Papers

Large reasoning models have demonstrated remarkable performance on complex reasoning tasks, yet the excessive length of their chain-of-thought outputs remains a major practical bottleneck due to high computation cost and poor deployability. Existing compression methods have achieved partial success but overlook a crucial phenomenon in the training process – the entropy conflict. During compression training, entropy decreases, leading to shorter reasoning but limited exploration, while accuracy-oriented objectives increase entropy, lengthening reasoning chains. This can cause the model to get stuck in a local dilemma. Our analysis further reveals the origin of the entropy conflict: many high-entropy tokens are logical connectors that receive larger gradients and are encouraged under the performance objective, while the compression objective simultaneously penalizes these potentially redundant connectors. This opposing pressure creates a direct source of entropy conflict. To address these issues, we adopt an entropy-guided training framework. As entropy descends, the model is guided toward efficient reasoning by encouraging concise thought steps; as entropy rises, exploration is reinforced under the compact reasoning mode to improve robustness. Experiments on six mathematical benchmarks show that our method compresses reasoning length to 20% of the original while maintaining or even surpassing baseline accuracy. Code and models will be released publicly.

Recent advances in speech-enabled AI, including Google's NotebookLM and OpenAI's speech-to-speech API, are driving widespread interest in voice interfaces globally. Despite this momentum, there exists no publicly available application-specific model evaluation that caters to Africa's linguistic diversity. We present AfriSpeech-MultiBench, the first domain-specific evaluation suite for over 100 African English accents across 10+ countries and seven application domains: Finance, Legal, Medical, General dialogue, Call Center, Named Entities and Hallucination Robustness. We benchmark a diverse range of open, closed, unimodal ASR and multimodal LLM-based speech recognition systems using both spontaneous and non-spontaneous speech conversation drawn from various open African accented English speech datasets. Our empirical analysis reveals systematic variation: open-source ASR models excels in spontaneous speech contexts but degrades on noisy, non-native dialogue; multimodal LLMs are more accent-robust yet struggle with domain-specific named entities; proprietary models deliver high accuracy on clean speech but vary significantly by country and domain. Models fine-tuned on African English achieve competitive accuracy with lower latency, a practical advantage for deployment, hallucinations still remain a big problem for most SOTA models. By releasing this comprehensive benchmark, we empower practitioners and researchers to select voice technologies suited to African use-cases, fostering inclusive voice applications for underserved communities.

The automatic movie dubbing model generates vivid speech from given scripts, replicating a speaker's timbre from a brief timbre prompt while ensuring lip-sync with the silent video. Existing approaches simulate a simplified workflow where actors dub directly without preparation, overlooking the critical director-actor interaction. In contrast, authentic workflows involve a dynamic collaboration: directors actively engage with actors, guiding them to internalize the context cues, specifically emotion, before performance. To address this issue, we propose a new Retrieve-Augmented Director-Actor Interaction Learning scheme to achieve authentic movie dubbing, termed Authentic-Dubber, which contains three novel mechanisms: (1) We construct a multimodal Reference Footage library to simulate the learning footage provided by directors. Note that we integrate Large Language Models (LLMs) to achieve deep comprehension of emotional representations across multimodal signals. (2) To emulate how actors efficiently and comprehensively internalize director-provided footage during dubbing, we propose an Emotion-Similarity-based Retrieval-Augmentation strategy. This strategy retrieves the most relevant multimodal information that aligns with the target silent video. (3) We develop a Progressive Graph-based speech generation approach that incrementally incorporates the retrieved multimodal emotional knowledge, thereby simulating the actor's final dubbing process. The above mechanisms enable the Authentic-Dubber to faithfully replicate the authentic dubbing workflow, achieving comprehensive improvements in emotional expressiveness. Both subjective and objective evaluations on the V2C Animation benchmark dataset validate the effectiveness. The code and demos are available at https://github.com/AI-S2-Lab/Authentic-Dubber.

