As the insight of knowledge storage in language models deepens, the ability to perform CRUD (Create, Read, Update, Delete) operations on language models becomes increasingly indispensable for satisfying the demands of managing rapidly updating knowledge. Considering the high cost of fine-tuning language models, model editing methods with low cost are usually required to manipulate models’ knowledge. The evidence suggests that modules carrying knowledge in a Transformer module are primarily the MLP blocks, thus we propose iReVa, a method that explicitly initializes and retrofits key-value pairs into MLP blocks to construct a new mapping of a piece of knowledge without damaging the irrelevant knowledge. In comparison to existing methods, iReVa reveals better interpretability and a stronger capacity for carrying traceable edits. Experiment results on a series of GPT series models show our prominent performance on edit success and generalization without influencing specificity. We also made the first attempt to conduct a knowledge withdrawal test of iReVa. Our codes are available at https://github.com/timberflow/iReVa.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Motivated by the surge of large language models, there has been a push to formally characterize the symbolic abilities intrinsic to the transformer architecture. A programming language, called RASP, has been proposed, which can be directly compiled into transformer weights to implement these algorithms. However, the tasks that can be implemented in RASP are often uncommon to learn from natural unsupervised data, showing a mismatch between theoretical capabilities of the transformer architecture, and the practical learnability of these capabilities from unsupervised data. We propose tracr-injection, a method that allows us to distill algorithms written in RASP directly into a pre-trained language model. We showcase our method by injecting 3 different algorithms into a language model. We show how our method creates an interpretable subspace within the model’s residual stream, which can be decoded into the variables present in the code of the RASP algorithm. Additionally, we found that the proposed method can improve out-of-distribution performance compared to our baseline, indicating that indeed a more symbolic mechanism is taking place in the inner workings of the model. We release the code used to run our experiments.
Do Language Models Understand the Cognitive Tasks Given to Them? Investigations with the N-Back Paradigm
PDF ↗Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5’s declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least in part to a limitation in task comprehension and task set maintenance. We challenge the best-performing model with progressively harder versions of the task (up to 10-back) and experiment with alternative prompting strategies, before analyzing model attentions. Our larger aim is to contribute to the ongoing conversation around refining methodologies for the cognitive evaluation of language models.
Large Language Models (LLMs), such as the GPT-4 and LLaMA families, have demonstrated considerable success across diverse tasks, including multiple-choice questions (MCQs). However, these models exhibit a positional bias, particularly an even worse “anchored bias” in the GPT-2 family, where they consistently favour the first choice ‘A’ in MCQs during inference. This anchored bias challenges the integrity of GPT-2’s decision-making process, as it skews performance based on the position rather than the content of the choices in MCQs. In this study, we utilise the mechanistic interpretability approach to identify the internal modules within GPT-2 models responsible for this bias. We focus on the Multi-Layer Perceptron (MLP) layers and attention heads, using the “logit lens” method to trace and modify the specific value vectors that contribute to the bias. By updating these vectors within MLP and recalibrating attention patterns to neutralise the preference for the first choice ‘A’, we effectively mitigate the anchored bias. Our interventions not only mitigate the bias but also improve the overall MCQ prediction accuracy for the GPT-2 family across various datasets. This work represents the first comprehensive mechanistic analysis of anchored bias from the failing cases in MCQs within the GPT-2 models, introducing targeted, minimal-intervention strategies that significantly enhance GPT2 model robustness and accuracy in MCQs. Our code is available at https://github.com/ruizheliUOA/Anchored_Bias_GPT2.
To enhance the interpretability of multimodal unified representations, many studies have focused on discrete unified representations. These efforts typically start with contrastive learning and gradually extend to the disentanglement of modal information, achieving solid multimodal discrete unified representations. However, existing research often overlooks two critical issues: 1) The use of Euclidean distance for quantization in discrete representations often overlooks the important distinctions among different dimensions of features, resulting in redundant representations after quantization; 2) Different modalities have unique characteristics, and a uniform alignment approach does not fully exploit these traits. To address these issues, we propose Training-free Optimization of Codebook (TOC) and Fine and Coarse cross-modal Information Disentangling (FCID). These methods refine the unified discrete representations from pretraining and perform fine- and coarse-grained information disentanglement tailored to the specific characteristics of each modality, achieving significant performance improvements over previous state-of-the-art models. The code is available at https://github.com/haihuangcode/CMG.
