Large language models (LLMs) are trained on extensive datasets that encapsulate substantial world knowledge. However, their outputs often include confidently stated inaccuracies. Earlier works suggest that LLMs encode truthfulness as a distinct linear feature, termed the "truth direction", which can classify truthfulness reliably. We address several open questions about the truth direction: (i) whether LLMs universally exhibit consistent truth directions; (ii) whether sophisticated probing techniques are necessary to identify truth directions; and (iii) how the truth direction generalizes across diverse contexts. Our findings reveal that not all LLMs exhibit consistent truth directions, with stronger representations observed in more capable models, particularly in the context of logical negation. Additionally, we demonstrate that truthfulness probes trained on declarative atomic statements can generalize effectively to logical transformations, question-answering tasks, in-context learning, and external knowledge sources. Finally, we explore the practical application of truthfulness probes in selective question-answering, illustrating their potential to improve user trust in LLM outputs. These results advance our understanding of truth directions and provide new insights into the internal representations of LLM beliefs. Our code is public at https://github.com/colored-dye/truthfulness_probe_generalization
The assessment of evaluation metrics (meta-evaluation) is crucial for determining the suitability of existing metrics in text-to-image (T2I) generation tasks. Human-based meta-evaluation is costly and time-intensive, and automated alternatives are scarce. We address this gap and propose CROC: a scalable framework for automated Contrastive Robustness Checks that systematically probes and quantifies metric robustness by synthesizing contrastive test cases across a comprehensive taxonomy of image properties. With CROC, we generate a pseudo-labeled dataset (CROC$^{syn}$) of over one million contrastive prompt-image pairs to enable a fine-grained comparison of evaluation metrics. We also use the dataset to train CROCScore, a new metric that achieves state-of-the-art performance among open-source methods, demonstrating an additional key application of our framework. To complement this dataset, we introduce a human-supervised benchmark (CROC$^{hum}$) targeting especially challenging categories. Our results highlight robustness issues in existing metrics: for example, many fail on prompts involving negation, and all tested open-source metrics fail on at least 25% of cases involving correct identification of body parts.
Formal mathematical reasoning remains a critical challenge for artificial intelligence, hindered by limitations of existing benchmarks in scope and scale. To address this, we present FormalMATH, a large-scale Lean4 benchmark comprising 5,560 formally verified problems spanning from high-school Olympiad challenges to undergraduate-level theorems across diverse domains (e.g., algebra, applied mathematics, calculus, number theory, and discrete mathematics). To mitigate the inefficiency of manual formalization, we introduce a novel human-in-the-loop autoformalization pipeline that integrates: (1) specialized large language models (LLMs) for statement autoformalization, (2) multi-LLM semantic verification, and (3) negation-based disproof filtering strategies using off-the-shelf LLM-based provers. This approach reduces expert annotation costs by retaining 72.09% of statements before manual verification while ensuring fidelity to the original natural-language problems. Our evaluation of state-of-the-art LLM-based theorem provers reveals significant limitations: even the strongest models achieve only 16.46% success rate under practical sampling budgets, exhibiting pronounced domain bias (e.g., excelling in algebra but failing in calculus) and over-reliance on simplified automation tactics. Notably, we identify a counterintuitive inverse relationship between natural-language solution guidance and proof success in chain-of-thought reasoning scenarios, suggesting that human-written informal reasoning introduces noise rather than clarity in the formal reasoning settings. We believe that FormalMATH provides a robust benchmark for benchmarking formal mathematical reasoning.