Large language models perform strongly on general tasks but remain constrained in specialized settings such as music, particularly in the music-entertainment domain, where corpus scale, purity, and the match between data and training objectives are critical. We address this by constructing a large, music-related natural language corpus (40B tokens) that combines open source and in-house data, and by implementing a domain-first data pipeline: a lightweight classifier filters and weights in-domain text, followed by multi-stage cleaning, de-duplication, and privacy-preserving masking. We further integrate multi-source music text with associated metadata to form a broader, better-structured foundation of domain knowledge. On the training side, we introduce reference-model (RM)-based token-level soft scoring for quality control: a unified loss-ratio criterion is used both for data selection and for dynamic down-weighting during optimization, reducing noise gradients and amplifying task-aligned signals, thereby enabling more effective music-domain continued pretraining and alignment. To assess factuality, we design the MusicSimpleQA benchmark, which adopts short, single-answer prompts with automated agreement scoring. Beyond the benchmark design, we conduct systematic comparisons along the axes of data composition. Overall, this work advances both the right corpus and the right objective, offering a scalable data-training framework and a reusable evaluation tool for building domain LLMs in the music field.

Grammatical Error Correction (GEC) is an important aspect of natural language processing. Arabic has a complicated morphological and syntactic structure, posing a greater challenge than other languages. Even though modern neural models have improved greatly in recent years, the majority of previous attempts used individual models without taking into account the potential benefits of combining different systems. In this paper, we present one of the first multi-system approaches for correcting grammatical errors in Arabic, the Arab Enhanced Edit Selection System Complication (ArbESC+). Several models are used to collect correction proposals, which are represented as numerical features in the framework. A classifier determines and implements the appropriate corrections based on these features. In order to improve output quality, the framework uses support techniques to filter overlapping corrections and estimate decision reliability. A combination of AraT5, ByT5, mT5, AraBART, AraBART+Morph+GEC, and Text editing systems gave better results than a single model alone, with F0.5 at 82.63% on QALB-14 test data, 84.64% on QALB-15 L1 data, and 65.55% on QALB-15 L2 data. As one of the most significant contributions of this work, it's the first Arab attempt to integrate linguistic error correction. Improving existing models provides a practical step towards developing advanced tools that will benefit users and researchers of Arabic text processing.

Entity Linking (EL), the task of mapping textual entity mentions to their corresponding entries in knowledge bases, constitutes a fundamental component of natural language understanding. Recent advancements in Large Language Models (LLMs) have demonstrated remarkable potential for enhancing EL performance. Prior research has leveraged LLMs to improve entity disambiguation and input representation, yielding significant gains in accuracy and robustness. However, these approaches typically apply LLMs to isolated stages of the EL task, failing to fully integrate their capabilities throughout the entire process. In this work, we introduce DeepEL, a comprehensive framework that incorporates LLMs into every stage of the entity linking task. Furthermore, we identify that disambiguating entities in isolation is insufficient for optimal performance. To address this limitation, we propose a novel self-validation mechanism that utilizes global contextual information, enabling LLMs to rectify their own predictions and better recognize cohesive relationships among entities within the same sentence. Extensive empirical evaluation across ten benchmark datasets demonstrates that DeepEL substantially outperforms existing state-of-the-art methods, achieving an average improvement of 2.6% in overall F1 score and a remarkable 4% gain on out-of-domain datasets. These results underscore the efficacy of deep LLM integration in advancing the state-of-the-art in entity linking.

LLMs still struggle with hallucination, especially when confronted with symbolic triggers like modifiers, negation, numbers, exceptions, and named entities. Yet, we lack a clear understanding of where these symbolic hallucinations originate, making it crucial to systematically handle such triggers and localize the emergence of hallucination inside the model. While prior work explored localization using statistical techniques like LSC and activation variance analysis, these methods treat all tokens equally and overlook the role symbolic linguistic knowledge plays in triggering hallucinations. So far, no approach has investigated how symbolic elements specifically drive hallucination failures across model layers, nor has symbolic linguistic knowledge been used as the foundation for a localization framework. We propose the first symbolic localization framework that leverages symbolic linguistic and semantic knowledge to meaningfully trace the development of hallucinations across all model layers. By focusing on how models process symbolic triggers, we analyze five models using HaluEval and TruthfulQA. Our symbolic knowledge approach reveals that attention variance for these linguistic elements explodes to critical instability in early layers (2-4), with negation triggering catastrophic variance levels, demonstrating that symbolic semantic processing breaks down from the very beginning. Through the lens of symbolic linguistic knowledge, despite larger model sizes, hallucination rates remain consistently high (78.3%-83.7% across Gemma variants), with steep attention drops for symbolic semantic triggers throughout deeper layers. Our findings demonstrate that hallucination is fundamentally a symbolic linguistic processing failure, not a general generation problem, revealing that symbolic semantic knowledge provides the key to understanding and localizing hallucination mechanisms in LLMs.

