Papers

Detecting reclaimed slurs represents a fundamental challenge for hate speech detection systems, as the same lexcal items can function either as abusive expressions or as in-group affirmations depending on social identity and context. In this work, we address Subtask B of the MultiPRIDE shared task at EVALITA 2026 by proposing a hierarchical approach to modeling the slur reclamation process. Our core assumption is that members of the LGBTQ+ community are more likely, on average, to employ certain slurs in a eclamatory manner. Based on this hypothesis, we decompose the task into two stages. First, using a weakly supervised LLM-based annotation, we assign fuzzy labels to users indicating the likelihood of belonging to the LGBTQ+ community, inferred from the tweet and the user bio. These soft labels are then used to train a BERT-like model to predict community membership, encouraging the model to learn latent representations associated with LGBTQ+ identity. In the second stage, we integrate this latent space with a newly initialized model for the downstream slur reclamation detection task. The intuition is that the first model encodes user-oriented sociolinguistic signals, which are then fused with representations learned by a model pretrained for hate speech detection. Experimental results on Italian and Spanish show that our approach achieves performance statistically comparable to a strong BERT-based baseline, while providing a modular and extensible framework for incorporating sociolinguistic context into hate speech modeling. We argue that more fine-grained hierarchical modeling of user identity and discourse context may further improve the detection of reclaimed language. We release our code at https://github.com/LucaTedeschini/multipride.

When humans and large language models (LLMs) process the same text, activations in the LLMs correlate with brain activity measured, e.g., with functional magnetic resonance imaging (fMRI). Moreover, it has been shown that, as the training of an LLM progresses, the performance in predicting brain activity from its internal activations improves more in the left hemisphere than in the right one. The aim of the present work is to understand which kind of competence acquired by the LLMs underlies the emergence of this left-right asymmetry. Using the OLMo-2 7B language model at various training checkpoints and fMRI data from English participants, we compare the evolution of the left-right asymmetry in the correlation between brain activity and model predictions alongside performance on several benchmarks. We observe that the asymmetry co-emerges with the formal linguistic abilities of the LLM. These abilities are demonstrated in two ways: by the model's capacity to assign a higher probability to an acceptable sentence than to a grammatically unacceptable one within a minimal contrasting pair, and by its ability to produce well-formed text. By contrast, the left-right asymmetry does not align with the performance on arithmetic or Dyck language tasks; nor with text-based tasks involving world knowledge and reasoning. We generalize these results to another family of LLMs (Pythia) and two other languages, French and Chinese. Our observations indicate that the left-right asymmetry in brain predictivity matches the progress in formal linguistic competence.

Data containing personal information is increasingly used to train, fine-tune, or query Large Language Models (LLMs). Text is typically scrubbed of identifying information prior to use, often with tools such as Microsoft's Presidio or Anthropic's PII purifier. These tools have traditionally been evaluated on their ability to remove specific identifiers (e.g., names), yet their effectiveness at preventing re-identification remains unclear. We introduce RAT-Bench, a comprehensive benchmark for text anonymization tools based on re-identification risk. Using U.S. demographic statistics, we generate synthetic text containing various direct and indirect identifiers across domains, languages, and difficulty levels. We evaluate a range of NER- and LLM-based text anonymization tools and, based on the attributes an LLM-based attacker is able to correctly infer from the anonymized text, we report the risk of re-identification in the U.S. population, while properly accounting for the disparate impact of identifiers. We find that, while capabilities vary widely, even the best tools are far from perfect in particular when direct identifiers are not written in standard ways and when indirect identifiers enable re-identification. Overall we find LLM-based anonymizers, including new iterative anonymizers, to provide a better privacy-utility trade-off albeit at a higher computational cost. Importantly, we also find them to work well across languages. We conclude with recommendations for future anonymization tools and will release the benchmark and encourage community efforts to expand it, in particular to other geographies.

