With new large language models (LLMs) emerging frequently, it is important to consider the potential value of model-agnostic approaches that can provide interpretability across a variety of architectures. While recent advances in LLM interpretability show promise, many rely on complex, model-specific methods with high computational costs. To address these limitations, we propose NormXLogit, a novel technique for assessing the significance of individual input tokens. This method operates based on the input and output representations associated with each token. First, we demonstrate that the norm of word embeddings can be utilized as a measure of token importance. Second, we reveal a significant relationship between a token’s importance and how predictive its representation is of the model’s final output. Extensive analyses indicate that our approach outperforms existing gradient-based methods in terms of faithfulness and offers competitive performance compared to leading architecture-specific techniques.
论文检索
输入标题、作者或关键词，从 8,216 篇学术成果中精准定位
Retrieval-Augmented Generation (RAG) is an effective approach to enhance the factual accuracy of large language models (LLMs) by retrieving information from external databases, which are typically composed of diverse sources, to supplement the limited internal knowledge of LLMs. However, the standard RAG often risks retrieving incorrect information, as it relies solely on relevance between a query and a document, overlooking the heterogeneous reliability of these sources. To address this issue, we propose Reliability-Aware RAG (RA-RAG), a new multi-source RAG framework that estimates the reliability of sources and leverages this information to prioritize highly reliable and relevant documents, ensuring more robust and accurate response generation. Specifically, RA-RAG first estimates source reliability by cross-checking information across multiple sources. It then retrieves documents from the top-\kappa reliable and relevant sources and aggregates their information using weighted majority voting (WMV), where the selective retrieval ensures scalability while not compromising the performance. Comprehensive experiments show that RA-RAG consistently outperforms baselines in scenarios with heterogeneous source reliability while scaling efficiently as the number of sources increases. Furthermore, we demonstrate the ability of RA-RAG to estimate real-world sources’ reliability, highlighting its practical applicability. Our code and data are available at RA-RAG.
Current coreference resolution systems are typically tailored for short- or medium-sized texts and struggle to scale to very long documents due to architectural limitations and implied memory costs.However, a few available solutions can be applied by inputting documents split into smaller windows. This is inherently similar to what happens in the cross-document setting, in which systems infer coreference relations between mentions that are found in separate documents.In this paper, we unify these two challenging settings under the general framework of cross-context coreference, and introduce xCoRe, a new unified approach designed to efficiently handle short-, long-, and cross-document coreference resolution.xCoRe adopts a three-step pipeline that first identifies mentions, then creates clusters within individual contexts, and finally merges clusters across contexts.In our experiments, we show that our formulation enables joint training on shared long- and cross-document resources, increasing data availability and particularly benefiting the challenging cross-document task.Our model achieves new state-of-the-art results on cross-document benchmarks and strong performance on long-document data, while retaining top-tier results on traditional datasets, positioning it as a robust, versatile solution that can be applied across all end-to-end coreference settings.We release our models and code at http://github.com/sapienzanlp/xcore.
MPCG: Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs
PDF ↗Misinformation evolves as it spreads, shifting in language, framing, and moral emphasis to adapt to new audiences. However, current misinformation detection approaches implicitly assume that misinformation is static. We introduce MPCG, a multi-round, persona-conditioned framework that simulates how claims are iteratively reinterpreted by agents with distinct ideological perspectives. Our approach uses an uncensored large language model (LLM) to generate persona-specific claims across multiple rounds, conditioningeach generation on outputs from the previous round, enabling the study of misinformation evolution. We evaluate the generated claims through human and LLM-based annotations, cognitive effort metrics (readability, perplexity), emotion evocation metrics (sentiment analysis, morality), clustering, feasibility, and downstream classification. Results show strong agreement between human and GPT-4o-mini annotations, with higher divergence in fluency judgments. Generated claims require greater cognitive effort than the original claims and consistently reflect persona-aligned emotional and moral framing. Clustering and cosine similarity analyses confirmsemantic drift across rounds while preserving topical coherence. Feasibility results show a 77% feasibility rate, confirming suitability for downstream tasks. Classification results reveal that commonly used misinformation detectors experience macro-F1 performance drops of up to 49.7%. The code is available at https://github.com/bcjr1997/MPCG.
