Despite significant advances in large language models (LLMs), their knowledge memorization capabilities remain underexplored, due to the lack of standardized and high-quality test ground. In this paper, we introduce a novel, real-world and large-scale knowledge injection benchmark that evolves continuously over time without requiring human intervention. Specifically, we propose WikiDYK, which leverages recently-added and human-written facts from Wikipedia's "Did You Know..." entries. These entries are carefully selected by expert Wikipedia editors based on criteria such as verifiability and clarity. Each entry is converted into multiple question-answer pairs spanning diverse task formats from easy cloze prompts to complex multi-hop questions. WikiDYK contains 12,290 facts and 77,180 questions, which is also seamlessly extensible with future updates from Wikipedia editors. Extensive experiments using continued pre-training reveal a surprising insight: despite their prevalence in modern LLMs, Causal Language Models (CLMs) demonstrate significantly weaker knowledge memorization capabilities compared to Bidirectional Language Models (BiLMs), exhibiting a 23% lower accuracy in terms of reliability. To compensate for the smaller scales of current BiLMs, we introduce a modular collaborative framework utilizing ensembles of BiLMs as external knowledge repositories to integrate with LLMs. Experiment shows that our framework further improves the reliability accuracy by up to 29.1%.
Fine-tuning large language models (LLMs) on a mixture of diverse datasets poses challenges due to data imbalance and heterogeneity. Existing methods often address these issues across datasets (globally) but overlook the imbalance and heterogeneity within individual datasets (locally), which limits their effectiveness. We introduce Hierarchical Balancing Optimization (HBO), a novel method that enables LLMs to autonomously adjust data allocation during fine-tuning both across datasets (globally) and within each individual dataset (locally). HBO employs a bilevel optimization strategy with two types of actors: a Global Actor, which balances data sampling across different subsets of the training mixture, and several Local Actors, which optimizes data usage within each subset based on difficulty levels. These actors are guided by reward functions derived from the LLM's training state, which measure learning progress and relative performance improvement. We evaluate HBO on three LLM backbones across nine diverse tasks in multilingual and multitask setups. Results show that HBO consistently outperforms existing baselines, achieving significant accuracy gains. Our in-depth analysis further demonstrates that both the global actor and local actors of HBO effectively adjust data usage during fine-tuning. HBO provides a comprehensive solution to the challenges of data imbalance and heterogeneity in LLM fine-tuning, enabling more effective training across diverse datasets.
Evaluating machine translation (MT) for low-resource languages poses a persistent challenge, primarily due to the limited availability of high quality reference translations. This issue is further exacerbated in languages with multiple dialects, where linguistic diversity and data scarcity hinder robust evaluation. Large Language Models (LLMs) present a promising solution through reference-free evaluation techniques; however, their effectiveness diminishes in the absence of dialect-specific context and tailored guidance. In this work, we propose a comprehensive framework that enhances LLM-based MT evaluation using a dialect guided approach. We extend the ONUBAD dataset by incorporating Sylheti-English sentence pairs, corresponding machine translations, and Direct Assessment (DA) scores annotated by native speakers. To address the vocabulary gap, we augment the tokenizer vocabulary with dialect-specific terms. We further introduce a regression head to enable scalar score prediction and design a dialect-guided (DG) prompting strategy. Our evaluation across multiple LLMs shows that the proposed pipeline consistently outperforms existing methods, achieving the highest gain of +0.1083 in Spearman correlation, along with improvements across other evaluation settings. The dataset and the code are available at https://github.com/180041123-Atiq/MTEonLowResourceLanguage.
Recent progress in large language models (LLMs) has outpaced the development of effective evaluation methods. Traditional benchmarks rely on task-specific metrics and static datasets, which often suffer from fairness issues, limited scalability, and contamination risks. In this paper, we introduce Teach2Eval, an indirect evaluation framework inspired by the Feynman Technique. Instead of directly testing LLMs on predefined tasks, our method evaluates a model's multiple abilities to teach weaker student models to perform tasks effectively. By converting open-ended tasks into standardized multiple-choice questions (MCQs) through teacher-generated feedback, Teach2Eval enables scalable, automated, and multi-dimensional assessment. Our approach not only avoids data leakage and memorization but also captures a broad range of cognitive abilities that are orthogonal to current benchmarks. Experimental results across 26 leading LLMs show strong alignment with existing human and model-based dynamic rankings, while offering additional interpretability for training guidance.
Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model's predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4$\times$ speedup.
Despite their remarkable success and deployment across diverse workflows, language models sometimes produce untruthful responses. Our limited understanding of how truthfulness is mechanistically encoded within these models jeopardizes their reliability and safety. In this paper, we propose a method for identifying representations of truthfulness at the neuron level. We show that language models contain truth neurons, which encode truthfulness in a subject-agnostic manner. Experiments conducted across models of varying scales validate the existence of truth neurons, confirming that the encoding of truthfulness at the neuron level is a property shared by many language models. The distribution patterns of truth neurons over layers align with prior findings on the geometry of truthfulness. Selectively suppressing the activations of truth neurons found through the TruthfulQA dataset degrades performance both on TruthfulQA and on other benchmarks, showing that the truthfulness mechanisms are not tied to a specific dataset. Our results offer novel insights into the mechanisms underlying truthfulness in language models and highlight potential directions toward improving their trustworthiness and reliability.
Content moderation research has recently made significant advances, but still fails to serve the majority of the world's languages due to the lack of resources, leaving millions of vulnerable users to online hostility. This work presents a large-scale human-annotated multi-task benchmark dataset for abusive language detection in Tigrinya social media with joint annotations for three tasks: abusiveness, sentiment, and topic classification. The dataset comprises 13,717 YouTube comments annotated by nine native speakers, collected from 7,373 videos with a total of over 1.2 billion views across 51 channels. We developed an iterative term clustering approach for effective data selection. Recognizing that around 64% of Tigrinya social media content uses Romanized transliterations rather than native Ge'ez script, our dataset accommodates both writing systems to reflect actual language use. We establish strong baselines across the tasks in the benchmark, while leaving significant challenges for future contributions. Our experiments reveal that small, specialized multi-task models outperform the current frontier models in the low-resource setting, achieving up to 86% accuracy (+7 points) in abusiveness detection. We make the resources publicly available to promote research on online safety.
Large Language Models (LLMs) have shown impressive capabilities across various tasks but remain vulnerable to meticulously crafted jailbreak attacks. In this paper, we identify a critical safety gap: while LLMs are adept at detecting jailbreak prompts, they often produce unsafe responses when directly processing these inputs. Inspired by this insight, we propose SAGE (Self-Aware Guard Enhancement), a training-free defense strategy designed to align LLMs' strong safety discrimination performance with their relatively weaker safety generation ability. SAGE consists of two core components: a Discriminative Analysis Module and a Discriminative Response Module, enhancing resilience against sophisticated jailbreak attempts through flexible safety discrimination instructions. Extensive experiments demonstrate SAGE's effectiveness and robustness across various open-source and closed-source LLMs of different sizes and architectures, achieving an average 99% defense success rate against numerous complex and covert jailbreak methods while maintaining helpfulness on general benchmarks. We further conduct mechanistic interpretability analysis through hidden states and attention distributions, revealing the underlying mechanisms of this detection-generation discrepancy. Our work thus contributes to developing future LLMs with coherent safety awareness and generation behavior. Our code and datasets are publicly available at https://github.com/NJUNLP/SAGE.
We present GenderBench -- a comprehensive evaluation suite designed to measure gender biases in LLMs. GenderBench includes 14 probes that quantify 19 gender-related harmful behaviors exhibited by LLMs. We release GenderBench as an open-source and extensible library to improve the reproducibility and robustness of benchmarking across the field. We also publish our evaluation of 12 LLMs. Our measurements reveal consistent patterns in their behavior. We show that LLMs struggle with stereotypical reasoning, equitable gender representation in generated texts, and occasionally also with discriminatory behavior in high-stakes scenarios, such as hiring.
Identifying hate speech content in the Arabic language is challenging due to the rich quality of dialectal variations. This study introduces a multilabel hate speech dataset in the Arabic language. We have collected 10000 Arabic tweets and annotated each tweet, whether it contains offensive content or not. If a text contains offensive content, we further classify it into different hate speech targets such as religion, gender, politics, ethnicity, origin, and others. A text can contain either single or multiple targets. Multiple annotators are involved in the data annotation task. We calculated the inter-annotator agreement, which was reported to be 0.86 for offensive content and 0.71 for multiple hate speech targets. Finally, we evaluated the data annotation task by employing a different transformers-based model in which AraBERTv2 outperformed with a micro-F1 score of 0.7865 and an accuracy of 0.786.