This study advances aspect-based sentiment analysis (ABSA) for Persian-language user reviews in the tourism domain, addressing challenges of low-resource languages. We propose a hybrid BERT-based model with Top-K routing and auxiliary losses to mitigate routing collapse and improve efficiency. The pipeline includes: (1) overall sentiment classification using BERT on 9,558 labeled reviews, (2) multi-label aspect extraction for six tourism-related aspects (host, price, location, amenities, cleanliness, connectivity), and (3) integrated ABSA with dynamic routing. The dataset consists of 58,473 preprocessed reviews from the Iranian accommodation platform Jabama, manually annotated for aspects and sentiments. The proposed model achieves a weighted F1-score of 90.6% for ABSA, outperforming baseline BERT (89.25%) and a standard hybrid approach (85.7%). Key efficiency gains include a 39% reduction in GPU power consumption compared to dense BERT, supporting sustainable AI deployment in alignment with UN SDGs 9 and 12. Analysis reveals high mention rates for cleanliness and amenities as critical aspects. This is the first ABSA study focused on Persian tourism reviews, and we release the annotated dataset to facilitate future multilingual NLP research in tourism.

Standard evaluation in NLP typically indicates that system A is better on average than system B, but it provides little info on how to improve performance and, what is worse, it should not come as a surprise if B ends up being better than A on outside data. We propose an evaluation methodology for sequence labeling tasks grounded on error analysis that provides both quantitative and qualitative information on where systems must be improved and predicts how models will perform on a different distribution. The key is to create test sets that, contrary to common practice, do not rely on gathering large amounts of real-world in-distribution scraped data, but consists in handcrafting a small set of linguistically motivated examples that exhaustively cover the range of span attributes (such as shape, length, casing, sentence position, etc.) a system may encounter in the wild. We demonstrate this methodology on a benchmark for anglicism identification in Spanish. Our methodology provides results that are diagnostic (because they help identify systematic weaknesses in performance), actionable (because they can inform which model is better suited for a given scenario) and predictive: our method predicts model performance on external datasets with a median correlation of 0.85.

Despite their impressive performance, self-supervised speech models often struggle to generalize to new languages and tend to forget previously acquired knowledge during continual training. To address this, we propose Lamer-SSL, a parameter-efficient framework that integrates a Layer-Aware MixturE of LoRA Experts (Lamer) module with a replay strategy. The Lamer module enables flexible balancing between shared and language-specific representations, while layer-aware expert allocation assigns more experts to deeper layers where semantic information is richer. Meanwhile, the replay strategy retains prior knowledge using minimal data, mitigating forgetting during continual training. Experiments on automatic speech recognition (ASR) and language identification (LID) demonstrate that Lamer-SSL extends self-supervised models to new languages effectively while maintaining strong performance on previously learned languages with only 2.14% parameters being trainable.

Effectively retrieving, reasoning, and understanding multimodal information remains a critical challenge for agentic systems. Traditional Retrieval-augmented Generation (RAG) methods rely on linear interaction histories, which struggle to handle long-context tasks, especially those involving information-sparse yet token-heavy visual data in iterative reasoning scenarios. To bridge this gap, we introduce VimRAG, a framework tailored for multimodal Retrieval-augmented Reasoning across text, images, and videos. Inspired by our systematic study, we model the reasoning process as a dynamic directed acyclic graph that structures the agent states and retrieved multimodal evidence. Building upon this structured memory, we introduce a Graph-Modulated Visual Memory Encoding mechanism, with which the significance of memory nodes is evaluated via their topological position, allowing the model to dynamically allocate high-resolution tokens to pivotal evidence while compressing or discarding trivial clues. To implement this paradigm, we propose a Graph-Guided Policy Optimization strategy. This strategy disentangles step-wise validity from trajectory-level rewards by pruning memory nodes associated with redundant actions, thereby facilitating fine-grained credit assignment. Extensive experiments demonstrate that VimRAG consistently achieves state-of-the-art performance on diverse multimodal RAG benchmarks. The code is available at https://github.com/Alibaba-NLP/VRAG.