Future superhuman models will surpass the ability of humans and humans will only be able to weakly supervise superhuman models. To alleviate the issue of lacking high-quality data for model alignment, some works on weak-to-strong generalization (W2SG) finetune a strong pretrained model with a weak supervisor so that it can generalize beyond weak supervision. However, the invariable use of weak supervision in existing methods exposes issues in robustness, with a proportion of weak labels proving harmful to models. In this paper, we propose a selective W2SG framework to avoid using weak supervision when unnecessary. We train a binary classifier P(IK) to identify questions that a strong model can answer and use its self-generated labels for alignment. We further refine weak labels with a graph smoothing method. Extensive experiments on three benchmarks show that our method consistently outperforms competitive baselines. Further analyses show that P(IK) can generalize across tasks and difficulties, which indicates selective W2SG can help superalignment.

In this research, we combine Transformer-based relation extraction with matching of knowledge graphs (KGs) and apply them to answering multiple-choice questions (MCQs) while maintaining the traceability of the output process. KGs are structured representations of factual knowledge consisting of entities and relations. Due to the high construction cost, they had been regarded as static databases with validated links. However, the recent development of Transformer-based relation extraction (RE) methods has enabled us to generate KGs dynamically by giving them natural language texts, and thereby opened the possibility for representing the meaning of the input sentences with the created KGs. Using this effect, we propose a method that answers MCQs in the "fill-in-the-blank" format, taking care of the point that RE methods generate KGs that represent false information if provided with factually incorrect texts. We measure the truthfulness of each question sentence by (i) converting the sentence into a relational graph using an RE method and (ii) verifying it against factually correct KGs under the closed-world assumption. The experimental results demonstrate that our method correctly answers up to around 70% of the questions, while providing traceability of the procedure. We also highlight that the question category has a vast influence on the accuracy.

Aspect-Based Sentiment Analysis (ABSA) predicts sentiment polarity for specific aspect terms, a task made difficult by conflicting sentiments across aspects and the sparse context of short texts. Prior graph-based approaches model only pairwise dependencies, forcing them to construct multiple graphs for different relational views. These introduce redundancy, parameter overhead, and error propagation during fusion, limiting robustness in short-text, low-resource settings. We present HyperABSA, a dynamic hypergraph framework that induces aspect-opinion structures through sample-specific hierarchical clustering. To construct these hyperedges, we introduce a novel acceleration-fallback cutoff for hierarchical clustering, which adaptively determines the level of granularity. Experiments on three benchmarks (Lap14, Rest14, MAMS) show consistent improvements over strong graph baselines, with substantial gains when paired with RoBERTa backbones. These results position dynamic hypergraph construction as an efficient, powerful alternative for ABSA, with potential extensions to other short-text NLP tasks.

Deploying LLMs for financial information retrieval demands accuracy, cost efficiency, and reproducibility, yet most approaches require expensive fine-tuning. PRISM is a training-free framework that composes three deployable modules (prompt engineering, in-context learning, and a candidate refinement module) with a boundary-condition study of multi-agent coordination for financial document retrieval. On FiQA-2018, stacked reranking and $k$-reciprocal encoding raise Recall@100 from 0.7783 to 0.8326. On FinAgentBench, a Cohere-refined top-50 pool yields a higher observed end-to-end NDCG@5, below this evaluation's minimum detectable effect, while cutting per-query cost by about 35%. Added complexity does not always pay off: $k$-reciprocal encoding stops helping once the upstream stage is already high-recall, and multi-agent coordination is not a default for fine-grained chunk ranking. Matched pairs that hold prompt, model, and exemplars fixed lose 0.0702 and 0.0681 combined NDCG@5 to their single-path controls, and the whole gap sits at the chunk stage. We release an expert-annotated FinAgentBench evaluation set, together with a label-robustness sweep over 16 alternative ground-truth constructions, restoring a reproducible evaluation path after the original evaluator went offline. Without fine-tuning, PRISM placed third on the official FinAgentBench private leaderboard (leaderboard basis, NDCG@5 0.71181), reaches NDCG@10 0.6197 on FiQA-2018, and 98% accuracy on FinanceBench. We report per-query latency, token, and dollar accounting across roughly 40 configurations, present the quality-cost-latency Pareto frontier, and introduce FACET, a transparent SLA-specific scalarization for choosing among non-dominated configurations. Code and dataset are released at https://bit.ly/prism-ailens.

