Papers

Retrieving the correct set of files from a large codebase is a crucial step in Automated Program Repair (APR). High recall is necessary to ensure that the relevant files are included, but simply increasing the number of retrieved files introduces noise and degrades efficiency. To address this tradeoff, we propose PatchRecall, a hybrid retrieval approach that balances recall with conciseness. Our method combines two complementary strategies: (1) codebase retrieval, where the current issue description is matched against the codebase to surface potentially relevant files, and (2) history-based retrieval, where similar past issues are leveraged to identify edited files as candidate targets. Candidate files from both strategies are merged and reranked to produce the final retrieval set. Experiments on SWE-Bench demonstrate that PatchRecall achieves higher recall without significantly increasing retrieved file count, enabling more effective APR.

Legal consultation question answering (Legal CQA) presents unique challenges compared to traditional legal QA tasks, including the scarcity of high-quality training data, complex task composition, and strong contextual dependencies. To address these, we construct JurisCQAD, a large-scale dataset of over 43,000 real-world Chinese legal queries annotated with expert-validated positive and negative responses, and design a structured task decomposition that converts each query into a legal element graph integrating entities, events, intents, and legal issues. We further propose JurisMA, a modular multi-agent framework supporting dynamic routing, statutory grounding, and stylistic optimization. Combined with the element graph, the framework enables strong context-aware reasoning, effectively capturing dependencies across legal facts, norms, and procedural logic. Trained on JurisCQAD and evaluated on a refined LawBench, our system significantly outperforms both general-purpose and legal-domain LLMs across multiple lexical and semantic metrics, demonstrating the benefits of interpretable decomposition and modular collaboration in Legal CQA.

The exponential growth of user-generated movie reviews on digital platforms has made accurate text sentiment classification a cornerstone task in natural language processing. Traditional models, including standard BERT and recurrent architectures, frequently struggle to capture long-distance semantic dependencies and resolve ambiguous emotional expressions in lengthy review texts. This paper proposes a novel hybrid framework that seamlessly integrates dynamic adaptive multi-head attention with supervised contrastive learning into a BERT-based Transformer encoder. The dynamic adaptive attention module employs a global context pooling vector to dynamically regulate the contribution of each attention head, thereby focusing on critical sentiment-bearing tokens while suppressing noise. Simultaneously, the supervised contrastive learning branch enforces tighter intra-class compactness and larger inter-class separation in the embedding space. Extensive experiments on the IMDB dataset demonstrate that the proposed model achieves competitive performance with an accuracy of 94.67%, outperforming strong baselines by 1.5–2.5 percentage points. The framework is lightweight, efficient, and readily extensible to other text classification tasks.

Recent advances in large language models (LLMs) have enabled promising progress in diagnosis prediction from electronic health records (EHRs). However, existing LLM-based approaches tend to overfit to historically observed diagnoses, often overlooking novel yet clinically important conditions that are critical for early intervention. To address this, we propose EviCare, an in-context reasoning framework that integrates deep model guidance into LLM-based diagnosis prediction. Rather than prompting LLMs directly with raw EHR inputs, EviCare performs (1) deep model inference for candidate selection, (2) evidential prioritization for set-based EHRs, and (3) relational evidence construction for novel diagnosis prediction. These signals are then composed into an adaptive in-context prompt to guide LLM reasoning in an accurate and interpretable manner. Extensive experiments on two real-world EHR benchmarks (MIMIC-III and MIMIC-IV) demonstrate that EviCare achieves significant performance gains, which consistently outperforms both LLM-only and deep model-only baselines by an average of 20.65% across precision and accuracy metrics. The improvements are particularly notable in challenging novel diagnosis prediction, yielding average improvements of 30.97%.

