretraining. Our approach is architecture-agnostic and scalable, offering a promising path toward more interpretable and fine-grained analysis of representation learning throughout pretraining.

Weiqi Wang, Tianqing Fang, Wenxuan Ding et al.

The task of zero-shot commonsense question answering evaluates models on their capacity to reason about general scenarios beyond those presented in specific datasets. Existing approaches for tackling this task leverage external knowledge from CommonSense Knowledge Bases (CSKBs) by pretraining the model on synthetic QA pairs constructed from CSKBs. In these approaches, negative examples (distractors) are formulated by randomly sampling from CSKBs using fairly primitive keyword constraints. However, two bottlenecks limit these approaches: the inherent incompleteness of CSKBs limits the semantic coverage of synthetic QA pairs, and the lack of human annotations makes the sampled negative examples potentially uninformative and contradictory. To tackle these limitations above, we propose Conceptualization-Augmented Reasoner (CAR), a zero-shot commonsense question-answering framework that fully leverages the power of conceptualization. Specifically, CAR abstracts a commonsense knowledge triple to many higher-level instances, which increases the coverage of CSKB and expands the ground-truth answer space, reducing the likelihood of selecting false-negative distractors. Extensive experiments demonstrate that CAR more robustly generalizes to answering questions about zero-shot commonsense scenarios than existing methods, including large language models, such as GPT3.5 and ChatGPT. Our codes, data, and model checkpoints are available at https://github.com/HKUST-KnowComp/CAR.

Shivalika Singh, Angelika Romanou, Clémentine Fourrier et al.

Cultural biases in multilingual datasets pose significant challenges for their effectiveness as global benchmarks. These biases stem not only from differences in language but also from the cultural knowledge required to interpret questions, reducing the practical utility of translated datasets like MMLU. Furthermore, translation often introduces artefacts that can distort the meaning or clarity of questions in the target language. A common practice in multilingual evaluation is to rely on machine-translated evaluation sets, but simply translating a dataset is insufficient to address these challenges. In this work, we trace the impact of both of these issues on multilingual evaluations and ensuing model performances. Our large-scale evaluation of state-of-the-art open and proprietary models illustrates that progress on MMLU depends heavily on learning Western-centric concepts, with 28% of all questions requiring culturally sensitive knowledge. Moreover, for questions requiring geographic knowledge, an astounding 84.9% focus on either North American or European regions. Rankings of model evaluations change depending on whether they are evaluated on the full portion or the subset of questions annotated as culturally sensitive, showing the distortion to model rankings when blindly relying on translated MMLU. We release Global MMLU, an improved MMLU with evaluation coverage across 42 languages -- with improved overall quality by engaging with compensated professional and community annotators to verify translation quality while also rigorously evaluating cultural biases present in the original dataset. This comprehensive Global MMLU set also includes designated subsets labeled as culturally sensitive and culturally agnostic to allow for more holistic, complete evaluation.

Debjit Paul, Robert West, Antoine Bosselut et al.

Large language models (LLMs) have been shown to perform better when asked to reason step-by-step before answering a question. However, it is unclear to what degree the model's final answer is faithful to the stated reasoning steps. In this paper, we perform a causal mediation analysis on twelve LLMs to examine how intermediate reasoning steps generated by the LLM influence the final outcome and find that LLMs do not reliably use their intermediate reasoning steps when generating an answer. To address this issue, we introduce FRODO, a framework to tailor small-sized LMs to generate correct reasoning steps and robustly reason over these steps. FRODO consists of an inference module that learns to generate correct reasoning steps using an implicit causal reward function and a reasoning module that learns to faithfully reason over these intermediate inferences using a counterfactual and causal preference objective. Our experiments show that FRODO significantly outperforms four competitive baselines. Furthermore, FRODO improves the robustness and generalization ability of the reasoning LM, yielding higher performance on out-of-distribution test sets. Finally, we find that FRODO's rationales are more faithful to its final answer predictions than standard supervised fine-tuning.