Automatic ICD coding from clinical text is a critical task in medical NLP but remains hindered by the extreme long-tail distribution of diagnostic codes. Thousands of rare and zero-shot ICD codes are severely underrepresented in datasets like MIMIC-III, leading to low macro-F1 scores. In this work, we propose a data-centric framework that generates high-quality synthetic discharge summaries to mitigate this imbalance. Our method constructs realistic multi-label code sets anchored on rare codes by leveraging real-world co-occurrence patterns, ICD descriptions, synonyms, taxonomy, and similar clinical notes. Using these structured prompts, we generate 90,000 synthetic notes covering 7,902 ICD codes, significantly expanding the training distribution. We fine-tune two state-of-the-art transformer-based models, PLM-ICD and GKI-ICD, on both the original and extended datasets. Experiments show that our approach modestly improves macro-F1 while maintaining strong micro-F1, outperforming prior SOTA. While the gain may seem marginal relative to the computational cost, our results demonstrate that carefully crafted synthetic data can enhance equity in long-tail ICD code prediction.

Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces. In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed Stealth Fine-Tuning. Our method elicits harmful reasoning traces through segment-level interference and reuses the self-generated outputs as supervised fine-tuning data. To facilitate this, we introduce a turn-based weighted loss that minimizes distribution shift. In our experiment, with only 499 samples and under 3 hours on a single A100 (QLoRA), Stealth Fine-Tuning outperforms IDEATOR by 38.66% ASR while preserving general reasoning ability, as the tuned model retains the original representation distribution. Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses. \textcolor{red}{Disclaimer: This paper contains content that may be disturbing or offensive.}

Despite recent progress in 3D-LLMs, they remain limited in accurately grounding language to visual and spatial elements in 3D environments. This limitation stems in part from training data that focuses on language reasoning rather than spatial understanding due to scarce 3D resources, leaving inherent grounding biases unresolved. To address this, we propose 3D scene editing as a key mechanism to generate visual counterfactuals that mitigate these biases through fine-grained spatial manipulation, without requiring costly scene reconstruction or large-scale 3D data collection. Furthermore, to make these edits targeted and directly address the specific weaknesses of the model, we introduce DEER-3D, an error-driven framework that diagnoses grounding failures and generates targeted counterfactual training supervision via a structured "Decompose, Diagnose, Edit, and Retrain" loop. Specifically, given a grounding failure, DEER-3D first identifies the predicate-level error (e.g., attribute or spatial relation). It then performs minimal predicate-aligned scene edits, such as recoloring or repositioning, and constructs aligned question-answer pairs that explicitly target the failed predicate, forming targeted counterfactual training examples. We evaluate our editing pipeline across multiple benchmarks for 3D grounding and scene understanding tasks, consistently demonstrating improvements across all grounding datasets through iterative refinement (4-6% gains). DEER-3D underscores the effectiveness of targeted, error-driven scene editing in bridging linguistic reasoning with spatial grounding in 3D LLMs.