Olfaction lies at the intersection of chemical structure, neural encoding, and linguistic perception, yet existing representation methods fail to fully capture this pathway. Current approaches typically model only isolated segments of the olfactory pathway, overlooking the complete chain from molecule to receptors to linguistic descriptions. Such fragmentation yields learned embeddings that lack both biological grounding and semantic interpretability. We propose NOSE (Neural Olfactory-Semantic Embedding), a representation learning framework that aligns three modalities along the olfactory pathway: molecular structure, receptor sequence, and natural language description. Rather than simply fusing these signals, we decouple their contributions via orthogonal constraints, preserving the unique encoded information of each modality. To address the sparsity of olfactory language, we introduce a weak positive sample strategy to calibrate semantic similarity, preventing erroneous repulsion of similar odors in the feature space. Extensive experiments demonstrate that NOSE achieves state-of-the-art (SOTA) performance and excellent zero-shot generalization, confirming the strong alignment between its representation space and human olfactory intuition.Code and data are available at https://github.com/Xianyusyy/NOSE

Instruction tuning relies on large instruction-response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.

Large Language Models (LLMs) struggle with knowledge-intensive tasks due to hallucinations and fragmented reasoning over dispersed information. While Retrieval-Augmented Generation (RAG) grounds generation in external sources, existing methods often treat evidence as isolated units, failing to reconstruct the logical chains that connect these dots. Inspired by Complementary Learning Systems (CLS), we propose CodaRAG, a framework that evolves retrieval from passive lookup into active associative discovery. CodaRAG operates via a three-stage pipeline: (1) Knowledge Consolidation to unify fragmented extractions into a stable memory substrate; (2) Associative Navigation to traverse the graph via multi-dimensional pathways-semantic, contextualized, and functional-explicitly recovering dispersed evidence chains; and (3) Interference Elimination to prune hyper-associative noise, ensuring a coherent, high-precision reasoning context. On GraphRAG-Bench, CodaRAG achieves absolute gains of 7-10% in retrieval recall and 3-11% in generation accuracy. These results demonstrate CodaRAG's superior ability to systematically robustify associative evidence retrieval for factual, reasoning, and creative tasks.

Alan Turing is considered as a founder of current computer science together with Kurt Godel, Alonzo Church and John von Neumann. In this paper multiple new research results are presented. It is demonstrated that there would not be Alan Turing's achievements without earlier seminal contributions by Georg Cantor in the set theory and foundations of mathematics. It is proposed to introduce the measure of undecidability of problems unsolvable by Turing machines based on probability distribution of its input data, i.e., to provide the degree of unsolvabilty based on the number of undecidable instances of input data versus decidable ones. It is proposed as well to extend the Turing's work on infinite logics and Oracle machines to a whole class of super-Turing models of computation. Next, the three new complexity classes for TM undecidable problems have been defined: U-complete (Universal complete), D-complete (Diagonalization complete) and H-complete (Hypercomputation complete) classes. The above has never been defined explicitly before by other scientists, and has been inspired by Cook/Levin NP-complete class for intractable problems. Finally, an equivalent to famous P is not equal to NP unanswered question for NP-complete class, has been answered negatively for U-complete class of complexity for undecidable problems.

In recent years, aspect-based sentiment analysis (ABSA) has made rapid progress and shown strong practical value. However, existing research and benchmarks are largely concentrated on high-resource languages, leaving fine-grained sentiment extraction in low-resource languages under-explored. To address this gap, we constructed the first Low-resource languages Aspect-based Sentiment Quadruple dataset, named LASQ, which includes two low-resource languages: Uzbek and Uyghur. Secondly, it includes a fine-grained target-aspect-opinion-sentiment quadruple extraction task. To facilitate future research, we designed a grid-tagging model that integrates syntactic knowledge. This model incorporates part-of-speech (POS) and dependency knowledge into the model through our designed Syntax Knowledge Embedding Module (SKEM), thereby alleviating the lexical sparsity problem caused by agglutinative languages. Experiments on LASQ demonstrate consistent gains over competitive baselines, validating both the dataset's utility and the effectiveness of the proposed modeling approach.

Inferring nationality from personal names is a critical capability for equity and bias monitoring, personalization, and a valuable tool in biomedical and sociological research. However, existing name-based nationality classifiers are typically trained on relatively small or source-specific labeled datasets, which can introduce coverage gaps and limit performance for underrepresented countries. While large language models (LLMs) demonstrate strong zero-shot performance for name-based nationality prediction, their computational cost and latency make them impractical for real-time, large-scale deployment. In this work, we created a large-scale name-nationality dataset from the Open Academic Graph (OAG) and introduce a framework that leverages LLMs as dataset enrichers rather than inference engines. We augment low-resource countries with LLM-generated names and evaluate on real and synthetic-tail test sets. We find that augmentation produces large gains when evaluation includes synthetic tail names and still offers a modest lift on tail-country metrics otherwise. Overall, NameBERT models achieve significantly higher accuracy than state-of-the-art baselines across both in- and out-of-domain tasks, while remaining efficient for large-scale inference compared to LLMs.