Chain-Talker: Chain Understanding and Rendering for Empathetic Conversational Speech Synthesis
PDF ↗Conversational Speech Synthesis (CSS) aims to align synthesized speech with the emotional and stylistic context of user-agent interactions to achieve empathy. Current generative CSS models face interpretability limitations due to insufficient emotional perception and redundant discrete speech coding. To address the above issues, we present Chain-Talker, a three-stage framework mimicking human cognition: Emotion Understanding derives context-aware emotion descriptors from dialogue history; Semantic Understanding generates compact semantic codes via serialized prediction; and Empathetic Rendering synthesizes expressive speech by integrating both components. To support emotion modeling, we develop CSS-EmCap, an LLM-driven automated pipeline for generating precise conversational speech emotion captions. Experiments on three benchmark datasets demonstrate that Chain-Talker produces more expressive and empathetic speech than existing methods, with CSS-EmCap contributing to reliable emotion modeling. The code and demos are available at: https://github.com/AI-S2-Lab/Chain-Talker.
Large language models (LLMs) often generate hallucinated content, making it crucial to identify and quantify inconsistencies in their outputs. We introduce HaluMap, a post-hoc framework that detects hallucinations by mapping entailment and contradiction relations between source inputs and generated outputs using a natural language inference (NLI) model. To improve reliability, we propose a calibration step leveraging intra-text relations to refine predictions. HaluMap outperforms state-of-the-art NLI-based methods by five percentage points compared to other training-free approaches, while providing clear, interpretable explanations. As a training-free and model-agnostic approach, HaluMap offers a practical solution for verifying LLM outputs across diverse NLP tasks. The resources of this paper are available at https://github.com/caisa-lab/acl25-halumap.
Most existing GUI agents typically depend on non-vision inputs like HTML source code or accessibility trees, limiting flexibility across diverse software environments and platforms. Current multimodal large language models (MLLMs), though excel at using vision to ground real-world objects, often struggle with accurately localizing GUI elements – a critical requirement for effective GUI automation – due to the semantic gap between real-world objects and GUI elements. In this work, we introduce Ponder & Press, a divide-and-conquer framework for general computer control that uses only visual input. Our approach combines a general-purpose MLLM as an ‘interpreter’, responsible for translating high-level user instructions into detailed action descriptions, with a GUI-specific MLLM as a ‘locator’ that precisely locates GUI elements for action placement. By leveraging a purely visual input, our agent offers a versatile, human-like interaction paradigm applicable to various applications. Ponder & Press locator outperforms existing models by +22.5% on the ScreenSpot GUI grounding benchmark. More offline and interactive agent benchmarks across various GUI environments – including web pages, desktop software, and mobile UIs – demonstrate that the Ponder & Press framework achieves state-of-the-art performance, highlighting the potential of visual GUI agents.
PersonaBench: Evaluating AI Models on Understanding Personal Information through Accessing (Synthetic) Private User Data
PDF ↗Personalization is essential for AI assistants, especially in private AI settings where models are expected to interpret users’ personal data (e.g., conversations, app usage) to understand their background, preferences, and social context. However, due to privacy concerns, existing academic research lacks direct access to such data, making benchmarking difficult. To fill this gap, we propose a synthetic data pipeline that generates realistic user profiles and private documents, enabling the creation of PersonaBench—a benchmark for evaluating models’ ability to understand personal information. Using this benchmark, we assess Retrieval-Augmented Generation (RAG) pipelines on personalized questions and find that current models struggle to accurately extract and answer questions even when provided with the full set of user documents, highlighting the need for improved personalization methods.
Analysing the differences in how events are represented across texts, or verifying whether the language model generations hallucinate, requires the ability to systematically compare their content. To support such comparison, structured representation that captures fine-grained information plays a vital role.In particular, identifying distinct atomic facts and the discourse relations connecting them enables deeper semantic comparison. Our proposed approach combines structured discourse information extraction with a classifier, FDSpotter, for factual consistency verification. We show that adversarial discourse relations pose challenges for language models, but fine-tuning on our annotated data, DiscInfer, achieves competitive performance. Our proposed approach advances factual consistency verification by grounding in linguistic structure and decomposing it into interpretable components. We demonstrate the effectiveness of our method on the evaluation of two tasks: data-to-text generation and text summarisation. Our code and dataset will be publicly available on GitHub.