Large Vision-Language Models (LVLMs) have achieved strong performance on vision-language tasks, particularly Visual Question Answering (VQA). While prior work has explored unimodal biases in VQA, the problem of selection bias in Multiple-Choice Question Answering (MCQA), where models may favor specific option tokens (e.g., “A”) or positions, remains underexplored. In this paper, we investigate both the presence and nature of selection bias in LVLMs through fine-grained MCQA benchmarks spanning easy, medium, and hard difficulty levels, defined by the semantic similarity of the options. We further propose an inference-time logit-level debiasing method that estimates an ensemble bias vector from general and contextual prompts and applies confidence-adaptive corrections to the model’s output. Our method mitigates bias without retraining and is compatible with frozen LVLMs. Extensive experiments across several state-of-the-art models reveal consistent selection biases that intensify with task difficulty, and show that our mitigation approach significantly reduces bias while improving accuracy in challenging settings. This work offers new insights into the limitations of LVLMs in MCQA and presents a practical approach to improve their robustness in fine-grained visual reasoning. Datasets and code are available at: https://github.com/Atabuzzaman/Selection-Bias-of-LVLMs
Idiosyncratic Versus Normative Modeling of Atypical Speech Recognition: Dysarthric Case Studies
PDF ↗State-of-the-art automatic speech recognition (ASR) models like Whisper perform poorly on atypical speech, such as that produced by individuals with dysarthria. Past works for atypical speech have mostly investigated fully personalized (or idiosyncratic) models, but modeling strategies that can both generalize and handle idiosyncrasy could be more effective for capturing atypical speech. To investigate this, we compare four strategies: (a) *normative* models trained on typical speech (no personalization), (b) *idiosyncratic* models completely personalized to individuals, (c) *dysarthric-normative* models trained on other dysarthric speakers, and (d) *dysarthric-idiosyncratic* models which combine strategies by first modeling normative patterns before adapting to individual speech. In this case study, we find the dysarthric-idiosyncratic model performs better than the idiosyncratic approach while requiring less than half as much personalized data (36.43 WER with 128 train size vs. 36.99 with 256). Further, we found that tuning the speech encoder alone (as opposed to the LM decoder) yielded the best results, reducing word error rate from 71% to 32% on average. Our findings highlight the value of leveraging both normative (cross-speaker) and idiosyncratic (speaker-specific) patterns to improve ASR for underrepresented speech populations. [GitHub: VishnuRaja98/Dysarthric-Speech-Transcription](https://github.com/VishnuRaja98/Dysarthric-Speech-Transcription)
While virtually all existing work on Automated Essay Scoring (AES) models an essay as a word sequence, we put forward the novel view that an essay can be modeled as a graph and subsequently propose GAT-AES, a graph-attention network approach to AES. GAT-AES models the interactions among essay traits in a principled manner by (1) representing each essay trait as a trait node in the graph and connecting each pair of trait nodes with directed edges, and (2) allowing neighboring nodes to influence each other by using a convolutional operator to update node representations. Unlike competing approaches, which can only model one-hop dependencies, GAT-AES allows us to easily model multi-hop dependencies. Experimental results demonstrate that GAT-AES achieves the best multi-trait scoring results to date on the ASAP++ dataset. Further analysis shows that GAT-AES outperforms not only alternative graph neural networks but also approaches that use trait-attention mechanisms to model trait dependencies.