Negation is a universal feature of natural languages. In Brazilian Portuguese, the most commonly used negation particle is n\~ao, which can scope over nouns or verbs. When it scopes over a verb, n\~ao can occur in three positions: pre-verbal (NEG1), double negation (NEG2), or post-verbal (NEG3), e.g., n\~ao gosto, n\~ao gosto n\~ao, gosto n\~ao ("I do not like it"). From a variationist perspective, these structures are different forms of expressing negation. Pragmatically, they serve distinct communicative functions, such as politeness and modal evaluation. Despite their grammatical acceptability, these forms differ in frequency. NEG1 dominates across Brazilian regions, while NEG2 and NEG3 appear more rarely, suggesting its use is contextually restricted. This low-frequency challenges research, often resulting in subjective, non-generalizable interpretations of verbal negation with n\~ao. To address this, we developed negativas, a tool for automatically identifying NEG1, NEG2, and NEG3 in transcribed data. The tool's development involved four stages: i) analyzing a dataset of 22 interviews from the Falares Sergipanos database, annotated by three linguists, ii) creating a code using natural language processing (NLP) techniques, iii) running the tool, iv) evaluating accuracy. Inter-annotator consistency, measured using Fleiss' Kappa, was moderate (0.57). The tool identified 3,338 instances of n\~ao, classifying 2,085 as NEG1, NEG2, or NEG3, achieving a 93% success rate. However, negativas has limitations. NEG1 accounted for 91.5% of identified structures, while NEG2 and NEG3 represented 7.2% and 1.2%, respectively. The tool struggled with NEG2, sometimes misclassifying instances as overlapping structures (NEG1/NEG2/NEG3).
Task arithmetic has emerged as a promising approach for editing models by representing task-specific knowledge as composable task vectors. However, existing methods rely on network linearization to derive task vectors, leading to computational bottlenecks during training and inference. Moreover, linearization alone does not ensure weight disentanglement, the key property that enables conflict-free composition of task vectors. To address this, we propose TaLoS which allows to build sparse task vectors with minimal interference without requiring explicit linearization and sharing information across tasks. We find that pre-trained models contain a subset of parameters with consistently low gradient sensitivity across tasks, and that sparsely updating only these parameters allows for promoting weight disentanglement during fine-tuning. Our experiments prove that TaLoS improves training and inference efficiency while outperforming current methods in task addition and negation. By enabling modular parameter editing, our approach fosters practical deployment of adaptable foundation models in real-world applications.
Contrastive Language-Image Pre-Training (CLIP) is a popular method for learning multimodal latent spaces with well-organized semantics. Despite its wide range of applications, CLIP's latent space is known to fail at handling complex visual-textual interactions. Recent works attempt to address its shortcomings with data-centric or algorithmic approaches. But what if the problem is more fundamental, and lies in the geometry of CLIP? Toward this end, we rigorously analyze CLIP's latent space properties, and prove that no CLIP-like joint embedding space exists which can correctly do any two of the following at the same time: 1. represent basic descriptions and image content, 2. represent attribute binding, 3. represent spatial location and relationships, 4. represent negation. Informed by this analysis, we propose Dense Cosine Similarity Maps (DCSMs) as a principled and interpretable scoring method for CLIP-like models, which solves the fundamental limitations of CLIP by retaining the semantic topology of the image patches and text tokens. This method improves upon the performance of classical CLIP-like joint encoder models on a wide array of benchmarks. We share our code and data here for reproducibility: https://github.com/Raphoo/DCSM_Ideal_CLIP
Negation is a fundamental aspect of human communication, yet it remains a challenge for Language Models (LMs) in Information Retrieval (IR). Despite the heavy reliance of modern neural IR systems on LMs, little attention has been given to their handling of negation. In this study, we reproduce and extend the findings of NevIR, a benchmark study that revealed most IR models perform at or below the level of random ranking when dealing with negation. We replicate NevIR's original experiments and evaluate newly developed state-of-the-art IR models. Our findings show that a recently emerging category-listwise Large Language Model (LLM) re-rankers-outperforms other models but still underperforms human performance. Additionally, we leverage ExcluIR, a benchmark dataset designed for exclusionary queries with extensive negation, to assess the generalisability of negation understanding. Our findings suggest that fine-tuning on one dataset does not reliably improve performance on the other, indicating notable differences in their data distributions. Furthermore, we observe that only cross-encoders and listwise LLM re-rankers achieve reasonable performance across both negation tasks.