Multi-label sentiment classification plays a vital role in natural language processing by detecting multiple emotions within a single text. However, existing datasets like GoEmotions often suffer from severe class imbalance, which hampers model performance, especially for underrepresented emotions. To address this, we constructed a balanced multi-label sentiment dataset by integrating the original GoEmotions data, emotion-labeled samples from Sentiment140 using a RoBERTa-base-GoEmotions model, and manually annotated texts generated by GPT-4 mini. Our data balancing strategy ensured an even distribution across 28 emotion categories. Based on this dataset, we developed an enhanced multi-label classification model that combines pre-trained FastText embeddings, convolutional layers for local feature extraction, bidirectional LSTM for contextual learning, and an attention mechanism to highlight sentiment-relevant words. A sigmoid-activated output layer enables multi-label prediction, and mixed precision training improves computational efficiency. Experimental results demonstrate significant improvements in accuracy, precision, recall, F1-score, and AUC compared to models trained on imbalanced data, highlighting the effectiveness of our approach.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a core training stage in recent large language models (LLMs). Its reliance on non-public, high-value prompt sets raises concerns about unauthorized data use, creating a need for exposure auditing. A natural tool is membership inference attacks (MIAs), but existing methods detect fitting to a fixed target string. This does not apply to RLVR, which generates responses from the model itself and reinforces successful ones, thus hindering the auditing of data exposure. We show that it remains detectable: RLVR reshapes the model's response distribution on training prompts, producing behavioral traces that can be surfaced through targeted auditing. We propose Divergence-in-Behavior Auditing (DIBA), a white-box query-level auditing framework for RLVR. DIBA compares a fine-tuned model against its pre-RLVR checkpoint along two axes: reward-side evidence capturing changes in verifiable task success, and policy-side evidence capturing prompt-conditioned behavioral drift. By aggregating over multiple stochastic rollouts, DIBA produces a stable query-level auditing signal. Under a white-box setting, DIBA consistently outperforms strong transferred likelihood-based baselines, including calibrated and self-generated variants, achieving around 0.8 AUC and an order-of-magnitude stronger TPR@0.1%FPR. We further show that RLVR auditing is stronger when training leaves non-trivial prompt-specific traces and weaker when the base model already performs well on the prompt. Under a practical grey-box setting, transfer is often robust across model sizes under the same RLVR algorithm, but more varied across algorithms, and can remain useful under distribution shift with carefully chosen shadow data.

AI Scientific Assistant Core (AISAC) is a transparent, modular multi-agent runtime developed at Argonne National Laboratory to support long-horizon, evidence-grounded scientific reasoning. Rather than proposing new agent algorithms or claiming autonomous scientific discovery, AISAC contributes a governed execution substrate that operationalizes key requirements for deploying agentic AI in scientific practice, including explicit role semantics, budgeted context management, traceable execution, and reproducible interaction with tools and knowledge. AISAC enforces four structural guarantees for scientific reasoning: (1) declarative agent registration with runtime-enforced role semantics and automatic system prompt generation; (2) budgeted orchestration via explicit per-turn context and delegation depth limits; (3) role-aligned memory access across episodic, dialogue, and evidence layers; and (4) trace-driven transparency through persistent execution records and a live event-stream interface. These guarantees are implemented through hybrid persistent memory (SQLite and dual FAISS indices), governed retrieval with agent-scoped RAG, structured tool execution with schema validation, and a configuration-driven bootstrap mechanism that enables project specific extension without modifying the shared core. AISAC is currently deployed across multiple scientific workflows at Argonne, including combustion science, materials research, and energy process safety, demonstrating its use as a reusable substrate for domain-specialized AI scientific assistants.

Recent advancements in multimodal out-of-context (OOC) misinformation detection have made remarkable progress in checking the consistencies between different modalities for supporting or refuting image-text pairs. However, existing OOC misinformation detection methods tend to emphasize the role of internal consistency, ignoring the significant of external consistency between image-text pairs and external evidence. In this paper, we propose HiEAG, a novel Hierarchical Evidence-Augmented Generation framework to refine external consistency checking through leveraging the extensive knowledge of multimodal large language models (MLLMs). Our approach decomposes external consistency checking into a comprehensive engine pipeline, which integrates reranking and rewriting, apart from retrieval. Evidence reranking module utilizes Automatic Evidence Selection Prompting (AESP) that acquires the relevant evidence item from the products of evidence retrieval. Subsequently, evidence rewriting module leverages Automatic Evidence Generation Prompting (AEGP) to improve task adaptation on MLLM-based OOC misinformation detectors. Furthermore, our approach enables explanation for judgment, and achieves impressive performance with instruction tuning. Experimental results on different benchmark datasets demonstrate that our proposed HiEAG surpasses previous state-of-the-art (SOTA) methods in the accuracy over all samples.