The recent rise of reasoning-tuned Large Language Models (LLMs)—which generate chains of thought (CoTs) before giving the final answer—has attracted significant attention and offers new opportunities for gaining insights into human label variation, which refers to plausible differences in how multiple annotators label the same data instance.Prior work has shown that LLM-generated explanations can help align model predictions with human label distributions, but typically adopt a *reverse* paradigm: producing explanations based on given answers. In contrast, CoTs provide a *forward* reasoning path that may implicitly embed rationales for each answer option, before generating the answers. We thus propose a novel LLM-based pipeline enriched with linguistically-grounded discourse segmenters to extract supporting and opposing statements for each answer option from CoTs with improved accuracy. We also propose a rank-based HLV evaluation framework that prioritizes the ranking of answers over exact scores, which instead favor direct comparison of label distributions.Our method outperforms a direct generation method as well as baselines on three datasets, and shows better alignment of ranking methods with humans, highlighting the effectiveness of our approach.
Non-English dialogue datasets are scarce, and models are often trained or evaluated on translations of English-language dialogues, an approach which can introduce artifacts that reduce their naturalness and cultural appropriateness. This work proposes Dialogue Act Script (DAS), a structured framework for encoding, localizing, and generating multilingual dialogues from abstract intent representations. Rather than translating dialogue utterances directly, DAS enables the generation of new dialogues in the target language that are culturally and contextually appropriate. By using structured dialogue act representations, DAS supports flexible localization across languages, mitigating translationese and enabling more fluent, naturalistic conversations. Human evaluations across Italian, German, and Chinese show that DAS-generated dialogues consistently outperform those produced by both machine and human translators on measures of cultural relevance, coherence, and situational appropriateness.
Conversational agents powered by large language models (LLMs) are rapidly becoming integral to our daily interactions, generating unprecedented amounts of conversational data. Such datasets offer a powerful lens into societal interests, trending topics, and collective concerns. Yet existing approaches typically treat these interactions as independent, missing critical insights that could emerge from aggregating and reasoning across large-scale conversation logs. In this paper, we introduce Aggregative Question Answering, a novel task requiring models to reason explicitly over thousands of user-chatbot interactions to answer aggregational queries, such as identifying emerging concerns among specific demographics. To enable research in this direction, we construct a benchmark, WildChat-AQA, comprising 6,027 aggregative questions derived from 182,330 real-world chatbot conversations. Experiments show that existing methods either struggle to reason effectively or incur prohibitive computational costs, underscoring the need for new approaches capable of extracting collective insights from large-scale conversational data.
We present NN-Rank, an algorithm for ranking source languages for cross-lingual transfer, which leverages hidden representations from multilingual models and unlabeled target-language data. We experiment with two pretrained multilingual models and two tasks: part-of-speech tagging (POS) and named entity recognition (NER). We consider 51 source languages and evaluate on 56 and 72 target languages for POS and NER, respectively. When using in-domain data, NN-Rank beats state-of-the-art baselines that leverage lexical and linguistic features, with average improvements of up to 35.56 NDCG for POS and 18.14 NDCG for NER. As prior approaches can fall back to language-level features if target language data is not available, we show that NN-Rank remains competitive using only the Bible, an out-of-domain corpus available for a large number of languages. Ablations on the amount of unlabeled target data show that, for subsets consisting of as few as 25 examples, NN-Rank produces high-quality rankings which achieve 92.8% of the NDCG achieved using all available target data for ranking.
REVIVING YOUR MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffing
PDF ↗LLMs are frequently fine-tuned or unlearned to adapt to new tasks or eliminate undesirable behaviors. While existing evaluation methods assess performance after such interventions, there remains no general approach for detecting unintended side effects—such as unlearning biology content degrading performance on chemistry tasks, particularly when these effects are unpredictable or emergent. To address this issue, we introduce MNEME, \textit{Model diffiNg for Evaluating Mechanistic Effects}, a framework for identifying these side effects using sparse model diffing. MNEME compares base and fine-tuned models on out-of-distribution (OOD) data (e.g., The Pile, LMSYS-Chat-1M), without access to fine-tuning data, to isolate behavioral shifts.Applied to five LLMs across three scenarios, WMDP knowledge unlearning, emergent misalignment, and benign fine-tuning, MNEME achieves up to 95% accuracy in predicting side effects, aligning with known benchmarks and requiring no custom heuristics. Our results demonstrate that sparse probing and diffing offer a scalable and automated lens into fine-tuning-induced model changes, providing practical tools for understanding and managing LLM behavior.
