Since the middle of the 20th century, a fierce battle is being fought between symbolic and distributed approaches to language and cognition. The success of deep learning models, and LLMs in particular, has been alternatively taken as showing that the distributed camp has won, or dismissed as an irrelevant engineering development. In this position paper, I argue that deep learning models for language actually represent a synthesis between the two traditions. This is because 1) deep learning architectures allow for both distributed/continuous/fuzzy and symbolic/discrete/categorical-like representations and processing; 2) models trained on language make use of this flexibility. In particular, I review recent research in interpretability that showcases how a substantial part of morphosyntactic knowledge is encoded in a near-discrete fashion in LLMs. This line of research suggests that different behaviors arise in an emergent fashion, and models flexibly alternate between the two modes (and everything in between) as needed. This is possibly one of the main reasons for their wild success; and it makes them particularly interesting for the study of language. Is it time for peace?
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
Metrics like FactScore and VeriScore that evaluate long-form factuality operate by decomposing an input response into atomic claims and then individually verifying each claim. While effective and interpretable, these methods incur numerous LLM calls and can take upwards of 100s to evaluate a single response, limiting their practicality in large-scale evaluation and training scenarios. To address this, we propose VeriFastScore, which leverages synthetic data to fine-tune Llama3.1 8B for simultaneously extracting and verifying all verifiable claims within a given text based on evidence from Google Search. We show that this task cannot be solved via few-shot prompting with closed LLMs due to its complexity: the model receives \sim\mathbf{4K} tokens of evidence on average and needs to concurrently decompose claims, judge their verifiability, and verify them against noisy evidence. However, our fine-tuned VeriFastScore model demonstrates strong correlation with the original VeriScore pipeline at both the example level (r=\mathbf{0.80}) and system level (r=\mathbf{0.94}) while achieving an overall speedup of \mathbf{6.6}\times (9.9 \times excluding evidence retrieval) over VeriScore. To facilitate future factuality research, we publicly release our VeriFastScore model and synthetic datasets.
Graphical User Interface (GUI) interaction, which aims to develop an intelligent GUI agent that executes user instructions to perform tasks such as installing applications by controlling digital devices, has gained significant attention due to its practical value. Although current advanced multimodal large language models (LLMs) provide GUI agents with robust perception and reasoning capabilities, they often struggle with the precise localization of small elements. To tackle this problem, we propose InReAct, a multimodal GUI agent framework that unifies observing, thinking, and acting for precise and interpretable decision-making. It is trained via a two-stage process: curriculum learning to progressively build perception, grounding, and reasoning abilities, followed by reinforcement learning to refine pixel-level grounding with an outcome-based reward. We introduce a rule-based reward function that jointly optimizes action-type selection and pixel-level localization accuracy. Experimental results on multiple datasets demonstrate the superiority of InReAct in both grounding and navigation tasks.
Unveiling Multimodal Processing: Exploring Activation Patterns in Multimodal LLMs for Interpretability and Efficiency
PDF ↗Recent Multimodal Large Language Models (MLLMs) have achieved remarkable advancements, yet their internal mechanisms for concurrently processing diverse modalities like text, image, and audio remain largely opaque. In this paper, we propose a methodology to convert dense MLLMs into fine-grained Mixture-of-Experts (MoE) architectures. This allows us to visually investigate their multimodal activation patterns through expert activation frequency heatmaps. Conducting comprehensive experiments on representative MLLMs, we analyze the similarities and differences in internal neuron activations when handling distinct modalities. Specifically, we examine the distribution of high-frequency activated experts, the distinct roles of high-frequency (e.g., fundamental logic) and low-frequency (e.g., domain-specific concepts) multimodal shared experts, and the prevalence and localization of modality-specific experts. Furthermore, we explore leveraging these discovered activation discrepancies to guide sparse activation and model pruning. Experimental results demonstrate that our approach substantially outperforms random expert pruning and can achieve comparable or even superior performance to the original unpruned models while utilizing significantly fewer active parameters. Our work not only sheds light on the multimodal processing mechanisms within MLLMs but also provides a practical pathway toward developing more interpretable and efficient multimodal systems.
MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation
PDF ↗Knowledge conflict often arises in retrieval-augmented generation (RAG) systems, where retrieved documents may be inconsistent with one another or contradict the model’s parametric knowledge.Existing benchmarks for investigating the phenomenon have notable limitations, including a narrow focus on the question answering setup, heavy reliance on entity substitution techniques, and a restricted range of conflict types. To address these issues, we propose a knowledge graph (KG)-based framework that generates varied and subtle conflicts between two similar yet distinct contexts, while ensuring interpretability through the explicit relational structure of KGs.Experimental results on our benchmark, MAGIC, provide intriguing insights into the inner workings of LLMs regarding knowledge conflict: both open-source and proprietary models struggle with conflict detection—especially when multi-hop reasoning is required—and often fail to pinpoint the exact source of contradictions.Finally, we present in-depth analyses that serve as a foundation for improving LLMs in integrating diverse, sometimes even conflicting, information.