Terminology substitution errors in clinical notes, where one medical term is replaced by a linguistically valid but clinically different term, pose a persistent challenge for automated error detection in healthcare. We introduce BLUEmed, a multi-agent debate framework augmented with hybrid Retrieval-Augmented Generation (RAG) that combines evidence-grounded reasoning with multi-perspective verification for clinical error detection. BLUEmed decomposes each clinical note into focused sub-queries, retrieves source-partitioned evidence through dense, sparse, and online retrieval, and assigns two domain expert agents distinct knowledge bases to produce independent analyses; when the experts disagree, a structured counter-argumentation round and cross-source adjudication resolve the conflict, followed by a cascading safety layer that filters common false-positive patterns. We evaluate BLUEmed on a clinical terminology substitution detection benchmark under both zero-shot and few-shot prompting with multiple backbone models spanning proprietary and open-source families. Experimental results show that BLUEmed achieves the best accuracy (69.13%), ROC-AUC (74.45%), and PR-AUC (72.44%) under few-shot prompting, outperforming both single-agent RAG and debate-only baselines. Further analyses across six backbone models and two prompting strategies confirm that retrieval augmentation and structured debate are complementary, and that the framework benefits most from models with sufficient instruction-following and clinical language understanding.

Diffusion models promise efficient parallel text generation but rely on bidirectional attention, creating a structural mismatch with pre-trained Autoregressive (AR) models. This incompatibility precludes reusing robust AR priors, necessitating prohibitive pre-training from scratch. To bridge this gap, we propose FLUID, a framework that efficiently adapts AR backbones to the diffusion paradigm. By enforcing Strictly Causal Alignment, FLUID enables seamless initialization from standard GPT-style checkpoints, circumventing the need for massive pre-training. Furthermore, we introduce Elastic Horizons, an entropy-driven mechanism that dynamically modulates denoising strides based on local information density rather than fixed schedules. Experiments demonstrate that FLUID achieves state-of-the-art performance while reducing training costs by orders of magnitude, effectively reconciling established AR foundations with efficient parallel generation. Our code is available at https://github.com/Oli-lab-nun/FLUID/tree/main.

Reinforcement Learning with Verifiable Rewards (RLVR) is an essential paradigm that enhances the reasoning capabilities of Large Language Models (LLMs). However, existing methods typically rely on static policy optimization schemes that misalign with the model's evolving reasoning capabilities. To address this issue, we propose Adaptive Power-Mean Policy Optimization (APMPO), which comprises two main innovations: Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC). Specifically, PMPO introduces a generalized power-mean objective. This enables the model to adaptively transition from the signal-amplifying behavior of the arithmetic mean to the consistency-enforcing behavior of the geometric mean. FAC adaptively adjusts clipping bounds based on real-time reward statistics to overcome the limitations of static mechanisms. Capitalizing on these innovations, APMPO improves learning dynamics and reasoning performance. Extensive experiments on nine datasets across three reasoning tasks showcase the superiority of APMPO over state-of-the-art RLVR-based baselines. For instance, APMPO boosts the average Pass@1 score on mathematical reasoning benchmarks by 3.0 points compared to GRPO when using Qwen2.5-3B-Instruct.