Retrieval-augmented generation (RAG) has become a dominant paradigm for grounding large language models (LLMs) with external evidence in knowledge-intensive question answering. A core design choice is how to fuse retrieved samples into the LLMs, where existing internal fusion approaches broadly fall into query-based fusion, parametric fusion, and latent-based fusion. Despite their effectiveness at modest retrieval scales, these methods often fail to scale gracefully as the number of retrieved candidates k increases: Larger k improves evidence coverage, yet realistic top-k retrieval inevitably contains irrelevant or redundant content and increases the inference cost. To address these limitations, we propose ReFilter, a novel latent-based fusion framework that performs token-level filtering and fusion. ReFilter consists of three key components: a context encoder for encoding context features, a gated filter for weighting each token, and a token fusion module for integrating the weighted token feature into the LLM's hidden states. Our experiments across four general-domain QA benchmarks show that ReFilter consistently achieves the best average performance under both in-domain adaptation and out-of-domain transfer. ReFilter further generalizes to five biomedical QA benchmarks in zero-shot transfer without domain fine-tuning, reaching 70.01% average accuracy with Qwen2.5-14B-Instruct.

We present MedXIAOHE, a medical vision-language foundation model designed to advance general-purpose medical understanding and reasoning in real-world clinical applications. MedXIAOHE achieves state-of-the-art performance across diverse medical benchmarks and surpasses leading closed-source multimodal systems on multiple capabilities. To achieve this, we propose an entity-aware continual pretraining framework that organizes heterogeneous medical corpora to broaden knowledge coverage and reduce long-tail gaps (e.g., rare diseases). For medical expert-level reasoning and interaction, MedXIAOHE incorporates diverse medical reasoning patterns via reinforcement learning and tool-augmented agentic training, enabling multi-step diagnostic reasoning with verifiable decision traces. To improve reliability in real-world use, MedXIAOHE integrates user-preference rubrics, evidence-grounded reasoning, and low-hallucination long-form report generation, with improved adherence to medical instructions. We release this report to document our practical design choices, scaling insights, and evaluation framework, hoping to inspire further research.

Knowledge Distillation (KD) for Large Language Models (LLMs) has become increasingly important as models grow in size and complexity. While existing distillation approaches focus on imitating teacher behavior, they often overlook the original learning environment that shaped the teacher's knowledge. Inspired by the experiential learning theory and inverse reinforcement learning, we propose Experiential Knowledge Distillation ($\mathcal{X}$-KD), a novel and general framework that enables student models to learn in the teacher's original learning environment. $\mathcal{X}$-KD adopts the Approximated Variational Reward Imitation Learning (AVRIL) framework to jointly model the teacher's original reward function and perform policy distillation, encouraging consistency between the student policy and the original reward function. Our derivation demonstrates that $\mathcal{X}$-KD follows the supervised learning framework and applies to both sequence-level and divergence-based distillation methods, underlining the simplicity and flexibility of our approach. Empirical results show that $\mathcal{X}$-KD outperforms the generalized KD and MiniLLM baselines on abstractive summarization, machine translation, and arithmetic reasoning tasks. Additionally, $\mathcal{X}$-KD achieves better performance-diversity trade-off and data efficiency than baseline KD approaches.

Large language models (LLMs) are increasingly deployed as autonomous agents for multi-turn decision-making tasks. However, current agents typically rely on fixed cognitive patterns: non-thinking models generate immediate responses, while thinking models engage in deep reasoning uniformly. This rigidity is inefficient for long-horizon tasks, where cognitive demands vary significantly from step to step, with some requiring strategic planning and others only routine execution. In this paper, we introduce CogRouter, a framework that trains agents to dynamically adapt cognitive depth at each step. Grounded in ACT-R theory, we design four hierarchical cognitive levels ranging from instinctive responses to strategic planning. Our two-stage training approach includes Cognition-aware Supervised Fine-tuning (CoSFT) to instill stable level-specific patterns, and Cognition-aware Policy Optimization (CoPO) for step-level credit assignment via confidence-aware advantage reweighting. The key insight is that appropriate cognitive depth should maximize the confidence of the resulting action. Experiments on ALFWorld and ScienceWorld demonstrate that CogRouter achieves state-of-the-art performance with superior efficiency. With Qwen2.5-7B, it reaches an 82.3% success rate, outperforming GPT-4o (+40.3%), OpenAI-o3 (+18.3%), and GRPO (+14.0%), while using 62% fewer tokens.

