Papers

The COVID-19 pandemic causes severe social and economic disruption around the world, raising various subjects that are discussed over social media. Identifying pandemic-related named entities as expressed on social media is fundamental and important to understand the discussions about the pandemic. However, there is limited work on named entity recognition on this topic due to the following challenges: 1) COVID-19 texts in social media are informal and their annotations are rare and insufficient to train a robust recognition model, and 2) named entity recognition in COVID-19 requires extensive domain-specific knowledge. To address these issues, we propose a novel entity knowledge augmentation approach for COVID-19, which can also be applied in general biomedical named entity recognition in both informal text format and formal text format. Experiments carried out on the COVID-19 tweets dataset and PubMed dataset show that our proposed entity knowledge augmentation improves NER performance in both fully-supervised and few-shot settings. Our source code is publicly available: https://github.com/kkkenshi/LLM-EKA/tree/master

Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While recent studies document instances of LLM deception, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce a new simulation framework, LH-Deception, for a systematic, empirical quantification of deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. LH-Deception is designed as a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed-source and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal emergent, long-horizon phenomena, such as “chains of deception", which are invisible to static, single-turn evaluations. Our findings provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.

Understanding how patients perceive their physicians is essential to improving trust, communication, and satisfaction. Patients increasingly consult large language models (LLMs) to summarize physician reviews and shape provider choices, yet the national landscape of patient-perceived physician traits remains poorly characterized. We present an LLM-based pipeline that extracts ten patient-perceived physician trait scores from review text: five Big-Five-style and five patient-oriented dimensions. From one million U.S. physicians, we analyze 4.1 million reviews of 226,999 physicians. We validate the pipeline through multi-model comparison and human expert benchmarking. LLM and human-rater trait scores from reviews are consistent. Trait scores correlate strongly with review rating scores yet retain substantial independent variance. Two national-scale patterns emerge: male physicians receive higher trait scores across all traits, with the largest gap in clinical competence; specialty differences are driven by encounter context, with surgical specialties leading interpersonal qualities and psychiatry lowest. Cluster analysis identifies four physician archetypes, from "Uniform High" (33.8%, high across traits) to "Uniform Low" (22.6%, low across traits). This map of LLM-derived physician traits exposes how LLMs read the U.S. clinical workforce. Pending clinical validation, it opens future research on fairness, bias, and how LLM-mediated provider search shapes patient choice.

Natural disaster assessment relies on accurate and rapid access to information, with social media emerging as a valuable real-time source. However, existing datasets suffer from class imbalance and limited samples, making effective model development a challenging task. This paper explores augmentation techniques to address these issues on the CrisisMMD multimodal dataset. For visual data, we apply diffusion-based methods, namely Real Guidance and DiffuseMix. For text data, we explore back-translation, paraphrasing with transformers, and image caption-based augmentation. We evaluated these across unimodal, multimodal, and multi-view learning setups. Results show that selected augmentations improve classification performance, particularly for underrepresented classes, while multi-view learning introduces potential but requires further refinement. This study highlights effective augmentation strategies for building more robust disaster assessment systems.

Test-Time Scaling (TTS) improves LLM reasoning by exploring multiple candidate responses and then operating over this set to find the best output. A tacit premise behind TTS is that sufficiently diverse candidate pools enhance reliability. In this work, we show that this assumption in TTS introduces a previously unrecognized failure mode. When candidate diversity is curtailed, even by a modest amount, TTS becomes much more likely to produce unsafe outputs. We present a reference-guided diversity reduction protocol (RefDiv) that serves as a diagnostic attack to stress test TTS pipelines. Through extensive experiments across open-source models (e.g. Qwen3, Mistral, Llama3.1, Gemma3) and two widely used TTS strategies (Monte Carlo Tree Search and Best-of-N), constraining diversity consistently signifies the rate at which TTS produces unsafe results. The effect is often stronger than that produced by prompts directly with high adversarial intent scores. This observed phenomenon also transfers across TTS strategies and to closed-source models (e.g. OpenAI o3-mini and Gemini-2.5-Pro), thus indicating that this is a general and extant property of TTS rather than a model-specific artifact. Additionally, we find that numerous widely used safety guardrail classifiers (e.g. Llama-Guard), are unable to flag the adversarial input prompts generated by RefDiv, demonstrating that existing defenses offer limited protection against this diversity-driven failure mode.