Tim R. Davidson, Veniamin Veselovsky, Martin Josifoski et al. · stanford

We introduce an approach to evaluate language model (LM) agency using negotiation games. This approach better reflects real-world use cases and addresses some of the shortcomings of alternative LM benchmarks. Negotiation games enable us to study multi-turn, and cross-model interactions, modulate complexity, and side-step accidental evaluation data leakage. We use our approach to test six widely used and publicly accessible LMs, evaluating performance and alignment in both self-play and cross-play settings. Noteworthy findings include: (i) only closed-source models tested here were able to complete these tasks; (ii) cooperative bargaining games proved to be most challenging to the models; and (iii) even the most powerful models sometimes "lose" to weaker opponents

Silin Gao, Jane Dwivedi-Yu, Ping Yu et al. · berkeley, meta-ai

To achieve faithful reasoning that aligns with human expectations, large language models (LLMs) need to ground their reasoning to real-world knowledge (e.g., web facts, math and physical rules). Tools help LLMs access this external knowledge, but there remains challenges for fine-tuning LLM agents (e.g., Toolformer) to invoke tools in multi-step reasoning problems, where inter-connected tool calls require holistic and efficient tool usage planning. In this work, we propose a new method for LLMs to better leverage tools in multi-step reasoning. Our method, Chain-of-Abstraction (CoA), trains LLMs to first decode reasoning chains with abstract placeholders, and then call domain tools to reify each reasoning chain by filling in specific knowledge. This planning with abstract chains enables LLMs to learn more general reasoning strategies, which are robust to shifts of domain knowledge (e.g., math results) relevant to different reasoning questions. It also allows LLMs to perform decoding and calling of external tools in parallel, which avoids the inference delay caused by waiting for tool responses. In mathematical reasoning and Wiki QA domains, we show that our method consistently outperforms previous chain-of-thought and tool-augmented baselines on both in-distribution and out-of-distribution test sets, with an average ~6% absolute QA accuracy improvement. LLM agents trained with our method also show more efficient tool use, with inference speed being on average ~1.4x faster than baseline tool-augmented LLMs.

Angelika Romanou, Negar Foroutan, Anna Sotnikova et al.

The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (\ie, multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts. Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed.

Badr AlKhamissi, Greta Tuckute, Antoine Bosselut et al.

Large language models (LLMs) exhibit remarkable capabilities on not just language tasks, but also various tasks that are not linguistic in nature, such as logical reasoning and social inference. In the human brain, neuroscience has identified a core language system that selectively and causally supports language processing. We here ask whether similar specialization for language emerges in LLMs. We identify language-selective units within 18 popular LLMs, using the same localization approach that is used in neuroscience. We then establish the causal role of these units by demonstrating that ablating LLM language-selective units -- but not random units -- leads to drastic deficits in language tasks. Correspondingly, language-selective LLM units are more aligned to brain recordings from the human language system than random units. Finally, we investigate whether our localization method extends to other cognitive domains: while we find specialized networks in some LLMs for reasoning and social capabilities, there are substantial differences among models. These findings provide functional and causal evidence for specialization in large language models, and highlight parallels with the functional organization in the brain.

Khanh Cao Nguyen, Mike Zhang, Syrielle Montariol et al.

Skill Extraction involves identifying skills and qualifications mentioned in documents such as job postings and resumes. The task is commonly tackled by training supervised models using a sequence labeling approach with BIO tags. However, the reliance on manually annotated data limits the generalizability of such approaches. Moreover, the common BIO setting limits the ability of the models to capture complex skill patterns and handle ambiguous mentions. In this paper, we explore the use of in-context learning to overcome these challenges, on a benchmark of 6 uniformized skill extraction datasets. Our approach leverages the few-shot learning capabilities of large language models (LLMs) to identify and extract skills from sentences. We show that LLMs, despite not being on par with traditional supervised models in terms of performance, can better handle syntactically complex skill mentions in skill extraction tasks.

Li Mi, Chang Xu, Javiera Castillo-Navarro et al.