Although multimodal large language models (MLLMs) show promise in generating chart rendering code, chart editing presents a greater challenge. This difficulty stems from its nature as a labor-intensive task for humans that also demands MLLMs to integrate chart understanding, complex reasoning, and precise intent interpretation. While many MLLMs claim such editing capabilities, current assessments typically rely on limited case studies rather than robust evaluation methodologies, highlighting the urgent need for a comprehensive evaluation framework. In this work, we propose ChartEdit, a new high-quality benchmark designed for chart editing tasks. This benchmark comprises $1,405$ diverse editing instructions applied to $233$ real-world charts, with each instruction-chart instance having been manually annotated and validated for accuracy. Utilizing ChartEdit, we evaluate the performance of 10 mainstream MLLMs across two types of experiments, assessing them at both the code and chart levels. The results suggest that large-scale models can generate code to produce images that partially match the reference images. However, their ability to generate accurate edits according to the instructions remains limited. The state-of-the-art (SOTA) model achieves a score of only $59.96$, highlighting significant challenges in precise modification. In contrast, small-scale models, including chart-domain models, struggle both with following editing instructions and generating overall chart images, underscoring the need for further development in this area. Code is available at https://github.com/xxlllz/ChartEdit.
Precise recognition of search intent in Retrieval-Augmented Generation (RAG) systems remains a challenging goal, especially under resource constraints and for complex queries with nested structures and dependencies. This paper presents QCompiler, a neuro-symbolic framework inspired by linguistic grammar rules and compiler design, to bridge this gap. It theoretically designs a minimal yet sufficient Backus-Naur Form (BNF) grammar $G[q]$ to formalize complex queries. Unlike previous methods, this grammar maintains completeness while minimizing redundancy. Based on this, QCompiler includes a Query Expression Translator, a Lexical Syntax Parser, and a Recursive Descent Processor to compile queries into Abstract Syntax Trees (ASTs) for execution. The atomicity of the sub-queries in the leaf nodes ensures more precise document retrieval and response generation, significantly improving the RAG system's ability to address complex queries.
Large Language Models (LLMs) have achieved impressive progress in natural language processing, but their limited ability to retain long-term context constrains performance on document-level or multi-turn tasks. Retrieval-Augmented Generation (RAG) mitigates this by retrieving relevant information from an external corpus. However, existing RAG systems often rely on embedding-based retrieval trained on corpus-level semantic similarity, which can lead to retrieving content that is semantically similar in form but misaligned with the question's true intent. Furthermore, recent RAG variants construct graph- or hierarchy-based structures to improve retrieval accuracy, resulting in significant computation and storage overhead. In this paper, we propose an embedding-free retrieval framework. Our method leverages the logical inferencing ability of LLMs in retrieval using iterative search space refinement guided by our novel importance measure and extend our retrieval results with logically related information without explicit graph construction. Experiments on long-context QA benchmarks, including NovelQA and Marathon, show that our approach outperforms strong baselines while reducing storage and runtime by over an order of magnitude.
Model editing techniques are essential for efficiently updating knowledge in large language models (LLMs). However, the effectiveness of existing approaches degrades in massive editing scenarios, particularly when evaluated with practical metrics or in context-rich settings. We attribute these failures to embedding collisions among knowledge items, which undermine editing reliability at scale. To address this, we propose NAMET (Noise-aware Model Editing in Transformers), a simple yet effective method that introduces noise during memory extraction via a one-line modification to MEMIT. Extensive experiments across six LLMs and three datasets demonstrate that NAMET consistently outperforms existing methods when editing thousands of facts.
Human preference plays a crucial role in the refinement of large language models (LLMs). However, collecting human preference feedback is costly and most existing datasets neglect the correlation between personalization and preferences. To address this issue, we introduce Fair-PP, a synthetic dataset of personalized preferences targeting social equity, derived from real-world social survey data, which includes 28 social groups, 98 equity topics, and 5 personal preference dimensions. Leveraging GPT-4o-mini, we engage in role-playing based on seven representative persona portrayals guided by existing social survey data, yielding a total of 238,623 preference records. Through Fair-PP, we also contribute (i) An automated framework for generating preference data, along with a more fine-grained dataset of personalized preferences; (ii) analysis of the positioning of the existing mainstream LLMs across five major global regions within the personalized preference space; and (iii) a sample reweighting method for personalized preference alignment, enabling alignment with a target persona while maximizing the divergence from other personas. Empirical experiments show our method outperforms the baselines.