Despite significant progress, large language models (LLMs) still struggle with long contexts due to memory limitations and their inability to tackle complex and long-context tasks. Additionally, LLMs often suffer from a lack of transparency and are prone to producing hallucinations. To address these challenges, we propose JERR, a novel framework designed to enhance long-context comprehension via graph-based reasoning in LLMs. JERR integrates three key components: synopsis extraction, graph construction, and relational reasoning. First, synopsis is extracted by chunking text strategically, allowing the model to summarize and understand information more efficiently. Second, we build a directed acyclic graph (DAG) to resolve redundancy, ensuring logical consistency and clarity. Finally, we incorporate Monte Carlo Tree Search (MCTS) to help the model navigate complex reasoning paths, ensuring more accurate and interpretable outputs. This framework provides a novel solution that enables LLMs to handle extended contexts and complex reasoning tasks with improved reliability and transparency. Experimental results show that JERR consistently outperforms all baselines on the ROUGE and F1 metrics, achieving the highest scores on the LLM-Rater evaluation.
Large Vision–Language Models (LVLMs) have garnered substantial interest owing to their impressive ability to interpret visual inputs and converse with users.Nevertheless, LVLMs still suffer from object hallucination – generating descriptions for objects that are absent from the image, which undermines reliability and hinders real-world deployment. We propose DAPE-BR, a positional-alignment scheme that (i) preserves the pretrained weight order while globally—- visual–text distances, (ii) embeds an isotropic fused patch-distance metric, and (iii) applies a patch-distance causal mask to enforce spatial causality. Extensive experiments on POPE, MMStar and SQA show that DAPE-BR consistently reduces hallucinations and boosts.
Large language models (LLMs) achieve impressive results over various tasks, and ever-expanding public repositories contain an abundance of pre-trained models. Therefore, identifying the best-performing LLM for a given task is a significant challenge. Previous works have suggested learning LLM representations to address this. However, these approaches present limited scalability and require costly retraining to encompass additional models and datasets. Moreover, the produced representation utilizes distinct spaces that cannot be easily interpreted. This work presents an efficient, training-free approach to representing LLMs as linear operators within the prompts’ semantic task space, thus providing a highly interpretable representation of the models’ application. Our method utilizes closed-form computation of geometrical properties and ensures exceptional scalability and real-time adaptability to dynamically expanding repositories. We demonstrate our approach on success prediction and model selection tasks, achieving competitive or state-of-the-art results with notable performance in out-of-sample scenarios.
ZEBRA: Leveraging Model-Behavioral Knowledge for Zero-Annotation Preference Dataset Construction
PDF ↗Recent efforts in LLM alignment have focused on constructing large-scale preference datasets via human or Artificial Intelligence(AI) annotators. However, such approaches rely on instance-wise supervision, incurring substantial annotation cost and limited interpretability. In this paper, we propose **ZEBRA**—a model behavior-wise zero-annotation framework that constructs preference data by leveraging model behavior knowledge derived from benchmark performances.ZEBRA binarizes response pairs by evaluating the quality and similarity of their origin models, entirely bypassing instance-level annotation. This allows scalable, controllable, and cost-effective alignment data generation. Empirical results show that ZEBRA achieves alignment performance comparable to instance-supervised methods, despite requiring no manual or model-based labeling.
Representation learning in knowledge graphs (KGs) has predominantly focused on static data, yet many real-world knowledge graphs are inherently dynamic. For instance, the fact (The CEO of Apple, holds position, Steve Jobs) was valid until 2011, after which it changed, emphasizing the need to incorporate temporal information into knowledge representation. In this paper, we propose 3DG-TE, a novel temporal KG embedding method inspired by 3D Gaussian Splatting, where entities, relations, and timestamps are modeled as 3D Gaussian distributions with learnable structured covariance. This approach optimizes the Gaussian distributions of entities, relations, and timestamps to improve the overall KG representation. To effectively capture temporal-relational interactions, we design structured covariances that form composite transformation operators: relations induce rotational transformations, while timestamps regulate adaptive scaling. We also design a compound scoring function that integrates mean positions and structured covariance, preserving geometric interpretability. Experimental results on three benchmark TKG datasets demonstrate that 3DG-TE outperforms state-of-the-art baselines in temporal link prediction tasks. Theoretical analysis further confirms our model’s ability to capture key relation patterns.