Reward models are crucial for aligning large language models (LLMs) with human values and intentions. Existing approaches follow either Generative (GRMs) or Discriminative (DRMs) paradigms, yet both suffer from limitations: GRMs typically demand costly point-wise supervision, while DRMs produce uncalibrated relative scores that lack probabilistic interpretation. To address these challenges, we introduce a novel reward modeling paradigm: Probabilistic Reward Model (PRM). Instead of modeling reward as a deterministic scalar, our approach treats it as a random variable, learning a full probability distribution for the quality of each response. To make this paradigm practical, we present its closed-form, discrete realization: the Ordinal Probabilistic Reward Model (OPRM), which discretizes the quality score into a finite set of ordinal ratings. Building on OPRM, we propose a data-efficient training strategy called Region Flooding Tuning (RgFT). It enables rewards to better reflect absolute text quality by incorporating quality-level annotations, which guide the model to concentrate the probability mass within corresponding rating sub-regions. Experiments on various reward model benchmarks show that our method improves accuracy by $\textbf{2.9%}\sim\textbf{7.4%}$ compared to prior reward models, demonstrating strong performance and data efficiency. Analysis of the score distribution provides evidence that our method captures not only relative rankings but also absolute quality.

Reward-maximizing RL methods have shown to be capable of enhancing the reasoning performance of LLMs, but often lead to reduced generation diversity. Recent works address this issue by adopting GFlowNets, training LLMs to match a target distribution while jointly learning its partition function. In contrast to prior works that treat this partition function solely as a normalizer, we reinterpret it as a per-prompt expected-reward (i.e., online accuracy) signal, leveraging this unused information to improve sample efficiency. Specifically, we first establish a theoretical relationship between the partition function and per-prompt accuracy estimates. Building on this key insight, we propose Partition Function-Guided RL (PACED-RL), a post-training framework that leverages accuracy estimates to prioritize informative question prompts during training, and further improves sample efficiency through an accuracy estimate error-prioritized replay. Crucially, both components reuse information already produced during GFlowNet training, effectively amortizing the compute overhead into the existing optimization process. Extensive experiments across diverse benchmarks demonstrate strong performance improvements over GRPO and prior GFlowNet approaches, highlighting PACED-RL as a promising direction for a more sample efficient distribution-matching training for LLMs.

Legal text generated by large language models (LLMs) can usually achieve reasonable factual accuracy, but it frequently fails to adhere to the specialised stylistic norms and linguistic conventions of legal writing. In order to improve stylistic quality, a crucial first step is to establish a reliable evaluation method. However, having legal experts manually develop such a metric is impractical, as the implicit stylistic requirements in legal writing practice are difficult to formalise into explicit rubrics. Meanwhile, existing automatic evaluation methods also fall short: reference-based metrics conflate semantic accuracy with stylistic fidelity, and LLM-as-a-judge evaluations suffer from opacity and inconsistency. To address these challenges, we introduce CLASE (Chinese LegAlese Stylistic Evaluation), a hybrid evaluation method that focuses on the stylistic performance of legal text. The method incorporates a hybrid scoring mechanism that combines 1) linguistic feature-based scores and 2) experience-guided LLM-as-a-judge scores. Both the feature coefficients and the LLM scoring experiences are learned from contrastive pairs of authentic legal documents and their LLM-restored counterparts. This hybrid design captures both surface-level features and implicit stylistic norms in a transparent, reference-free manner. Experiments on 200 Chinese legal documents show that CLASE achieves substantially higher alignment with human judgments than traditional metrics and pure LLM-as-a-judge methods. Beyond improved alignment, CLASE provides interpretable score breakdowns and suggestions for improvements, offering a scalable and practical solution for professional stylistic evaluation in legal text generation (Code and data for CLASE is available at: https://github.com/rexera/CLASE).

As LLMs scale, low-bit floating-point formats like MXFP and NVFP4 offer new opportunities for precision and efficiency. In this work, we evaluate HiFloat (HiF8 and HiF4), a family of formats tailored for Ascend NPUs. Through rigorous comparison across weight-activation and KV-cache tasks, we provide three key insights: (1) INT8 suits narrow-range data, while floating-point formats excel with high-variance data; (2) in 4-bit regimes, HiF4's hierarchical scaling prevents the accuracy collapse seen in integer formats; and (3) HiFloat is fully compatible with state-of-the-art post-training quantization frameworks. Overall, HiFloat provides a solution for high-efficiency LLM inference on NPUs.