Post-disaster reconnaissance reports contain critical evidence for understanding multi-hazard interactions, yet their unstructured narratives make systematic knowledge transfer difficult. Large language models (LLMs) offer new potential for analyzing these reports, but often generate unreliable or hallucinated outputs when domain grounding is absent. This study introduces the Mixture-of-Retrieval Agentic RAG (MoRA-RAG), a knowledge-grounded LLM framework that transforms reconnaissance reports into a structured foundation for multi-hazard reasoning. The framework integrates a Mixture-of-Retrieval mechanism that dynamically routes queries across hazard-specific databases while using agentic chunking to preserve contextual coherence during retrieval. It also includes a verification loop that assesses evidence sufficiency, refines queries, and initiates targeted searches when information remains incomplete. We construct HazardRecQA by deriving question-answer pairs from GEER reconnaissance reports, which document 90 global events across seven major hazard types. MoRA-RAG achieves up to 94.5 percent accuracy, outperforming zero-shot LLMs by 30 percent and state-of-the-art RAG systems by 10 percent, while reducing hallucinations across diverse LLM architectures. MoRA-RAG also enables open-weight LLMs to achieve performance comparable to proprietary models. It establishes a new paradigm for transforming post-disaster documentation into actionable, trustworthy intelligence for hazard resilience.

The 3rd Generation Partnership Project (3GPP) produces complex technical specifications essential to global telecommunications, yet their hierarchical structure, dense formatting, and multi-modal content make them difficult to process. While Large Language Models (LLMs) show promise, existing approaches fall short in handling complex queries, visual information, and document interdependencies. We present TelcoAI, an agentic, multi-modal Retrieval-Augmented Generation (RAG) system tailored for 3GPP documentation. TelcoAI introduces section-aware chunking, structured query planning, metadata-guided retrieval, and multi-modal fusion of text and diagrams. Evaluated on multiple benchmarks-including expert-curated queries-our system achieves $87\%$ recall, $83\%$ claim recall, and $92\%$ faithfulness, representing a $16\%$ improvement over state-of-the-art baselines. These results demonstrate the effectiveness of agentic and multi-modal reasoning in technical document understanding, advancing practical solutions for real-world telecommunications research and engineering.

This research digitizes and analyzes the Leidse hoogleraren en lectoren 1575-1815 books written between 1983 and 1985, which contain biographic data about professors and curators of Leiden University. It addresses the central question: how can we design an automated pipeline that integrates OCR, LLM-based interpretation, and database linking to harmonize data from historical document images with existing high-quality database records? We applied OCR techniques, generative AI decoding constraints that structure data extraction, and database linkage methods to process typewritten historical records into a digital format. OCR achieved a Character Error Rate (CER) of 1.08 percent and a Word Error Rate (WER) of 5.06 percent, while JSON extraction from OCR text achieved an average accuracy of 63 percent and, based on annotated OCR, 65 percent. This indicates that generative AI somewhat corrects low OCR performance. Our record linkage algorithm linked annotated JSON files with 94% accuracy and OCR-derived JSON files with 81%. This study contributes to digital humanities research by offering an automated pipeline for interpreting digitized historical documents, addressing challenges like layout variability and terminology differences, and exploring the applicability and strength of an advanced generative AI model.

The rapid advancement of large language models (LLMs) has accelerated their integration into clinical decision support, particularly in prescription review. To enable systematic and fine-grained evaluation, we developed RxBench, a comprehensive benchmark that covers common prescription review categories and consolidates 14 frequent types of prescription errors drawn from authoritative pharmacy references. RxBench consists of 1,150 single-choice, 230 multiple-choice, and 879 short-answer items, all reviewed by experienced clinical pharmacists. We benchmarked 18 state-of-the-art LLMs and identified clear stratification of performance across tasks. Notably, Gemini-2.5-pro-preview-05-06, Grok-4-0709, and DeepSeek-R1-0528 consistently formed the first tier, outperforming other models in both accuracy and robustness. Comparisons with licensed pharmacists indicated that leading LLMs can match or exceed human performance in certain tasks. Furthermore, building on insights from our benchmark evaluation, we performed targeted fine-tuning on a mid-tier model, resulting in a specialized model that rivals leading general-purpose LLMs in performance on short-answer question tasks. The main contribution of RxBench lies in establishing a standardized, error-type-oriented framework that not only reveals the capabilities and limitations of frontier LLMs in prescription review but also provides a foundational resource for building more reliable and specialized clinical tools.