Unsupervised reinforcement learning (RL) has emerged as a promising paradigm for enabling self-improvement in large language models (LLMs). However, existing unsupervised RL-based methods often lack the capacity to adapt to the model's evolving reasoning capabilities during training. Therefore, these methods can misdirect policy optimization in the absence of ground-truth supervision. To address this issue, we introduce FREIA, a novel RL-based algorithm built on two key innovations: (1) Free Energy-Driven Reward (FER) adapts rewards to balance consensus and exploration based on the Free Energy Principle. (2) Adaptive Advantage Shaping (AAS) adaptively adjusts learning signals based on the statistical characteristics of sampled rewards. Empirical evaluations on nine datasets across three reasoning tasks showcase that FREIA outperforms other unsupervised RL-based baselines. Notably, in mathematical reasoning tasks, FREIA surpasses other methods by an average of 0.5 to 3.5 points in Pass@1 using the DeepSeek-R1-Distill-Qwen-1.5B model.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in Traffic Accident Detection (TAD) and Traffic Accident Understanding (TAU). However, existing studies mainly focus on describing and interpreting accident videos, leaving room for deeper causal reasoning and integration of legal knowledge. Traffic Accident Responsibility Allocation (TARA) is a more challenging task that requires multi-step reasoning grounded in traffic regulations. To address this, we introduce AITP (Artificial Intelligence Traffic Police), a multimodal large language model for responsibility reasoning and allocation. AITP enhances reasoning via a Multimodal Chain-of-Thought (MCoT) mechanism and integrates legal knowledge through Retrieval-Augmented Generation (RAG). We further present DecaTARA, a decathlon-style benchmark unifying ten interrelated traffic accident reasoning tasks with 67,941 annotated videos and 195,821 question-answer pairs. Extensive experiments show that AITP achieves state-of-the-art performance across responsibility allocation, TAD, and TAU tasks, establishing a new paradigm for reasoning-driven multimodal traffic analysis.

Media narratives wield tremendous power in shaping public opinion, yet computational approaches struggle to capture the nuanced storytelling structures that communication theory emphasizes as central to how meaning is constructed. Existing approaches either miss subtle narrative patterns through coarse-grained analysis or require domain-specific taxonomies that limit scalability. To bridge this gap, we present a framework for inducing rich narrative schemas by jointly modeling events and characters via structured clustering. Our approach produces explainable narrative schemas that align with established framing theory while scaling to large corpora without exhaustive manual annotation.

Large language models (LLMs) demonstrate strong performance in math reasoning benchmarks, but their performance varies inconsistently across problems with varying levels of difficulty. This paper describes Adaptive Multi-Expert Reasoning (AMR), a framework that focuses on problem complexity by reasoning with dynamically adapted strategies. An agile routing system that focuses on problem text predicts problems' difficulty and uncertainty and guides a reconfigurable sampling mechanism to manage the breadth of generation. Three specialized experts create candidate responses, which are modified during multiple correction and finalization phases. A neural verifier assesses the correctness of responses, while a clustering-based aggregation technique identifies the final candidate answer based on a combination of consensus and answer quality. When evaluated on the GSM8K dataset, AMR achieved 75.28% accuracy while only using the original training data. This result outperformed the majority of comparable 7B models that were trained on synthetic data. This showcases that models using difficulty-based routing and uncertainty-driven aggregation are efficient and effective in improving math reasoning models' robustness.

Background: Large Language Models (LLMs) are transforming artificial intelligence applications in healthcare due to their ability to understand, generate, and summarize complex medical text. They offer valuable support to clinicians, researchers, and patients, yet their deployment in high-stakes clinical environments raises critical concerns regarding accuracy, reliability, and patient safety. Despite substantial attention in recent years, standardized benchmarking of LLMs for medical applications has been limited. Objective: This study addresses the need for a standardized comparative evaluation of LLMs in medical settings. Method: We evaluate multiple models, including ChatGPT, LLaMA, Grok, Gemini, and ChatDoctor, on core medical tasks such as patient note summarization and medical question answering, using the open-access datasets, MedMCQA, PubMedQA, and Asclepius, and assess performance through a combination of linguistic and task-specific metrics. Results: The results indicate that domain-specific models, such as ChatDoctor, excel in contextual reliability, producing medically accurate and semantically aligned text, whereas general-purpose models like Grok and LLaMA perform better in structured question-answering tasks, demonstrating higher quantitative accuracy. This highlights the complementary strengths of domain-specific and general-purpose LLMs depending on the medical task. Conclusion: Our findings suggest that LLMs can meaningfully support medical professionals and enhance clinical decision-making; however, their safe and effective deployment requires adherence to ethical standards, contextual accuracy, and human oversight in relevant cases. These results underscore the importance of task-specific evaluation and cautious integration of LLMs into healthcare workflows.