Temporal relation extraction (TRE) is a fundamental task in natural language processing (NLP) that involves identifying the temporal relationships between events in a document. Despite the advances in large language models (LLMs), their application to TRE remains limited. Most existing approaches rely on pairwise classification, where event pairs are classified in isolation, leading to computational inefficiency and a lack of global consistency in the resulting temporal graph. In this work, we propose a novel zero-shot method for TRE that generates a document’s complete temporal graph in a single step, followed by temporal constraint optimization to refine predictions and enforce temporal consistency across relations. Additionally, we introduce OmniTemp, a new dataset with complete annotations for all pairs of targeted events within a document. Through experiments and analyses, we demonstrate that our method outperforms existing zero-shot approaches and offers a competitive alternative to supervised TRE models.
Recent advances in test-time scaling have led to the emergence of thinking LLMs that exhibit self-reflective behaviors and multi-step reasoning. While RL drives this self-improvement paradigm, recent studies show that solely RL does not truly instill these new reasoning abilities - it merely draws out behaviors already present in the base models. This raises a question: How can we train models that don’t exhibit such thinking behavior to develop it in the first place? To this end, we propose ThinkTuning, a GRPO-based interactive training approach where we augment the rollouts of a student model with the guidance from a teacher model. A simple idea from classroom practice inspires our method: a teacher poses a problem, lets the student try an answer, then gives corrective feedback–enough to point the mind in the right direction and then show the solution. Each feedback reshapes the student’s thoughts, leading them to arrive at the correct solution. Similarly, we find that this type of implicit supervision through feedback from a teacher model of the same size improves the reasoning capabilities of the student model. Particularly, on average, our method shows 3.69% improvement over zero-shot baselines across benchmarks, and on MATH-500 and GPQA-Diamond, it shows 2.08% and 3.99% improvement over the vanilla-GRPO baseline.
In this paper, we introduce a novel weighted co-training approach that is guided by Large Language Models (LLMs). Namely, in our co-training approach, we use LLM labels on unlabeled data as target labels and co-train two encoder-only based networks that train each other over multiple iterations: first, all samples are forwarded through each network and historical estimates of each network’s confidence in the LLM label are recorded; second, a dynamic importance weight is derived for each sample according to each network’s belief (or confidence) in the quality of the LLM label for that sample; finally, the two networks exchange importance weights with each other—each network back-propagates all samples weighted with the importance weights coming from its peer network and updates its own parameters. By strategically utilizing LLM-generated guidance, our approach significantly outperforms conventional SSL methods, particularly in settings with abundant unlabeled data. Empirical results show that it achieves state-of-the-art performance on 4 out of 5 benchmark datasets and ranks first among 14 compared methods according to the Friedman test. Our results highlight a new direction in semi-supervised learning—where LLMs serve as knowledge amplifiers, enabling backbone co-training models to achieve SOTA performance efficiently.
Innovators often exhibit cognitive fixation on existing solutions or nascent ideas, hindering the exploration of novel alternatives. This paper introduces a methodology for constructing Functional Concept Graphs (FCGs), interconnected representations of functional elements that support abstraction, problem reframing, and analogical inspiration. Our approach yields large-scale, high-quality FCGs with explicit abstraction relations, overcoming limitations of prior work. We further present MUSE, an algorithm leveraging FCGs to generate creative inspirations for a given problem. We demonstrate our method by computing an FCG on 500K patents, which we release for further research. We introduced MUSE, a novel engine to find unexpected solutions to problems. This engine consists of the inspiration graph, whose problem and solution nodes were extracted from 500K patent descriptions. For a given problem, MUSE aims to enhance users’ creative problem solving by providing them with inspirations sampled from the inspiration graph. A user study indicates that participants exposed to MUSE’s inspirations generated more creative ideas, both in terms of absolute number (up to 19% increase over participants not given inspirations) and ratio (75%, compared to 49% for no inspirations).