Large language models (LLMs) are increasingly deployed in collaborative settings, yet little is known about how they coordinate when treated as black-box agents. We simulate 7500 multi-agent, multi-round discussions in an inductive coding task, generating over 125000 utterances that capture both final annotations and their interactional histories. We introduce process-level metrics: code stability, semantic self-consistency, and lexical confidence alongside sentiment and convergence measures, to track coordination dynamics. To probe deeper alignment signals, we analyze the evolving geometry of output embeddings, showing that intrinsic dimensionality declines over rounds, suggesting semantic compression. The results reveal that LLM groups converge lexically and semantically, develop asymmetric influence patterns, and exhibit negotiation-like behaviors despite the absence of explicit role prompting. This work demonstrates how black-box interaction analysis can surface emergent coordination strategies, offering a scalable complement to internal probe-based interpretability methods.

This paper investigates the automation of qualitative data analysis, focusing on inductive coding using large language models (LLMs). Unlike traditional approaches that rely on deductive methods with predefined labels, this research investigates the inductive process where labels emerge from the data. The study evaluates the performance of six open-source LLMs compared to human experts. As part of the evaluation, experts rated the perceived difficulty of the quotes they coded. The results reveal a peculiar dichotomy: human coders consistently perform well when labeling complex sentences but struggle with simpler ones, while LLMs exhibit the opposite trend. Additionally, the study explores systematic deviations in both human and LLM generated labels by comparing them to the golden standard from the test set. While human annotations may sometimes differ from the golden standard, they are often rated more favorably by other humans. In contrast, some LLMs demonstrate closer alignment with the true labels but receive lower evaluations from experts.

Multiple Choice Question (MCQ) answering is a widely used method for evaluating the performance of Large Language Models (LLMs). However, LLMs often exhibit selection bias in MCQ tasks, where their choices are influenced by factors like answer position or option symbols rather than the content. This bias undermines the reliability of MCQ as an evaluation framework. Most existing selection bias metrics require answer labels and measure divergences between prediction and answer distributions, but do not fully capture the consistency of a model's predictions across different orderings of answer choices. Existing selection bias mitigation strategies have notable limitations: majority voting, though effective, is computationally prohibitive; calibration-based methods require validation sets and often fail to generalize across datasets. To address these gaps, we propose three key contributions: (1) a new unsupervised label-free Permutation Bias Metric (PBM) that directly quantifies inconsistencies in model predictions across answer permutations, providing a more precise measure of selection bias, (2) an efficient majority voting approach called Batch Question-Context KV caching (BaQCKV), to significantly reduce computational costs while preserving bias mitigation effectiveness, and (3) an unsupervised Low-Rank Adaptation (LoRA-1) fine-tuning strategy based on our proposed metric and the BaQCKV that mitigates selection bias, providing a computationally efficient alternative that maintains model generalizability. Experiments across multiple MCQ benchmarks demonstrate that our approaches reduce bias, increasing consistency in accuracy while minimizing computational costs.

Large language models have recently demonstrated their exceptional capabilities in supporting and automating various tasks. Among the tasks worth exploring for testing large language model capabilities, we considered data preparation, a critical yet often labor-intensive step in data-driven processes. This paper investigates whether large language models can effectively support users in selecting and automating data preparation tasks. To this aim, we considered both general-purpose and fine-tuned tabular large language models. We prompted these models with poor-quality datasets and measured their ability to perform tasks such as data profiling and cleaning. We also compare the support provided by large language models with that offered by traditional data preparation tools. To evaluate the capabilities of large language models, we developed a custom-designed quality model that has been validated through a user study to gain insights into practitioners' expectations.

In goal-oriented dialogue tasks, the main challenge is to steer the interaction towards a given goal within a limited number of turns. Existing approaches either rely on elaborate prompt engineering, whose effectiveness is heavily dependent on human experience, or integrate policy networks and pre-trained policy models, which are usually difficult to adapt to new dialogue scenarios and costly to train. Therefore, in this paper, we present Nested Rollout Policy Adaptation for Goal-oriented Dialogue (NRPA-GD), a novel dialogue policy planning method that completely avoids specific model training by utilizing a Large Language Model (LLM) to simulate behaviors of user and system at the same time. Specifically, NRPA-GD constructs a complete evaluation mechanism for dialogue trajectories and employs an optimization framework of nested Monte Carlo simulation and policy self-adaptation to dynamically adjust policies during the dialogue process. The experimental results on four typical goal-oriented dialogue datasets show that NRPA-GD outperforms both existing prompt engineering and specifically pre-trained model-based methods. Impressively, NRPA-GD surpasses ChatGPT and pre-trained policy models with only a 0.6-billion-parameter LLM. The proposed approach further demonstrates the advantages and novelty of employing planning methods on LLMs to solve practical planning tasks.