Large Vision-Language Models (LVLMs) rely on attention-based retrieval of safety instructions to maintain alignment during generation. Existing attacks typically optimize image perturbations to maximize harmful output likelihood, but suffer from slow convergence due to gradient conflict between adversarial objectives and the model's safety-retrieval mechanism. We propose Attention-Guided Visual Jailbreaking, which circumvents rather than overpowers safety alignment by directly manipulating attention patterns. Our method introduces two simple auxiliary objectives: (1) suppressing attention to alignment-relevant prefix tokens and (2) anchoring generation on adversarial image features. This simple yet effective push-pull formulation reduces gradient conflict by 45% and achieves 94.4% attack success rate on Qwen-VL (vs. 68.8% baseline) with 40% fewer iterations. At tighter perturbation budgets ($\epsilon=8/255$), we maintain 59.0% ASR compared to 45.7% for standard methods. Mechanistic analysis reveals a failure mode we term safety blindness: successful attacks suppress system-prompt attention by 80%, causing models to generate harmful content not by overriding safety rules, but by failing to retrieve them.

In what way could a data breach involving government-issued IDs such as passports, driver's licenses, etc., rival a random voluntary disclosure on a nondescript social-media platform? At first glance, the former appears more significant, and that is a valid assessment. The disclosed data could contain an individual's date of birth and address; for all intents and purposes, a leak of that data would be disastrous. Given the threat, the latter scenario involving an innocuous online post seems comparatively harmless – or does it? From that post and others like it, a forensic linguist could stylometrically uncover equivalent pieces of information, estimating an age range for the author (adolescent or adult) and narrowing down their geographical location (specific country). While not an exact science – the determinations are statistical – stylometry can reveal comparable, though noticeably diluted, information about an individual. To prevent an ID from being breached, simply sharing it as little as possible suffices. Preventing the leakage of personal information from written text requires a more complex solution: adversarial stylometry. In this paper, we explore how performing homoglyph substitution – the replacement of characters with visually similar alternatives (e.g., "h" $\texttt{[U+0068]}$ $\rightarrow$ "h" $\texttt{[U+04BB]}$) – on text can degrade stylometric systems.

Existing tool-use benchmarks for LLM agents are overwhelmingly linear: our analysis of six benchmarks shows 55 to 100% of instances are simple chains of 2 to 5 steps. We introduce The Amazing Agent Race (AAR), a benchmark featuring directed acyclic graph (DAG) puzzles (or "legs") with fork-merge tool chains. We release 1,400 instances across two variants: sequential (800 legs) and compositional (600 DAG legs). Agents must navigate Wikipedia, execute multi-step tool chains, and aggregate results into a verifiable answer. Legs are procedurally generated from Wikipedia seeds across four difficulty levels with live-API validation. Three complementary metrics (finish-line accuracy, pit-stop visit rate, and roadblock completion rate) separately diagnose navigation, tool-use, and arithmetic failures. Evaluating three agent frameworks on 1,400 legs, the best achieves only 37.2% accuracy. Navigation errors dominate (27 to 52% of trials) while tool-use errors remain below 17%, and agent architecture matters as much as model scale (Claude Code matches Codex CLI at 37% with 6x fewer tokens). The compositional structure of AAR reveals that agents fail not at calling tools but at navigating to the right pages, a blind spot invisible to linear benchmarks. The project page can be accessed at: https://minnesotanlp.github.io/the-amazing-agent-race

Agentic code tasks such as fault localization and patch generation require processing long codebases under tight memory constraints, where the Key-Value (KV) cache becomes the primary inference bottleneck. Existing compression methods rely exclusively on attention signals to estimate token importance, systematically discarding structurally critical tokens such as call sites, branch conditions, and assignments that are essential for code understanding. We present CodeComp, a training-free KV cache compression framework that incorporates static program analysis into LLM inference via Code Property Graph priors extracted by Joern. Across bug localization and code generation benchmarks, CodeComp consistently outperforms attention-only compression baselines under equal memory budgets, recovering the majority of full-context accuracy under aggressive KV cache compression, while matching the patch generation quality of uncompressed full-context inference and integrating seamlessly into SGLang-based agentic coding pipelines without model modification.