Task vectors, which are derived from the difference between pre-trained and fine-tuned model weights, enable flexible task adaptation and model merging through arithmetic operations such as addition and negation. However, existing approaches often rely on heuristics with limited theoretical support, often leading to performance gaps comparing to direct task fine tuning. Meanwhile, although it is easy to manipulate saved task vectors with arithmetic for different purposes, such compositional flexibility demands high memory usage, especially when dealing with a huge number of tasks, limiting scalability. This work addresses these issues with a theoretically grounded framework that explains task vector arithmetic and introduces the task vector bases framework. Building upon existing task arithmetic literature, our method significantly reduces the memory cost for downstream arithmetic with little effort, while achieving competitive performance and maintaining compositional advantage, providing a practical solution for large-scale task arithmetic.
3D visual grounding (3DVG) aims to locate objects in a 3D scene with natural language descriptions. Supervised methods have achieved decent accuracy, but have a closed vocabulary and limited language understanding ability. Zero-shot methods mostly utilize large language models (LLMs) to handle natural language descriptions, yet suffer from slow inference speed. To address these problems, in this work, we propose a zero-shot method that reformulates the 3DVG task as a Constraint Satisfaction Problem (CSP), where the variables and constraints represent objects and their spatial relations, respectively. This allows a global reasoning of all relevant objects, producing grounding results of both the target and anchor objects. Moreover, we demonstrate the flexibility of our framework by handling negation- and counting-based queries with only minor extra coding efforts. Our system, Constraint Satisfaction Visual Grounding (CSVG), has been extensively evaluated on the public datasets ScanRefer and Nr3D datasets using only open-source LLMs. Results show the effectiveness of CSVG and superior grounding accuracy over current state-of-the-art zero-shot 3DVG methods with improvements of $+7.0\%$ ([email omitted] score) and $+11.2\%$ on the ScanRefer and Nr3D datasets, respectively. The code of our system is publicly available at https://github.com/sunsleaf/CSVG.
The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output. In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation.
Visual hallucination (VH) occurs when a multimodal large language model (MLLM) generates responses with incorrect visual details for prompts. Existing methods for generating VH test cases primarily rely on human annotations, typically in the form of triples: (image, question, answer). In this paper, we introduce VHExpansion, the first automated method for expanding VH test cases for MLLMs. Given an initial VH test case, VHExpansion automatically expands it by perturbing the question and answer through negation as well as modifying the image using both common and adversarial perturbations. Additionally, we propose a new evaluation metric, symmetric accuracy, which measures the proportion of correctly answered VH test-case pairs. Each pair consists of a test case and its negated counterpart. Our theoretical analysis shows that symmetric accuracy is an unbiased evaluation metric that remains unaffected by the imbalance of VH testing cases with varying answers when an MLLM is randomly guessing the answers, whereas traditional accuracy is prone to such imbalance. We apply VHExpansion to expand three VH datasets annotated manually and use these expanded datasets to benchmark seven MLLMs. Our evaluation shows that VHExpansion effectively identifies more VH test cases. Moreover, symmetric accuracy, being unbiased, leads to different conclusions about the vulnerability of MLLMs to VH compared to traditional accuracy metric. Finally, we show that fine-tuning MLLMs on the expanded VH dataset generated by VHExpansion mitigates VH more effectively than fine-tuning on the original, manually annotated dataset. Our code is available at: https://github.com/lycheeefish/VHExpansion.
Machine unlearning aims to selectively remove specific knowledge from a model. Current methods, such as task arithmetic, rely on fine-tuning models on the forget set, generating a task vector, and subtracting it from the original model. However, we argue the effectiveness of this approach is highly sensitive to hyperparameter selection, necessitating careful validation to identify the best model among many fine-tuned candidates. In this paper, we propose a novel method that leverages all given fine-tuned models rather than selecting a single one. By constructing task vectors from models trained with varied hyperparameters and merging only the components of the task vectors with consistent signs, we perform unlearning by negating the merged task vector from the original model. Given that existing methods also utilize multiple fine-tuned models, our approach delivers more effective unlearning without incurring additional computational costs. We demonstrate the effectiveness of our method on both vision-language models and standard image classification models, showing improved unlearning performance with minimal degradation on the retain set, outperforming state-of-the-art techniques.