FM2DS: Few-Shot Multimodal Multihop Data Synthesis with Knowledge Distillation for Question Answering
PDF ↗Multimodal multihop question answering (MMQA) requires reasoning over images and text from multiple sources, an essential task for many real-world applications. Despite advances in visual question answering, this multihop setting remains underexplored due to a lack of quality datasets. Existing methods focus on single-hop, single-modality, or short texts, limiting real-world applications like interpreting educational documents with long, multimodal content. To fill this gap, we introduce FM2DS, the first framework for creating a high-quality dataset for MMQA. Our approach consists of a 5-stage pipeline that involves acquiring relevant multimodal documents from Wikipedia, synthetically generating high-level questions and answers, and validating them through rigorous criteria to ensure data quality. We evaluate our methodology by training models on our synthesized dataset and testing on two benchmarks: MultimodalQA and WebQA. Our results demonstrate that, with an equal sample size, models trained on our synthesized data outperform those trained on human-collected data by 1.9 in exact match (EM) score on average. Additionally, we introduce M2QA-Bench with 1k samples, the first benchmark for MMQA on long documents, generated using FM2DS and refined by human annotators.
Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly large language models (LLMs), have enabled virtual screening methods that reduce costs and improve efficiency. However, the growing complexity of these tools has limited their accessibility to wet-lab researchers. Multi-agent systems offer a promising solution by combining the interpretability of LLMs with the precision of specialized models and tools. In this work, we present MADD, a multi-agent system that builds and executes customized hit identification pipelines from natural language queries. MADD employs four coordinated agents to handle key subtasks in de novo compound generation and screening. We evaluate MADD across seven drug discovery cases and demonstrate its superior performance compared to existing LLM-based solutions. Using MADD, we pioneer application of AI-first drug design to five biological targets and release the identified hit molecules. Finally, we introduce a new benchmark of query-molecule pairs and docking scores for over three million compounds to contribute to the agentic future of drug design.
How do Language Models Generate Slang: A Systematic Comparison between Human and Machine-Generated Slang Usages
PDF ↗Slang is a commonly used type of informal language that poses a daunting challenge to NLP systems. Recent advances in large language models (LLMs), however, have made the problem more approachable. While LLM agents are becoming more widely applied to intermediary tasks such as slang detection and slang interpretation, their generalizability and reliability are heavily dependent on whether these models have captured structural knowledge about slang that align well with human attested slang usages. To answer this question, we contribute a systematic comparison between human and machine-generated slang usages. Our evaluative framework focuses on three core aspects: 1) Characteristics of the usages that reflect systematic biases in how machines perceive slang, 2) Creativity reflected by both lexical coinages and word reuses employed by the slang usages, and 3) Informativeness of the slang usages when used as gold-standard examples for model distillation. By comparing human-attested slang usages from the Online Slang Dictionary (OSD) and slang generated by GPT-4o and Llama-3, we find significant biases in how LLMs perceive slang. Our results suggest that while LLMs have captured significant knowledge about the creative aspects of slang, such knowledge does not align with humans sufficiently to enable LLMs for extrapolative tasks such as linguistic analyses.
Multi-label classification (MLC) faces persistent challenges from label imbalance, spurious correlations, and distribution shifts, especially in rare label prediction. We propose the Causal Cooperative Game (CCG) framework, which models MLC as a multi-player cooperative process. CCG integrates explicit causal discovery via Neural Structural Equation Models, a counterfactual curiosity reward to guide robust feature learning, and a causal invariance loss to ensure generalization across environments, along with targeted rare label enhancement. Extensive experiments on benchmark datasets demonstrate that CCG significantly improves rare label prediction and overall robustness compared to strong baselines. Ablation and qualitative analyses further validate the effectiveness and interpretability of each component. Our work highlights the promise of combining causal inference and cooperative game theory for more robust and interpretable multi-label learning.
Understanding the core dimensions of conceptual semantics is fundamental to uncovering how meaning is organized in language and the brain. Existing approaches often rely on predefined semantic dimensions that offer only broad representations, overlooking finer conceptual distinctions. This paper proposes a novel framework to investigate the subdimensions underlying coarse-grained semantic dimensions. Specifically, we introduce a \textbf{D}isentangled \textbf{C}ontinuous \textbf{S}emantic \textbf{R}epresentation \textbf{M}odel (\textbf{DCSRM}) that decomposes word embeddings from large language models into multiple sub-embeddings, each encoding specific semantic information. Using these subembeddings, we identify a set of interpretable semantic subdimensions. To assess their neural plausibility, we apply voxel-wise encoding models to map these subdimensions to brain activation. Our work offers more fine-grained interpretable semantic subdimensions of conceptual meaning. Further analyses reveal that semantic dimensions are structured according to distinct principles, with polarity emerging as a key factor driving their decomposition into subdimensions. The neural correlates of the identified subdimensions support their cognitive and neuroscientific plausibility.