Embedding vision-language models (VLMs) are typically pretrained with short text windows (<77 tokens), which forces the truncation of long-format captions. Yet, the distribution of biomedical captions from large-scale open source literature reveals that a huge portion of captions far exceed 77 tokens. To this end, we investigate the impact of pretraining on long-format biomedical captions by extending the context length of text encoders in VLMs. We find that longer context (thus, enabling additional supervision provided in long-format captions) correlates with better retrieval and classification performance. Given this finding, we introduce BIOMEDICA-LongCAP, a dataset of 1M image-caption pairs enriched with context-aware descriptions from full-text articles, providing longer and additional textual supervision. Using BIOMEDICA-LongCAP, we train BMC-LongCLIP, a long-context biomedical VLM with a text encoder supporting windows of up to 512 tokens. Our model extends context capacity by 6.6x, reducing token waste from 55% to just 2.2%. On long-caption retrieval benchmarks, BMC-LongCLIP achieves up to +30% absolute gains in Recall@1 and +2% average improvements in classification, while also converging faster than short-context. Our results demonstrate that long-context modeling is a promising direction for advancing biomedical VLMs.

Multi-LLM collaboration promises accurate, robust, and context-aware solutions, yet existing approaches rely on implicit selection and output assessment without analyzing whether collaborating models truly complement or conflict. We introduce LLM Chemistry – a framework that measures when LLM combinations exhibit synergistic or antagonistic behaviors that shape collective performance beyond individual capabilities. We formalize the notion of chemistry among LLMs, propose algorithms that quantify it by analyzing interaction dependencies, and recommend optimal model ensembles accordingly. Our theoretical analysis shows that chemistry among collaborating LLMs is most evident under heterogeneous model profiles, with its outcome impact shaped by task type, group size, and complexity. Evaluation on classification, summarization, and program repair tasks provides initial evidence for these task-dependent effects, thereby reinforcing our theoretical results. This establishes LLM Chemistry as both a diagnostic factor in multi-LLM systems and a foundation for ensemble recommendation.

This study presents PsychoLexTherapy, a framework for simulating psychotherapeutic reasoning in Persian using small language models (SLMs). The framework tackles the challenge of developing culturally grounded, therapeutically coherent dialogue systems with structured memory for multi-turn interactions in underrepresented languages. To ensure privacy and feasibility, PsychoLexTherapy is optimized for on-device deployment, enabling use without external servers. Development followed a three-stage process: (i) assessing SLMs psychological knowledge with PsychoLexEval; (ii) designing and implementing the reasoning-oriented PsychoLexTherapy framework; and (iii) constructing two evaluation datasets-PsychoLexQuery (real Persian user questions) and PsychoLexDialogue (hybrid simulated sessions)-to benchmark against multiple baselines. Experiments compared simple prompting, multi-agent debate, and structured therapeutic reasoning paths. Results showed that deliberate model selection balanced accuracy, efficiency, and privacy. On PsychoLexQuery, PsychoLexTherapy outperformed all baselines in automatic LLM-as-a-judge evaluation and was ranked highest by human evaluators in a single-turn preference study. In multi-turn tests with PsychoLexDialogue, the long-term memory module proved essential: while naive history concatenation caused incoherence and information loss, the full framework achieved the highest ratings in empathy, coherence, cultural fit, and personalization. Overall, PsychoLexTherapy establishes a practical, privacy-preserving, and culturally aligned foundation for Persian psychotherapy simulation, contributing novel datasets, a reproducible evaluation pipeline, and empirical insights into structured memory for therapeutic reasoning.

Detecting media bias is crucial, specifically in the South Asian region. Despite this, annotated datasets and computational studies for Bangla political bias research remain scarce. Crucially because, political stance detection in Bangla news requires understanding of linguistic cues, cultural context, subtle biases, rhetorical strategies, code-switching, implicit sentiment, and socio-political background. To address this, we introduce the first benchmark dataset of 200 politically significant and highly debated Bangla news articles, labeled for government-leaning, government-critique, and neutral stances, alongside diagnostic analyses for evaluating large language models (LLMs). Our comprehensive evaluation of 28 proprietary and open-source LLMs shows strong performance in detecting government-critique content (F1 up to 0.83) but substantial difficulty with neutral articles (F1 as low as 0.00). Models also tend to over-predict government-leaning stances, often misinterpreting ambiguous narratives. This dataset and its associated diagnostics provide a foundation for advancing stance detection in Bangla media research and offer insights for improving LLM performance in low-resource languages.