Provenance in databases has been thoroughly studied for positive and for recursive queries, then for first-order (FO) queries, i.e., having negation but no recursion. Query evaluation can be understood as a two-player game where the opponents argue whether or not a tuple is in the query answer. This game-theoretic approach yields a natural provenance model for FO queries, unifying how and why-not provenance. Here, we study the fine-grain structure of game provenance. A game $G=(V,E)$ consists of positions $V$ and moves $E$ and can be solved by computing the well-founded model of a single, unstratifiable rule: \[ \text{win}(X) \leftarrow \text{move}(X, Y), \neg \, \text{win}(Y). \] In the solved game $G^{\lambda}$, the value of a position $x\,{\in}\,V$ is either won, lost, or drawn. This value is explained by the provenance $\mathscr{P}$(x), i.e., certain (annotated) edges reachable from $x$. We identify seven edge types that give rise to new kinds of provenance, i.e., potential, actual, and primary, and demonstrate that "not all moves are created equal". We describe the new provenance types, show how they can be computed while solving games, and discuss applications, e.g., for abstract argumentation frameworks.
Music-text multimodal systems have enabled new approaches to Music Information Research (MIR) applications such as audio-to-text and text-to-audio retrieval, text-based song generation, and music captioning. Despite the reported success, little effort has been put into evaluating the musical knowledge of Large Language Models (LLM). In this paper, we demonstrate that LLMs suffer from 1) prompt sensitivity, 2) inability to model negation (e.g. 'rock song without guitar'), and 3) sensitivity towards the presence of specific words. We quantified these properties as a triplet-based accuracy, evaluating the ability to model the relative similarity of labels in a hierarchical ontology. We leveraged the Audioset ontology to generate triplets consisting of an anchor, a positive (relevant) label, and a negative (less relevant) label for the genre and instruments sub-tree. We evaluated the triplet-based musical knowledge for six general-purpose Transformer-based models. The triplets obtained through this methodology required filtering, as some were difficult to judge and therefore relatively uninformative for evaluation purposes. Despite the relatively high accuracy reported, inconsistencies are evident in all six models, suggesting that off-the-shelf LLMs need adaptation to music before use.
Multi-hop query answering over a Knowledge Graph (KG) involves traversing one or more hops from the start node to answer a query. Path-based and logic-based methods are state-of-the-art for multi-hop question answering. The former is used in link prediction tasks. The latter is for answering complex logical queries. The logical multi-hop querying technique embeds the KG and queries in the same embedding space. The existing work incorporates First Order Logic (FOL) operators, such as conjunction ($\wedge$), disjunction ($\vee$), and negation ($\neg$), in queries. Though current models have most of the building blocks to execute the FOL queries, they cannot use the dense information of multi-modal entities in the case of Multi-Modal Knowledge Graphs (MMKGs). We propose RConE, an embedding method to capture the multi-modal information needed to answer a query. The model first shortlists candidate (multi-modal) entities containing the answer. It then finds the solution (sub-entities) within those entities. Several existing works tackle path-based question-answering in MMKGs. However, to our knowledge, we are the first to introduce logical constructs in querying MMKGs and to answer queries that involve sub-entities of multi-modal entities as the answer. Extensive evaluation of four publicly available MMKGs indicates that RConE outperforms the current state-of-the-art.