Cross-view geo-localization aims at localizing a ground-level query image by matching it to its corresponding geo-referenced aerial view. In real-world scenarios, the task requires accommodating diverse ground images captured by users with varying orientations and reduced field of views (FoVs). However, existing learning pipelines are orientation-specific or FoV-specific, demanding separate model training for different ground view variations. Such models heavily depend on the North-aligned spatial correspondence and predefined FoVs in the training data, compromising their robustness across different settings. To tackle this challenge, we propose ConGeo, a single- and cross-view Contrastive method for Geo-localization: it enhances robustness and consistency in feature representations to improve a model's invariance to orientation and its resilience to FoV variations, by enforcing proximity between ground view variations of the same location. As a generic learning objective for cross-view geo-localization, when integrated into state-of-the-art pipelines, ConGeo significantly boosts the performance of three base models on four geo-localization benchmarks for diverse ground view variations and outperforms competing methods that train separate models for each ground view variation.

Mete Ismayilzada, Defne Circi, Jonne Sälevä et al.

Large language models (LLMs) have demonstrated significant progress in various natural language generation and understanding tasks. However, their linguistic generalization capabilities remain questionable, raising doubts about whether these models learn language similarly to humans. While humans exhibit compositional generalization and linguistic creativity in language use, the extent to which LLMs replicate these abilities, particularly in morphology, is under-explored. In this work, we systematically investigate the morphological generalization abilities of LLMs through the lens of compositionality. We define morphemes as compositional primitives and design a novel suite of generative and discriminative tasks to assess morphological productivity and systematicity. Focusing on agglutinative languages such as Turkish and Finnish, we evaluate several state-of-the-art instruction-finetuned multilingual models, including GPT-4 and Gemini. Our analysis shows that LLMs struggle with morphological compositional generalization particularly when applied to novel word roots, with performance declining sharply as morphological complexity increases. While models can identify individual morphological combinations better than chance, their performance lacks systematicity, leading to significant accuracy gaps compared to humans.

Badr AlKhamissi, Greta Tuckute, Yingtian Tang et al.

Large language models (LLMs) exhibit remarkable similarity to neural activity in the human language network. However, the key properties of language shaping brain-like representations, and their evolution during training as a function of different tasks remain unclear. We here benchmark 34 training checkpoints spanning 300B tokens across 8 different model sizes to analyze how brain alignment relates to linguistic competence. Specifically, we find that brain alignment tracks the development of formal linguistic competence -- i.e., knowledge of linguistic rules -- more closely than functional linguistic competence. While functional competence, which involves world knowledge and reasoning, continues to develop throughout training, its relationship with brain alignment is weaker, suggesting that the human language network primarily encodes formal linguistic structure rather than broader cognitive functions. We further show that model size is not a reliable predictor of brain alignment when controlling for feature size and find that the correlation between next-word prediction, behavioral alignment and brain alignment fades once models surpass human language proficiency. Finally, using the largest set of rigorous neural language benchmarks to date, we show that language brain alignment benchmarks remain unsaturated, highlighting opportunities for improving future models. Taken together, our findings suggest that the human language network is best modeled by formal, rather than functional, aspects of language.

Mete Ismayilzada, Debjit Paul, Antoine Bosselut et al.

Creativity is the ability to produce novel, useful, and surprising ideas, and has been widely studied as a crucial aspect of human cognition. Machine creativity on the other hand has been a long-standing challenge. With the rise of advanced generative AI, there has been renewed interest and debate regarding AI's creative capabilities. Therefore, it is imperative to revisit the state of creativity in AI and identify key progresses and remaining challenges. In this work, we survey leading works studying the creative capabilities of AI systems, focusing on creative problem-solving, linguistic, artistic, and scientific creativity. Our review suggests that while the latest AI models are largely capable of producing linguistically and artistically creative outputs such as poems, images, and musical pieces, they struggle with tasks that require creative problem-solving, abstract thinking and compositionality and their generations suffer from a lack of diversity, originality, long-range incoherence and hallucinations. We also discuss key questions concerning copyright and authorship issues with generative models. Furthermore, we highlight the need for a comprehensive evaluation of creativity that is process-driven and considers several dimensions of creativity. Finally, we propose future research directions to improve the creativity of AI outputs, drawing inspiration from cognitive science and psychology.