The robustness and security of large language models (LLMs) has become a prominent research area. One notable vulnerability is the ability to bypass LLM safeguards by translating harmful queries into rare or underrepresented languages, a simple yet effective method of "jailbreaking" these models. Despite the growing concern, there has been limited research addressing the safeguarding of LLMs in multilingual scenarios, highlighting an urgent need to enhance multilingual safety. In this work, we investigate the correlation between various attack features across different languages and propose Multilingual Collaborative Defense (MCD), a novel learning method that optimizes a continuous, soft safety prompt automatically to facilitate multilingual safeguarding of LLMs. The MCD approach offers three advantages: First, it effectively improves safeguarding performance across multiple languages. Second, MCD maintains strong generalization capabilities while minimizing false refusal rates. Third, MCD mitigates the language safety misalignment caused by imbalances in LLM training corpora. To evaluate the effectiveness of MCD, we manually construct multilingual versions of commonly used jailbreak benchmarks, such as MaliciousInstruct and AdvBench, to assess various safeguarding methods. Additionally, we introduce these datasets in underrepresented (zero-shot) languages to verify the language transferability of MCD. The results demonstrate that MCD outperforms existing approaches in safeguarding against multilingual jailbreak attempts while also exhibiting strong language transfer capabilities. Our code is available at https://github.com/HLiang-Lee/MCD.
Deep learning models have driven significant progress in predicting protein function and interactions at the protein level. While these advancements have been invaluable for many biological applications such as enzyme engineering and function annotation, a more detailed perspective is essential for understanding protein functional mechanisms and evaluating the biological knowledge captured by models. To address this demand, we introduce VenusX, the first large-scale benchmark for fine-grained functional annotation and function-based protein pairing at the residue, fragment, and domain levels. VenusX comprises three major task categories across six types of annotations, including residue-level binary classification, fragment-level multi-class classification, and pairwise functional similarity scoring for identifying critical active sites, binding sites, conserved sites, motifs, domains, and epitopes. The benchmark features over 878,000 samples curated from major open-source databases such as InterPro, BioLiP, and SAbDab. By providing mixed-family and cross-family splits at three sequence identity thresholds, our benchmark enables a comprehensive assessment of model performance on both in-distribution and out-of-distribution scenarios. For baseline evaluation, we assess a diverse set of popular and open-source models, including pre-trained protein language models, sequence-structure hybrids, structure-based methods, and alignment-based techniques. Their performance is reported across all benchmark datasets and evaluation settings using multiple metrics, offering a thorough comparison and a strong foundation for future research. Code and data are publicly available at https://github.com/ai4protein/VenusX.
Large Language Models (LLMs) possess extensive knowledge and commonsense reasoning capabilities, making them valuable for creating powerful agents. However, existing LLM agent frameworks have not fully utilized past experiences for improvement. This work introduces a new LLM-based agent framework called Retrospex, which addresses this challenge by analyzing past experiences in depth. Unlike previous approaches, Retrospex does not directly integrate experiences into the LLM's context. Instead, it combines the LLM's action likelihood with action values estimated by a Reinforcement Learning (RL) Critic, which is trained on past experiences through an offline ''retrospection'' process. Additionally, Retrospex employs a dynamic action rescoring mechanism that increases the importance of experience-based values for tasks that require more interaction with the environment. We evaluate Retrospex in ScienceWorld, ALFWorld and Webshop environments, demonstrating its advantages over strong, contemporary baselines.
Interpretability research now offers a variety of techniques for identifying abstract internal mechanisms in neural networks. Can such techniques be used to predict how models will behave on out-of-distribution examples? In this work, we provide a positive answer to this question. Through a diverse set of language modeling tasks--including symbol manipulation, knowledge retrieval, and instruction following--we show that the most robust features for correctness prediction are those that play a distinctive causal role in the model's behavior. Specifically, we propose two methods that leverage causal mechanisms to predict the correctness of model outputs: counterfactual simulation (checking whether key causal variables are realized) and value probing (using the values of those variables to make predictions). Both achieve high AUC-ROC in distribution and outperform methods that rely on causal-agnostic features in out-of-distribution settings, where predicting model behaviors is more crucial. Our work thus highlights a novel and significant application for internal causal analysis of language models.
It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying "true features" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Our work shows there remain fundamental issues with SAEs, but we are hopeful that that highlighting feature hedging will catalyze future advances that allow SAEs to achieve their full potential of interpreting LLMs at scale.