Enhancing Study-Level Inference from Clinical Trial Papers via Reinforcement Learning-Based Numeric Reasoning
PDF ↗Systematic reviews in medicine play a critical role in evidence-based decision-making by aggregating findings from multiple studies. A central bottleneck in automating this process is extracting numeric evidence and determining study-level conclusions for specific outcomes and comparisons. Prior work has framed this problem as a textual inference task by retrieving relevant content fragments and inferring conclusions from them. However, such approaches often rely on shallow textual cues and fail to capture the underlying numeric reasoning behind expert assessments.In this work, we conceptualise the problem as one of quantitative reasoning. Rather than inferring conclusions from surface text, we extract structured numerical evidence (e.g., event counts or standard deviations) and apply domain knowledge informed logic to derive outcome-specific conclusions. We develop a numeric reasoning system composed of a numeric data extraction model and an effect estimate component, enabling more accurate and interpretable inference aligned with the domain expert principles. We train the numeric data extraction model using different strategies, including supervised fine-tuning (SFT) and reinforcement learning (RL) with a new value reward model.When evaluated on the CochraneForest benchmark, our best-performing approach – using RL to train a small-scalenumber extraction model – yields up to a 21% absolute improvement in F1 score over retrieval-based systems and outperforms general-purpose LLMs of over 400B parameters by up to 9%.Our results demonstrate the promise of reasoning-driven approaches for automating systematic evidence synthesis.
The proliferation of multimodal content on social media presents significant challenges in understanding and moderating complex, context-dependent issues such as misinformation, hate speech, and propaganda. While efforts have been made to develop resources and propose new methods for automatic detection, limited attention has been given to label detection and the generation of explanation-based rationales for predicted labels. To address this challenge, we introduce MemeXplain, an explanation-enhanced dataset for propaganda memes in Arabic and hateful memes in English, making it the first large-scale resource for these tasks. To solve these tasks, we propose a novel multi-stage optimization approach and train Vision-Language Models (VLMs). Our results demonstrate that this approach significantly improves performance over the base model for both label detection and explanation generation, outperforming the current state-of-the-art with an absolute improvement of approximately 3% on ArMeme and 7% on Hateful Memes. For reproducibility and future research, we aim to make the MemeXplain dataset and scripts publicly available.
Large Language Models (LLMs) excel across diverse tasks, with post-training methods like Supervised Fine-Tuning (SFT), Preference Learning, and Model Merging enabling effective domain and task adaptation. While outcomes can vary with data orderings or component combinations, yet manual pipeline optimization is costly and labor-intensive. Existing approaches typically rely on manual design or focus narrowly on optimizing individual components, such as data ordering or merging parameters. We propose LaMDAgent, an LLM Agent-driven framework that autonomously constructs and optimizes end-to-end post-training pipelines by exploring various model improving methods, objects, and their applied orderings based on task-based feedback. LaMDAgent achieves a 9.0-point gain in tool-use accuracy without degrading instruction-following, and identifies high-performing strategies overlooked by manual design.We further analyze the impact of data and model scaling to reduce computational costs on the exploration, finding that model size scalings introduces new challenges, whereas scaling data size enables cost-effective pipeline discovery.
Lemmatization is crucial for NLP tasks in morphologically rich languages with ambiguous orthography like Arabic, but existing tools face challenges due to inconsistent standards and limited genre coverage. This paper introduces two novel approaches that frame lemmatization as classification into a Lemma-POS-Gloss (LPG) tagset, leveraging machine translation and semantic clustering. We also present a new Arabic lemmatization test set covering diverse genres, standardized alongside existing datasets. We evaluate character-level sequence-to-sequence models, which perform competitively and offer complementary value, but are limited to lemma prediction (not LPG) and prone to hallucinating implausible forms. Our results show that classification and clustering yield more robust, interpretable outputs, setting new benchmarks for Arabic lemmatization.