Karina Halevy, Anna Sotnikova, Badr AlKhamissi et al. · cmu

Model editing has emerged as a cost-effective strategy to update knowledge stored in language models. However, model editing can have unintended consequences after edits are applied: information unrelated to the edits can also be changed, and other general behaviors of the model can be wrongly altered. In this work, we investigate how model editing methods unexpectedly amplify model biases post-edit. We introduce a novel benchmark dataset, Seesaw-CF, for measuring bias-related harms of model editing and conduct the first in-depth investigation of how different weight-editing methods impact model bias. Specifically, we focus on biases with respect to demographic attributes such as race, geographic origin, and gender, as well as qualitative flaws in long-form texts generated by edited language models. We find that edited models exhibit, to various degrees, more biased behavior as they become less confident in attributes for Asian, African, and South American subjects. Furthermore, edited models amplify sexism and xenophobia in text generations while remaining seemingly coherent and logical. Finally, editing facts about place of birth, country of citizenship, or gender have particularly negative effects on the model's knowledge about unrelated features like field of work.

Li Mi, Syrielle Montariol, Javiera Castillo-Navarro et al.

Asking questions about visual environments is a crucial way for intelligent agents to understand rich multi-faceted scenes, raising the importance of Visual Question Generation (VQG) systems. Apart from being grounded to the image, existing VQG systems can use textual constraints, such as expected answers or knowledge triplets, to generate focused questions. These constraints allow VQG systems to specify the question content or leverage external commonsense knowledge that can not be obtained from the image content only. However, generating focused questions using textual constraints while enforcing a high relevance to the image content remains a challenge, as VQG systems often ignore one or both forms of grounding. In this work, we propose Contrastive Visual Question Generation (ConVQG), a method using a dual contrastive objective to discriminate questions generated using both modalities from those based on a single one. Experiments on both knowledge-aware and standard VQG benchmarks demonstrate that ConVQG outperforms the state-of-the-art methods and generates image-grounded, text-guided, and knowledge-rich questions. Our human evaluation results also show preference for ConVQG questions compared to non-contrastive baselines.

Charles Corbière, Simon Roburin, Syrielle Montariol et al.

While chain-of-thought (CoT) prompting improves reasoning in large language models, its effectiveness in vision-language models (VLMs) remains limited due to over-reliance on textual cues and memorized knowledge. To investigate the visual reasoning capabilities of VLMs in complex real-world scenarios, we introduce DrivingVQA, a visual question answering dataset derived from driving theory exams, which contains 3,931 multiple-choice problems with expert-written explanations and grounded entities relevant to the reasoning process. Leveraging this dataset, we propose RIV-CoT, a Retrieval-Based Interleaved Visual Chain-of-Thought method that enables VLMs to reason using visual crops corresponding to these relevant entities. Our experiments demonstrate that RIV-CoT improves answer accuracy by 3.1% and reasoning accuracy by 4.6% over vanilla CoT prompting. Furthermore, we demonstrate that our method effectively scales to the larger A-OKVQA reasoning dataset by leveraging automatically generated pseudo-labels, outperforming CoT prompting.

Mete Ismayilzada, Antonio Laverghetta, Simone A. Luchini et al.

While Large Language Models (LLMs) have demonstrated impressive performance across natural language generation tasks, their ability to generate truly creative content-characterized by novelty, diversity, surprise, and quality-remains limited. Existing methods for enhancing LLM creativity often focus narrowly on diversity or specific tasks, failing to address creativity's multifaceted nature in a generalizable way. In this work, we propose Creative Preference Optimization (CrPO), a novel alignment method that injects signals from multiple creativity dimensions into the preference optimization objective in a modular fashion. We train and evaluate creativity-augmented versions of several models using CrPO and MuCE, a new large-scale human preference dataset spanning over 200,000 human-generated responses and ratings from more than 30 psychological creativity assessments. Our models outperform strong baselines, including GPT-4o, on both automated and human evaluations, producing more novel, diverse, and surprising generations while maintaining high output quality. Additional evaluations on NoveltyBench further confirm the generalizability of our approach. Together, our results demonstrate that directly optimizing for creativity within preference frameworks is a promising direction for advancing the creative capabilities of LLMs without compromising output quality.