Multimodal Large Language Models (MLLMs) incur significant computational cost from processing numerous vision tokens through all LLM layers. Prior pruning methods operate either before the LLM, limiting generality due to diverse encoder-projector designs or within the LLM using heuristics that are incompatible with FlashAttention. We take a different approach: rather than identifying unimportant tokens, we treat the LLM itself as the optimal guide for compression. Observing that deeper layers naturally transmit vision-to-text information, we introduce Attention-Driven Self-Compression (ADSC), a simple, broadly applicable method that progressively reduces vision tokens using only the LLM's attention mechanism. Our method applies uniform token downsampling at selected layers, forming bottlenecks that encourage the model to reorganize and compress information into the remaining tokens. It requires no score computation, auxiliary modules, or attention modification, and remains fully compatible with FlashAttention. Applied to LLaVA-1.5, ADSC reduces FLOPs by 53.7% and peak KV-cache memory by 56.7%, while preserving 98.2% of the original model performance. Across multiple benchmarks, it outperforms prior pruning approaches in both efficiency and accuracy. Crucially, under high compression ratios, our method remains robust while heuristic-based techniques degrade sharply.

Self-attention is often viewed as probabilistic query-key lookup, motivating designs that preserve normalized attention scores and fixed positional semantics. We advocate a simpler and more unified perspective: an autoregressive attention head can be viewed as a dynamic two-layer MLP whose weights are instantiated from the context history. From this view, attention scores form an ever-growing hidden representation, and standard MLP activations such as ReLU or GLU naturally implement input-conditioned selection over a context-dependent memory pool rather than a probability distribution. Based on this formulation, we introduce HyperMLP and HyperGLU, which learn dynamic mixing in both feature space and sequence space, using a reverse-offset (lag) layout to align temporal mixing with autoregressive semantics. We provide theoretical characterizations of the expressivity and implications of this structure, and empirically show that HyperMLP/HyperGLU consistently outperform strong softmax-attention baselines under matched parameter budgets.

Psychological scale refinement traditionally relies on response-based methods such as factor analysis, item response theory, and network psychometrics to uncover latent constructs and optimize item composition. Although rigorous, these approaches require large samples and may be constrained by data availability and cross-cultural comparability. Recent advances in natural language processing suggest that the wording of scale items contains construct-relevant semantic organization, offering a complementary response-free perspective. We introduce a topic-modeling framework that operationalizes item-level semantic structure for systematic scale simplification. Items are encoded using contextual sentence embeddings and grouped through density-based clustering to identify semantic topics without predefining their number. Within each cluster, class-based term weighting identifies representative keywords that summarize each topic and support an interpretable semantic pre-structure of the item pool. Semantically related clusters are hierarchically merged, and representative items are selected using structure-aware membership criteria within an integrated reduction pipeline. We evaluated the framework on three widely used instruments: DASS, IPIP, and EPOCH, covering negative affect, personality traits, and well-being. Results showed that the identified semantic topics often formed factor-like groupings aligned with established constructs. Representative items selected through semantic membership probabilities reduced item counts by 60.5% on average while maintaining acceptable psychometric properties and preserving factor correlation patterns. Our framework formalizes semantic topic structure as an inspectable front-end for scale construction and simplification. To facilitate adoption, we provide a visualization-supported tool enabling one-click semantic analysis and structured simplification.

We present a decoder-only Conformer for automatic speech recognition (ASR) that processes speech and text in a single stack without external speech encoders or pretrained large language models (LLM). The model uses a modality-aware sparse mixture of experts (MoE): disjoint expert pools for speech and text with hard routing and top-1 selection, embedded in hybrid-causality Conformer blocks (bidirectional for speech, causal for text). Training combines CTC on speech positions with label-smoothed cross-entropy for text generation. Our 113M-parameter model consistently improves WER over a 139M AED baseline on Librispeech (2.8% vs. 3.2% test-clean; 5.6% vs. 6.0% test-other). On Common Voice 16.1 with a single multilingual model across five languages, our approach reduces average WER from 12.2% to 10.6%. To our knowledge, this is the first randomly initialized decoder-only ASR that surpasses strong AED baselines via modality-aware routing and sparse MoE, achieving better accuracy with fewer active parameters and without alignment/adaptation modules.