Large language models (LLMs) have made significant strides in text-to-SQL tasks; however, small language models (SLMs) are crucial due to their low resource consumption and efficient inference for real-world deployment. Due to resource limitations, SLMs struggle to accurately interpret natural language questions and may overlook critical constraints, leading to challenges such as generating SQL with incorrect logic or incomplete conditions. To address these issues, we propose PARSQL, a novel framework that leverages SQL parsing and reasoning. Specifically, we design PARSer, an SQL parser that extracts constraints from SQL to generate sub-SQLs for data augmentation and producing step-by-step SQL explanations (reason) via both rule-based and LLM-based methods. We define a novel text-to-reason task and incorporate it into multi-task learning, thereby enhancing text-to-SQL performance. Additionally, we employ an efficient SQL selection strategy that conducts direct similarity computation between the generated SQLs and their corresponding reasons to derive the final SQL for post-correction. Extensive experiments show that our PARSQL outperforms models with the same model size on the BIRD and Spider benchmarks. Notably, PARSQL-3B achieves 56.98% execution accuracy on BIRD, rivaling 7B models with significantly fewer parameters, setting a new state-of-the-art performance. Code can be found [here](https://github.com/yaxundai/parsql).
Analyzing ideological discourse even in the age of LLMs remains a challenge, as these models often struggle to capture the key elements that shape real-world narratives. Specifically, LLMs fail to focus on characteristic elements driving dominant discourses and lack the ability to integrate contextual information required for understanding abstract ideological views. To address these limitations, we propose a framework motivated by the theory of ideological discourse analysis to analyze news articles related to real-world events. Our framework represents the news articles using a relational structure−talking points, which captures the interaction between entities, their roles, and media frames along with a topic of discussion. It then constructs a vocabulary of repeating themes−prominent talking points, that are used to generate ideology-specific viewpoints (or partisan perspectives). We evaluate our framework’s ability to generate these perspectives through automated tasks−ideology and partisan classification tasks, supplemented by human validation. Additionally, we demonstrate straightforward applicability of our framework in creating event snapshots, a visual way of interpreting event discourse. We release resulting dataset and model to the community to support further research.
Determining and ranking the most salient entities in a text is critical for user-facing systems, especially as users increasingly rely on models to interpret long documents they only partially read. Graded entity salience addresses this need by assigning entities scores that reflect their relative importance in a text. Existing approaches fall into two main categories: subjective judgments of salience, which allow for gradient scoring but lack consistency, and summarization-based methods, which define salience as mention-worthiness in a summary, promoting explainability but limiting outputs to binary labels (entities are either summary-worthy or not). In this paper, we introduce a novel approach for graded entity salience that combines the strengths of both approaches. Using an English dataset spanning 12 spoken and written genres, we collect 5 summaries per document and calculate each entity’s salience score based on its presence across these summaries. Our approach shows stronger correlation with scores based on human summaries and alignments, and outperforms existing techniques, including LLMs. We release our data and code at https://github.com/jl908069/gum_sum_salience to support further research on graded salient entity extraction.
Recent studies on interpreting the hidden states of speech models have shown their ability to capture speaker-specific features, including gender. Does this finding also hold for speech translation (ST) models? If so, what are the implications for the speaker’s gender assignment in translation? We address these questions from an interpretability perspective, using probing methods to assess gender encoding across diverse ST models. Results on three language directions (English \rightarrow French/Italian/Spanish) indicate that while traditional encoder-decoder models capture gender information, newer architectures—integrating a speech encoder with a machine translation system via adapters—do not. We also demonstrate that low gender encoding capabilities result in systems’ tendency toward a masculine default, a translation bias that is more pronounced in newer architectures.
Mitigating entity bias is a critical challenge in Relation Extraction (RE), where models often rely excessively on entities, resulting in poor generalization. This paper presents a novel approach to address this issue by adapting a Variational Information Bottleneck (VIB) framework. Our method compresses entity-specific information while preserving task-relevant features. It achieves state-of-the-art performance on both general and financial domain RE datasets, excelling in in-domain settings (original test sets) and out-of-domain (modified test sets with type-constrained entity replacements). Our approach offers a robust, interpretable, and theoretically grounded methodology.