Personalized alignments towards individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Unlike traditional Reinforcement Learning from Human Feedback (RLHF), which relies on vast annotated datasets and expensive gradient updates, Drift operates in a training-free manner by steering a frozen LLM through few-shot preference modeling. Our approach represents user preferences as a composition of interpretable and predefined attributes, and employs a zero-shot rewarding mechanism based on contrastive system prompts. Experiments on both a synthetic persona dataset Perspective and a real human-annotated dataset PRISM demonstrate that Drift achieves performance comparable to standard RLHF methods while using only 50–100 examples. Our results show that Drift delivers not only computationally efficient but also interpretable personalization.
Large Language Models (LLMs), e.g. ChatGPT, have been widely adopted in real-world dialogue applications. However, LLMs’ robustness, especially in handling long complex dialogue sessions, including frequent motivation transfer, sophisticated cross-turn dependency, is criticized all along. Nevertheless, no existing benchmarks can fully reflect these weaknesses. We present MARS-Bench, a Multi-turn Athletic Real-world Scenario Dialogue Benchmark, designed to remedy the gap. MARS-Bench is constructed from play-by-play text commentary so to feature realistic dialogues specifically designed to evaluate three critical aspects of multi-turn conversations: ultra multi-turn, interactive multi-turn, and cross-turn tasks. Extensive experiments on MARS-Bench also reveal that closed-source LLMs significantly outperform open-source alternatives, explicit reasoning significantly boosts LLMs’ robustness on handling long complex dialogue sessions, and LLMs indeed face significant challenge when handling motivation transfer and sophisticated cross-turn dependency. Moreover, we provide mechanistic interpretability on how attention sinks due to special tokens lead to LLMs’ performance degradation when handling long complex dialogue sessions based on attention visualization experiment in Qwen2.5-7B-Instruction.
Large language models (LLMs) have shown promise in table Question Answering (Table QA). However, extending these capabilities to multi-table QA remains challenging due to unreliable schema linking across complex tables. Existing methods based on semantic similarity work well only on simplified hand-crafted datasets and struggle to handle complex, real-world scenarios with numerous and diverse columns. To address this, we propose a graph-based framework that leverages human-curated relational knowledge to explicitly encode schema links and join paths. Given a natural language query, our method searches on graph to construct interpretable reasoning chains, aided by pruning and sub-path merging strategies to enhance efficiency and coherence. Experiments on both standard benchmarks and a realistic, large-scale dataset demonstrate the effectiveness of our approach. To our knowledge, this is the first multi-table QA system applied to truly complex industrial tabular data.
Large Language Models (LLMs) have transformed listwise document reranking by enabling global reasoning over candidate sets, yet single models often struggle to balance fine-grained relevance scoring with holistic cross-document analysis. We propose DeepAgentRank (DeAR), an open-source framework that decouples these tasks through a dual-stage approach, achieving superior accuracy and interpretability. In Stage 1, we distill token-level relevance signals from a frozen 13B LLaMA teacher into a compact 3, 8B student model using a hybrid of cross-entropy, RankNet, and KL divergence losses, ensuring robust pointwise scoring. In Stage 2, we attach a second LoRA adapter and fine-tune on 20K GPT-4o-generated chain-of-thought permutations, enabling listwise reasoning with natural-language justifications. Evaluated on TREC-DL19/20, eight BEIR datasets, and NovelEval-2306, DeAR surpasses open-source baselines by +5.1 nDCG@5 on DL20 and achieves 90.97 nDCG@10 on NovelEval, outperforming GPT-4 by +3.09. Without fine-tuning on Wikipedia, DeAR also excels in open-domain QA, achieving 54.29 Top-1 accuracy on Natural Questions, surpassing baselines like MonoT5, UPR, and RankGPT. Ablations confirm that dual-loss distillation ensures stable calibration, making DeAR a highly effective and interpretable solution for modern reranking systems.
Cross-lingual topic modeling aims to uncover shared semantic themes across languages. Several methods have been proposed to address this problem, leveraging both traditional and neural approaches. While previous methods have achieved some improvements in topic diversity, they often struggle to ensure high topic coherence and consistent alignment across languages. We propose XTRA (Cross-Lingual Topic Modeling with Topic and Representation Alignments), a novel framework that unifies Bag-of-Words modeling with multilingual embeddings. XTRA introduces two core components: (1) representation alignment, aligning document-topic distributions via contrastive learning in a shared semantic space; and (2) topic alignment, projecting topic-word distributions into the same space to enforce cross-lingual consistency. This dual mechanism enables XTRA to learn topics that are interpretable (coherent and diverse) and well-aligned across languages. Experiments on multilingual corpora confirm that XTRA significantly outperforms strong baselines in topic coherence, diversity, and alignment quality.