Recent advances in large language models (LLMs) have inspired new paradigms for document reranking. While this paradigm better exploits the reasoning and contextual understanding capabilities of LLMs, most existing LLM-based rerankers rely on autoregressive generation, which limits their efficiency and flexibility. In particular, token-by-token decoding incurs high latency, while the fixed left-to-right generation order causes early prediction errors to propagate and is difficult to revise. To address these limitations, we explore the use of diffusion language models (dLLMs) for document reranking and propose DiffuRank, a reranking framework built upon dLLMs. Unlike autoregressive models, dLLMs support more flexible decoding and generation processes that are not constrained to a left-to-right order, and enable parallel decoding, which may lead to improved efficiency and controllability. Specifically, we investigate three reranking strategies based on dLLMs: (1) a pointwise approach that uses dLLMs to estimate the relevance of each query-document pair; (2) a logit-based listwise approach that prompts dLLMs to jointly assess the relevance of multiple documents and derives ranking lists directly from model logits; and (3) a permutation-based listwise approach that adapts the canonical decoding process of dLLMs to the reranking tasks. For each approach, we design corresponding training methods to fully exploit the advantages of dLLMs. We evaluate both zero-shot and fine-tuned reranking performance on multiple benchmarks. Experimental results show that dLLMs achieve performance comparable to, and in some cases exceeding, that of autoregressive LLMs with similar model sizes. These findings demonstrate the promise of diffusion-based language models as a compelling alternative to autoregressive architectures for document reranking.

Chain-of-Thought (CoT) has significantly enhanced the reasoning capabilities of Large Language Models (LLMs), especially when combined with reinforcement learning (RL) based post-training methods. While longer reasoning traces can improve answer quality and unlock abilities such as self-correction, they also incur high inference costs and often introduce redundant steps, known as overthinking. Recent research seeks to develop efficient reasoning strategies that balance reasoning length and accuracy, either through length-aware reward design or prompt-based calibration. However, these heuristic-based approaches may suffer from severe accuracy drop and be very sensitive to hyperparameters. To address these problems, we introduce CRT (Constraint-Rectified Training), a principled post-training framework based on reference-guarded constrained optimization, yielding a more stable and interpretable formulation for efficient reasoning. CRT alternates between minimizing reasoning length and rectifying accuracy only when performance falls below the reference, enabling stable and effective pruning of redundant reasoning. We further extend CRT with a two-stage training scheme that first discovers the shortest reliable reasoning patterns and then refines accuracy under a learnt length budget, preventing the re-emergence of verbose CoT. Our comprehensive evaluation shows that this framework consistently reduces token usage while maintaining answer quality at a robust and reliable level. Further analysis reveals that CRT improves reasoning efficiency not only by shortening responses but also by reducing internal language redundancy, leading to a new evaluation metric. Moreover, CRT-based training naturally yields a sequence of intermediate checkpoints that span a spectrum of explanation lengths while preserving correctness, enabling fine-grained control over reasoning verbosity without retraining.

Retrieval-augmented generation over semi-structured sources such as HTML is constrained by a mismatch between document structure and the flat, sequence-based interfaces of today's embedding and generative models. Retrieval pipelines often linearize documents into fixed-size chunks before indexing, which obscures section structure, lists, and tables, and makes it difficult to return small, citation-ready evidence without losing the surrounding context that makes it interpretable. We present a structure-aware retrieval pipeline that operates over tree-structured documents. The core idea is to represent candidates as subdocuments: precise, addressable selections that preserve structural identity while deferring the choice of surrounding context. We define a small set of document primitives–paths and path sets, subdocument extraction by pruning, and two contextualization mechanisms. Global contextualization adds the non-local scaffolding needed to make a selection intelligible (e.g., titles, headers, list and table structure). Local contextualization expands a seed selection within its structural neighborhood to obtain a compact, context-rich view under a target budget. Building on these primitives, we describe an embedding-based candidate generator that indexes sentence-seeded subdocuments and a query-time, document-aware aggregation step that amortizes shared structural context. We then introduce a contextual filtering stage that re-scores retrieved candidates using locally contextualized views. Across experiments on HTML question-answering benchmarks, we find that preserving structure while contextualizing selections yields higher-quality, more diverse citations under fixed budgets than strong passage-based baselines, while maintaining scalability.