Language models can be used to identify relationships between financial entities in text. However, while structured output mechanisms exist, prompting-based pipelines still incur autoregressive decoding costs and decouple graph construction from downstream optimization. We propose Relational Probing, which replaces the standard language-model head with a relation head that induces a relational graph directly from language-model hidden states and is trained jointly with the downstream task model for stock-trend prediction. This approach both learns semantic representations and preserves the strict structure of the induced relational graph. It enables language-model outputs to go beyond text, allowing them to be reshaped into task-specific formats for downstream models. To enhance reproducibility, we provide an operational definition of small language models (SLMs): models that can be fine-tuned end-to-end on a single 24GB GPU under specified batch-size and sequence-length settings. Experiments use Qwen3 backbones (0.6B/1.7B/4B) as upstream SLMs and compare against a co-occurrence baseline. Relational Probing yields consistent performance improvements at competitive inference cost.

Large Language Models (LLMs) can generate factually inaccurate content even if they have corresponding knowledge, which critically undermines their reliability. Existing approaches attempt to mitigate this by incorporating uncertainty in QA prompt during training, but these numerical scores lack the semantic richness for LLM to properly understand its internal states of trustworthiness and honestness, leading to insufficient factuality alignment. We introduce FAITH (Factuality Alignment through Integrating Trustworthiness and Honestness), a post-training framework for factuality alignment that integrates natural-language uncertainty signals with external knowledge. Specifically, we augment training datasets by computing confidence scores and semantic entropy from LLM outputs and mapping them into a knowledge state quadrant that describes the model's internal knowledge possession (trustworthiness) and answering behaviors (honestness) in natural language. Based on this enhanced data, we design a reward function that considers both correctness and uncertainty signals, and fine-tune the LLM using the Proximal Policy Optimization (PPO) algorithm. To further mitigate weakly grounded responses, we design a retrieval-augmented module that retrieves relevant external passages, improving the consistency between internal and external knowledge representations. Extensive experiments on four knowledge-intensive benchmarks demonstrate that FAITH enhances the factual accuracy and truthfulness of LLMs.

Multi-vector models dominate Visual Document Retrieval (VDR) due to their fine-grained matching capabilities, but their high storage and computational costs present a major barrier to practical deployment. In this paper, we propose ColChunk, a plug-and-play framework that introduces multimodal late chunking to construct efficient, contextualized multi-vectors. Unlike existing pruning or fixed-token approaches, ColChunk employs hierarchical clustering on patch-level embeddings, fused with a 2D position prior to ensure spatial-semantic coherence. This adaptive grouping allows for a content-aware representation that preserves global context while drastically reducing the vector count. Evaluations across 24 VDR datasets demonstrate ColChunk achieves over a 90% reduction in storage requirements while simultaneously delivering a 9-point average improvement in nDCG@5 across representative single-vector models. ColChunk provides a practical solution for balancing retrieval accuracy and efficiency in visual document systems.

The advancement of large language models (LLMs) has enhanced tabular question answering (Tabular QA), yet they struggle with open-domain queries exhibiting underspecified or uncertain expressions. To address this, we introduce the ODUTQA-MDC task and the first comprehensive benchmark to tackle it. This benchmark includes: (1) a large-scale ODUTQA dataset with 209 tables and 25,105 QA pairs; (2) a fine-grained labeling scheme for detailed evaluation; and (3) a dynamic clarification interface that simulates user feedback for interactive assessment. We also propose MAIC-TQA, a multi-agent framework that excels at detecting ambiguities, clarifying them through dialogue, and refining answers. Experiments validate our benchmark and framework, establishing them as a key resource for advancing conversational, underspecification-aware Tabular QA research.