AI-generated content (AIGC) technology has emerged as a prevalent alternative to create multimodal misinformation on social media platforms, posing unprecedented threats to societal safety. However, standard prompting leverages multimodal large language models (MLLMs) to identify the emerging misinformation, which ignores the misinformation attribution. To this end, we present Insight-A, exploring attribution with MLLM insights for detecting multimodal misinformation. Insight-A makes two efforts: I) attribute misinformation to forgery sources, and II) an effective pipeline with hierarchical reasoning that detects distortions across modalities. Specifically, to attribute misinformation to forgery traces based on generation patterns, we devise cross-attribution prompting (CAP) to model the sophisticated correlations between perception and reasoning. Meanwhile, to reduce the subjectivity of human-annotated prompts, automatic attribution-debiased prompting (ADP) is used for task adaptation on MLLMs. Additionally, we design image captioning (IC) to achieve visual details for enhancing cross-modal consistency checking. Extensive experiments demonstrate the superiority of our proposal and provide a new paradigm for multimodal misinformation detection in the era of AIGC.

Contemporary large language model deployments typically employ uniform prompting strategies across diverse query types, applying verbose response patterns to both complex analytical tasks and straightforward factual questions. This one-size-fits-all methodology leads to substantial token inefficiency, a concern amplified by the significant cost differential between input and output tokens–the latter commanding 4-8x higher prices across major providers. We present Dynamic Template Selection (DTS), which adaptively matches response templates to query complexity, achieving significant cost reductions without compromising response quality. We compared two routing approaches: a simple MLP that uses pre-computed embeddings and a more complex fine-tuned RoBERTa transformer. Through comprehensive evaluation on 1,000 MMLU questions, we find that the MLP router achieves 90.5% routing accuracy on held-out test data, marginally exceeding RoBERTa's performance (89.5%) despite utilizing 125M fewer parameters. Notably, our empirical analysis reveals provider-agnostic behavior in template selection–routing decisions generalize effectively across 3 major LLM providers (OpenAI GPT-4, Google Gemini, and Anthropic Claude), as validated through 9,000 production API calls. While routing accuracy remains consistent at 90.5% across providers, observed token reductions vary from 32.6% to 33.9%, reflecting provider-specific generation characteristics. This work contributes several key elements: formal problem formulation with theoretical grounding in machine learning, four algorithms with corresponding complexity analyses, and extensive empirical validation across production systems.

We investigate empathy-in-action – the willingness to sacrifice task efficiency to address human needs – as a linear direction in LLM activation space. Using contrastive prompts grounded in the Empathy-in-Action (EIA) benchmark, we test detection and steering across Phi-3-mini-4k (3.8B), Qwen2.5-7B (safety-trained), and Dolphin-Llama-3.1-8B (uncensored). Detection: All models show AUROC 0.996-1.00 at optimal layers. Uncensored Dolphin matches safety-trained models, demonstrating empathy encoding emerges independent of safety training. Phi-3 probes correlate strongly with EIA behavioral scores (r=0.71, p<0.01). Cross-model probe agreement is limited (Qwen: r=-0.06, Dolphin: r=0.18), revealing architecture-specific implementations despite convergent detection. Steering: Qwen achieves 65.3% success with bidirectional control and coherence at extreme interventions. Phi-3 shows 61.7% success with similar coherence. Dolphin exhibits asymmetric steerability: 94.4% success for pro-empathy steering but catastrophic breakdown for anti-empathy (empty outputs, code artifacts). Implications: The detection-steering gap varies by model. Qwen and Phi-3 maintain bidirectional coherence; Dolphin shows robustness only for empathy enhancement. Safety training may affect steering robustness rather than preventing manipulation, though validation across more models is needed.