Hallucinations, the tendency for large language models to provide responses with factually incorrect and unsupported claims, is a serious problem within natural language processing for which we do not yet have an effective solution to mitigate against. Existing benchmarks for medical QA rarely evaluate this behavior against a fixed evidence source. We ask how often hallucinations occur on textbook-grounded QA and how responses to medical QA prompts vary across models. We conduct two experiments, the first experiment to determine the prevalence of hallucinations for a prominent open source large language model (LLaMA-70B-Instruct) in medical QA given closed-source zero-shot prompts, and the second experiment to determine the prevalence of hallucinations and clinician preference to model responses. We observed, in experiment one, with the passages provided, LLaMA-70B-Instruct hallucinated in 19.7% of answers (95% CI 18.6 to 20.7) even though 98.8% of prompt responses received maximal plausibility, and observed in experiment two, across models, lower hallucination rates aligned with higher usefulness scores ($\rho=-0.71$, $p=0.058$). Clinicians produced high agreement (quadratic weighted $\kappa=0.92$) and ($\tau_b=0.06$ to $0.18$, $\kappa=0.57$ to $0.61$) for experiments 1 and 2 respectively. Our findings indicate that, across all scales and architectures tested, current large language models remain unfit for unsupervised clinical deployment, and that human expert oversight is both necessary and the dominant cost driver.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their ability to accurately assess their own confidence remains poorly understood. We present an empirical study investigating whether LLMs exhibit patterns reminiscent of the Dunning-Kruger effect – a cognitive bias where individuals with limited competence tend to overestimate their abilities. We evaluate four state-of-the-art models (Claude Haiku 4.5, Gemini 2.5 Pro, Gemini 2.5 Flash, and Kimi K2) across four benchmark datasets totaling 24,000 experimental trials. Our results reveal striking calibration differences: Kimi K2 exhibits severe overconfidence with an Expected Calibration Error (ECE) of 0.726 despite only 23.3% accuracy, while Claude Haiku 4.5 achieves the best calibration (ECE = 0.122) with 75.4% accuracy. These findings demonstrate that poorly performing models display markedly higher overconfidence – a pattern analogous to the Dunning-Kruger effect in human cognition. We discuss implications for safe deployment of LLMs in high-stakes applications.

The digital age has expanded social media and online forums, allowing free expression for nearly 45% of the global population. Yet, it has also fueled online harassment, bullying, and harmful behaviors like hate speech and toxic comments across social networks, messaging apps, and gaming communities. Studies show 65% of parents notice hostile online behavior, and one-third of adolescents in mobile games experience bullying. A substantial volume of abusive content is generated and shared daily, not only on the surface web but also within dark web forums. Creators of abusive comments often employ specific words or coded phrases to evade detection and conceal their intentions. To address these challenges, we propose a hybrid deep learning model that integrates BERT, CNN, and LSTM architectures with a ReLU activation function to detect abusive language across multiple online platforms, including YouTube comments, online forum discussions, and dark web posts. The model demonstrates strong performance on a diverse and imbalanced dataset containing 77,620 abusive and 272,214 non-abusive text samples (ratio 1:3.5), achieving approximately 99% across evaluation metrics such as Precision, Recall, Accuracy, F1-score, and AUC. This approach effectively captures semantic, contextual, and sequential patterns in text, enabling robust detection of abusive content even in highly skewed datasets, as encountered in real-world scenarios.

Mixture-of-Experts (MoE) models enable scalable performance but face severe memory constraints on edge devices. Existing offloading strategies struggle with I/O bottlenecks due to the dynamic, low-information nature of autoregressive expert activation. In this paper, we propose to repurpose Speculative Decoding (SD) not merely as a compute accelerator, but as an informative lookahead sensor for memory management, supported by our theoretical and empirical analyses. Hence, we introduce MoE-SpAc, an MoE inference framework that integrates a Speculative Utility Estimator to track expert demand, a Heterogeneous Workload Balancer to dynamically partition computation via online integer optimization, and an Asynchronous Execution Engine to unify the prefetching and eviction in the same utility space. Extensive experiments on seven benchmarks demonstrate that MoE-SpAc achieves a 42% improvement in TPS over the SOTA SD-based baseline, and an average 4.04x speedup over all standard baselines. Code is available at https://github.com/lshAlgorithm/MoE-SpAc .