Dual Horn clauses mirror key properties of Horn clauses. This paper explores the ``other side of the looking glass'' to reveal some expected and unexpected symmetries and their practical uses. We revisit Dual Horn clauses as enablers of a form of constructive negation that supports goal-driven forward reasoning and is valid both intuitionistically and classically. In particular, we explore the ability to falsify a counterfactual hypothesis in the context of a background theory expressed as a Dual Horn clause program. With Dual Horn clause programs, by contrast to negation as failure, the variable bindings in their computed answers provide explanations for the reasons why a statement is successfully falsified. Moreover, in the propositional case, by contrast to negation as failure as implemented with stable models semantics in ASP systems, and similarly to Horn clause programs, Dual Horn clause programs have polynomial complexity. After specifying their execution model with a metainterpreter, we devise a compilation scheme from Dual Horn clause programs to Horn clause programs, ensuring their execution with no performance penalty and we design the embedded SymLP language to support combined Horn clause and Dual Horn clause programs. As a (motivating) application, we cast LLM reasoning chains into propositional Horn and Dual Horn clauses that work together to constructively prove and disprove goals and enhance Generative AI with explainability of reasoning chains.
Text-to-image generative models often present issues regarding fairness with respect to certain sensitive attributes, such as gender or skin tone. This study aims to reproduce the results presented in "ITI-GEN: Inclusive Text-to-Image Generation" by Zhang et al. (2023a), which introduces a model to improve inclusiveness in these kinds of models. We show that most of the claims made by the authors about ITI-GEN hold: it improves the diversity and quality of generated images, it is scalable to different domains, it has plug-and-play capabilities, and it is efficient from a computational point of view. However, ITI-GEN sometimes uses undesired attributes as proxy features and it is unable to disentangle some pairs of (correlated) attributes such as gender and baldness. In addition, when the number of considered attributes increases, the training time grows exponentially and ITI-GEN struggles to generate inclusive images for all elements in the joint distribution. To solve these issues, we propose using Hard Prompt Search with negative prompting, a method that does not require training and that handles negation better than vanilla Hard Prompt Search. Nonetheless, Hard Prompt Search (with or without negative prompting) cannot be used for continuous attributes that are hard to express in natural language, an area where ITI-GEN excels as it is guided by images during training. Finally, we propose combining ITI-GEN and Hard Prompt Search with negative prompting.
An adverse drug effect (ADE) is any harmful event resulting from medical drug treatment. Despite their importance, ADEs are often under-reported in official channels. Some research has therefore turned to detecting discussions of ADEs in social media. Impressive results have been achieved in various attempts to detect ADEs. In a high-stakes domain such as medicine, however, an in-depth evaluation of a model's abilities is crucial. We address the issue of thorough performance evaluation in English-language ADE detection with hand-crafted templates for four capabilities: Temporal order, negation, sentiment, and beneficial effect. We find that models with similar performance on held-out test sets have varying results on these capabilities.
Minimal pairs are a well-established approach to evaluating the grammatical knowledge of language models. However, existing resources for minimal pairs address a limited number of languages and lack diversity of language-specific grammatical phenomena. This paper introduces the Russian Benchmark of Linguistic Minimal Pairs (RuBLiMP), which includes 45k pairs of sentences that differ in grammaticality and isolate a morphological, syntactic, or semantic phenomenon. In contrast to existing benchmarks of linguistic minimal pairs, RuBLiMP is created by applying linguistic perturbations to automatically annotated sentences from open text corpora and carefully curating test data. We describe the data collection protocol and present the results of evaluating 25 language models in various scenarios. We find that the widely used language models for Russian are sensitive to morphological and agreement-oriented contrasts but fall behind humans on phenomena requiring understanding of structural relations, negation, transitivity, and tense. RuBLiMP, the codebase, and other materials are publicly available.
This paper introduces a novel, entity-aware metric, termed as Radiological Report (Text) Evaluation (RaTEScore), to assess the quality of medical reports generated by AI models. RaTEScore emphasizes crucial medical entities such as diagnostic outcomes and anatomical details, and is robust against complex medical synonyms and sensitive to negation expressions. Technically, we developed a comprehensive medical NER dataset, RaTE-NER, and trained an NER model specifically for this purpose. This model enables the decomposition of complex radiological reports into constituent medical entities. The metric itself is derived by comparing the similarity of entity embeddings, obtained from a language model, based on their types and relevance to clinical significance. Our evaluations demonstrate that RaTEScore aligns more closely with human preference than existing metrics, validated both on established public benchmarks and our newly proposed RaTE-Eval benchmark.