Do Multimodal Large Language Models Truly See What We Point At? Investigating Indexical, Iconic, and Symbolic Gesture Comprehension
PDF ↗Understanding hand gestures is essential for human communication, yet it remains unclear how well multimodal large language models (MLLMs) comprehend them. In this paper, we examine MLLMs’ ability to interpret indexical gestures, which require external referential grounding, in comparison to iconic gestures, which depict imagery, and symbolic gestures, which are conventionally defined. We hypothesize that MLLMs, lacking real-world referential understanding, will struggle significantly with indexical gestures. To test this, we manually annotated five gesture type labels to 925 gesture instances from the Miraikan SC Corpus and analyzed gesture descriptions generated by state-of-the-art MLLMs, including GPT-4o. Our findings reveal a consistent weakness across models in interpreting indexical gestures, suggesting that MLLMs rely heavily on linguistic priors or commonsense knowledge rather than grounding their interpretations in visual or contextual cues.
How well do modern long-context language models understand literary fiction? We explore this question via the task of literary evidence retrieval, repurposing the RELiC dataset of Thai et al. (2022) to construct a benchmark where the entire text of a primary source (e.g., The Great Gatsby) is provided to an LLM alongside literary criticism with a missing quotation from that work. This setting, in which the model must generate the missing quotation, mirrors the human process of literary analysis by requiring models to perform both global narrative reasoning and close textual examination. We curate a high-quality subset of 292 examples through extensive filtering and human verification. Our experiments show that recent reasoning models, such as Gemini 2.5 Pro can exceed human expert performance (62.5% vs. 50% accuracy). In contrast, the best open-weight model achieves only 29.1% accuracy, highlighting a wide gap in interpretive reasoning between open and closed-weight models. Despite their speed and apparent accuracy, even the strongest models struggle with nuanced literary signals and overgeneration, signaling open challenges for applying LLMs to literary analysis. We release our dataset and evaluation code to encourage future work in this direction.
Whether large language models (LLMs) process language similarly to humans has been the subject of much theoretical and practical debate. We examine this question through the lens of the production-interpretation distinction found in human sentence processing and evaluate the extent to which instruction-tuned LLMs replicate this distinction. Using an empirically documented asymmetry between pronoun production and interpretation in humans for implicit causality verbs as a testbed, we find that some LLMs do quantitatively and qualitatively reflect human-like asymmetries between production and interpretation. We demonstrate that whether this behavior holds depends upon both model size-with larger models more likely to reflect human-like patterns and the choice of meta-linguistic prompts used to elicit the behavior. Our codes and results are available here.
Despite substantial research efforts evaluating how well large language models (LLMs) handle global cultural diversity, the mechanisms behind their cultural knowledge acquisition, particularly in multilingual settings, remain unclear. We study this question by investigating how cultural knowledge transfers across languages during the language adaptation of LLMs, a process where an LLM is continually pre-trained to learn another language. We introduce an interpretable framework to study this transfer, ensuring training data transparency and controlling transfer effects. Through a study of four non-Anglophonic cultures, we observe bidirectional cultural transfer between English and other high-resource languages, while low-resource languages primarily transfer knowledge to English with limited reverse flow. To explain this asymmetric phenomenon, we propose a frequency-based hypothesis: cultural knowledge appearing more frequently in the pretraining data transfers more easily, which is supported by empirical analysis of the training corpora. We hope our findings could inform future research on knowledge transfer and promote the development of culturally aware models, particularly for low-resource languages.
A large part of human communication relies on nonverbal cues such as facial expressions, eye contact, and body language. Unlike language or sign language, such nonverbal communication lacks formal rules, requiring complex reasoning based on commonsense understanding.Enabling current Video Large Language Models (VideoLLMs) to accurately interpret body language is a crucial challenge, as human unconscious actions can easily cause the model to misinterpret their intent.To address this, we propose a dataset, BQA, a body language question answering dataset, to validate whether the model can correctly interpret emotions from short clips of body language comprising 26 emotion labels of videos of body language.We evaluated various VideoLLMs on the BQA with and without Multimodal Chain of Thought (CoT) and revealed that understanding body language is challenging, and our analyses of the wrong answers by VideoLLMs show that certain VideoLLMs made largely biased answers depending on the age group and ethnicity of the individuals. We also found consistent error patterns in VideoLLMs.