Large Language Models (LLMs) are now good enough at coding that developers can describe intent in plain language and let the tool produce the first code draft, a workflow increasingly built into tools like GitHub Copilot, Cursor, and Replit. What is missing is a reliable way to tell which model written queries are safe to accept without sending everything to a human. We study the application of an LLM jury to run this review step. We first benchmark 15 open models on 82 MySQL text to SQL tasks using an execution grounded protocol to get a clean baseline of which models are strong. From the six best models we build unanimous committees of sizes 1 through 6 that see the prompt, schema, and candidate SQL and accept it only when every member says it is correct. This rule matches safety first deployments where false accepts are more costly than false rejects. We measure true positive rate, false positive rate and Youden J and we also look at committees per generator. Our results show that single model judges are uneven, that small unanimous committees of strong models can cut false accepts while still passing many good queries, and that the exact committee composition matters significantly.

Small datasets like MNIST have historically been instrumental in advancing machine learning research by providing a controlled environment for rapid experimentation and model evaluation. However, their simplicity often limits their utility for distinguishing between advanced neural network architectures. To address these challenges, Greydanus et al. introduced the MNIST-1D dataset, a one-dimensional adaptation of MNIST designed to explore inductive biases in sequential data. This dataset maintains the advantages of small-scale datasets while introducing variability and complexity that make it ideal for studying advanced architectures. In this paper, we extend the exploration of MNIST-1D by evaluating the performance of Residual Networks (ResNet), Temporal Convolutional Networks (TCN), and Dilated Convolutional Neural Networks (DCNN). These models, known for their ability to capture sequential patterns and hierarchical features, were implemented and benchmarked alongside previously tested architectures such as logistic regression, MLPs, CNNs, and GRUs. Our experimental results demonstrate that advanced architectures like TCN and DCNN consistently outperform simpler models, achieving near-human performance on MNIST-1D. ResNet also shows significant improvements, highlighting the importance of leveraging inductive biases and hierarchical feature extraction in small structured datasets. Through this study, we validate the utility of MNIST-1D as a robust benchmark for evaluating machine learning architectures under computational constraints. Our findings emphasize the role of architectural innovations in improving model performance and offer insights into optimizing deep learning models for resource-limited environments.

Being modeled as a single-label classification task for a long time, recent work has argued that Arabic Dialect Identification (ADI) should be framed as a multi-label classification task. However, ADI remains constrained by the availability of single-label datasets, with no large-scale multi-label resources available for training. By analyzing models trained on single-label ADI data, we show that the main difficulty in repurposing such datasets for Multi-Label Arabic Dialect Identification (MLADI) lies in the selection of negative samples, as many sentences treated as negative could be acceptable in multiple dialects. To address these issues, we construct a multi-label dataset by generating automatic multi-label annotations using GPT-4o and binary dialect acceptability classifiers, with aggregation guided by the Arabic Level of Dialectness (ALDi). Afterward, we train a BERT-based multi-label classifier using curriculum learning strategies aligned with dialectal complexity and label cardinality. On the MLADI leaderboard, our best-performing LAHJATBERT model achieves a macro F1 of 0.69, compared to 0.55 for the strongest previously reported system. Code and data are available at https://mohamedalaa9.github.io/lahjatbert/.

Language models (LMs) are known to be prone to response biases, which present as option preference biases in fixed-response questions. It is therefore imperative to develop low-cost and effective response bias correction methods to improve LM performance and enable more accurate evaluations of model abilities. Here, we propose a simple response bias correction strategy ($\texttt{RBCorr}$) and test it on 12 open-weight language models using yes-no, entailment, and multiple choice questions. We show that response bias is prevalent in LMs pre-correction and that $\texttt{RBCorr}$ effectively eliminates bias and boosts model performance. We also explore the generalizability of bias behavior across models, datasets, and prompt formats, showing that LogProbs-based correction is highly dependent on all three of these aspects. Overall, $\texttt{RBCorr}$ is an easy-to-use method that can boost the performance of smaller LMs and ensure that LM performance on closed-response benchmarks aligns more closely with their true capabilities.