Large language models are increasingly used as writing tools and pedagogical resources in English for Academic Purposes, but it remains unclear whether they encode culturally differentiated representations when generating academic text. This study tests whether Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating research article introductions conditioned by British and Chinese academic personas. A corpus of 270 texts was generated from 45 prompt templates crossed with six persona conditions in a 2 x 3 design. Logistic regression probes were trained on hidden-state activations across all 35 layers, with shuffled-label baselines, a surface-text skyline classifier, cross-family tests, and sentence-level baselines used as controls. Probe-selected token positions were annotated for structural, lexical, and stance features using the Stanza NLP pipeline. The nationality probe reached 0.968 cross-validated accuracy at Layer 18, with perfect held-out classification. Nationality encoding followed a non-monotonic trajectory across layers, with structural effects strongest in the middle to upper network and lexical-domain effects peaking earlier. At high-signal token positions, British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative or process-oriented vocabulary, while Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural or internationalisation vocabulary. However, sentence-level analysis found no significant nationality differences in the full generated surface text. The findings extend probing methodology to a sociolinguistic attribute and have practical implications for EAP and language pedagogy.

Generative listwise reranking leverages global context for superior retrieval but is plagued by intrinsic position bias, where models exhibit structural sensitivity to input order independent of relevance. Existing mitigations present a dilemma: inference-time aggregation incurs prohibitive latency, while training-based methods often fail to eradicate ingrained priors, particularly in compact models. To resolve this dilemma, we propose CapCal (Content-Agnostic Probability Calibration), a training-free framework that mechanically decouples positional bias from ranking decisions. By estimating the bias distribution via content-free placeholders, CapCal rectifies output logits through an entropy-adaptive contrastive mechanism. Evaluations across 10 benchmarks confirm that CapCal achieves superior performance among training-free methods while preserving single-pass efficiency. Notably, it unlocks the latent potential of lightweight models (e.g., 0.6B), delivering absolute NDCG gains exceeding 10 points and outperforming both permutation-based aggregation and data-augmentation baselines.

Researchers have explored different ways to improve large language models (LLMs)' capabilities via dummy token insertion in contexts. However, existing works focus solely on the dummy tokens themselves, but fail to leverage the inherent sentence-level structure of natural language. This is a critical oversight, as LLMs acquire linguistic capabilities through exposure to human-generated texts, which are inherently structured at the sentence level. Motivated by this gap, we propose an approach that inserts delimiters at sentence boundaries in LLM inputs, which not only integrates dummy tokens into the context, but also facilitates LLMs with sentence-by-sentence processing behavior during reasoning. Two concrete methods: (1). In-context learning and (2). Supervised fine-tuning are experimented using 7B models to 600B Deepseek-V3. Our results demonstrate consistent improvements across various tasks, with notable gains of up to 7.7% on GSM8k and 12.5% on DROP. Furthermore, the fine-tuned LLMs can incorporate sentence awareness evidenced by their internal representations. Our work establishes a simple yet effective technique for enhancing LLM's capabilities, offering promising directions for cognitive-inspired LLM enhancement paradigm.

Dysarthric speech severity assessment typically requires trained clinicians or supervised models built from labelled pathological speech, limiting scalability across languages and clinical settings. We present a training-free method that quantifies dysarthria severity by measuring degradation in phonological feature subspaces within frozen HuBERT representations. No supervised severity model is trained; feature directions are estimated from healthy control speech using a pretrained forced aligner. For each speaker, we extract phone-level embeddings via Montreal Forced Aligner, compute d-prime scores along phonological contrast directions (nasality, voicing, stridency, sonorance, manner, and four vowel features) derived exclusively from healthy controls, and construct a 12-dimensional phonological profile.Evaluating 890 speakers across 10 corpora, 5 languages (English, Spanish, Dutch, Mandarin, French), and 3 primary aetiologies (Parkinson's disease, cerebral palsy, ALS), we find that all five consonant d-prime features correlate significantly with clinical severity (random-effects meta-analysis rho = -0.50 to -0.56, p < 2e-4; pooled Spearman rho = -0.47 to -0.55 with bootstrap 95% CIs not crossing zero). The effect replicates within individual corpora, survives FDR correction, and remains robust to leave-one-corpus-out removal and alignment quality controls. Nasality d-prime decreases monotonically from control to severe in 6 of 7 severity-graded corpora. Mann-Whitney U tests confirm that all 12 features distinguish controls from severely dysarthric speakers (p < 0.001).The method requires no dysarthric training data and applies to any language with an existing MFA acoustic model (currently 29 languages). We release the full pipeline and phone feature configurations for six languages.