We present a gamified explainable AI (XAI) system for ethically aware consumer decision-making in the coffee domain. Each session comprises six rounds with three options per round. Two symbolic engines provide real-time reasons: a Kantian module flags rule violations (e.g., child labor, deforestation risk without shade certification, opaque supply chains, unsafe decaf), and a utilitarian module scores options via multi-criteria aggregation over normalized attributes (price, carbon, water, transparency, farmer income share, taste/freshness, packaging, convenience). A meta-explainer with a regret bound (0.2) highlights Kantian–utilitarian (mis)alignment and switches to a deontically clean, near-parity option when welfare loss is small. We release a structured configuration (attribute schema, certification map, weights, rule set), a policy trace for auditability, and an interactive UI.

Reinforcement learning with verifiable rewards (RLVR) has recently enhanced the reasoning capabilities of large language models (LLMs), particularly for mathematical problem solving. However, a fundamental limitation remains: as the sampling budget increases, the advantage of RLVR-trained models over their pretrained bases often diminishes or even vanishes, revealing a strong dependence on the base model's restricted search space. We attribute this phenomenon to the widespread use of the reverse Kullback-Leibler (KL) divergence regularizer, whose mode-seeking behavior keeps the policy trapped inside the base model's support region and hampers wider exploration. To address this issue, we propose RAPO (Rewards-Aware Policy Optimization), an algorithm to promote broader yet focused exploration. Our method (i) utilizes the forward KL penalty to replace the reverse KL penalty for out-of-distribution exploration, and (ii) reweights the reference policy to facilitate adaptive in-distribution exploration. We train Qwen2.5-3B and 7B models with RAPO on the 8K SimpleRL-Zero dataset, without supervised fine-tuning, and evaluate them on AIME2024 and AIME2025. Results show that RAPO consistently improves problem-solving performance. Notably, RAPO enables models to surpass the base model's performance ceiling and solves previously intractable problems, advancing the frontier of RLVR for challenging reasoning tasks.

This research explores the annotation of rhetorical relations in discourse using the INCEpTION tool and compares manual annotation with automatic approaches based on large language models. The study focuses on sports reports (specifically cricket news) and evaluates the performance of BERT, DistilBERT, and Logistic Regression models in classifying rhetorical relations such as elaboration, contrast, background, and cause-effect. The results show that DistilBERT achieved the highest accuracy, highlighting its potential for efficient discourse relation prediction. This work contributes to the growing intersection of discourse parsing and transformer-based NLP. (This paper was conducted as part of an academic requirement under the supervision of Prof. Dr. Ralf Klabunde, Linguistic Data Science Lab, Ruhr University Bochum.) Keywords: Rhetorical Structure Theory, INCEpTION, BERT, DistilBERT, Discourse Parsing, NLP.

Large Reasoning Models (LRMs) demonstrate strong performance on complex tasks but often suffer from excessive verbosity, known as "overthinking." Existing solutions via reinforcement learning (RL) typically penalize generated tokens to promote conciseness. However, these methods encounter two challenges: responses with fewer tokens do not always correspond to fewer reasoning steps, and models may develop hacking behavior in later stages of training by discarding reasoning steps to minimize token usage. In this work, we introduce Step Pruner (SP), an RL framework that steers LRMs toward more efficient reasoning by favoring compact reasoning steps. Our step-aware reward function prioritizes correctness while imposing penalties for redundant steps, and withholds rewards for incorrect responses to prevent the reinforcement of erroneous reasoning. Moreover, we propose a dynamic stopping mechanism: when the model's output no longer shortens, training is halted to prevent hacking behavior caused by the merging of steps. Extensive experiments across four reasoning benchmarks demonstrate that SP achieves state-of-the-art accuracy while significantly reducing response length. For instance, on AIME24, SP reduces token usage by 69.7%.

This paper explores the ability of large language models to generate and recognize deep cognitive frames, particularly in socio-political contexts. We demonstrate that LLMs are highly fluent in generating texts that evoke specific frames and can recognize these frames in zero-shot settings. Inspired by mechanistic interpretability research, we investigate the location of the `strict father' and `nurturing parent' frames within the model's hidden representation, identifying singular dimensions that correlate strongly with their presence. Our findings contribute to understanding how LLMs capture and express meaningful human concepts.

Personalized Conversational Information Retrieval (CIR) has seen rapid progress in recent years, driven by the development of Large Language Models (LLMs). Personalized CIR aims to enhance document retrieval by leveraging user-specific information, such as preferences, knowledge, or constraints, to tailor responses to individual needs. A key resource for this task is the TREC iKAT 2023 dataset, designed to evaluate personalization in CIR pipelines. Building on this resource, Mo et al. explored several strategies for incorporating Personal Textual Knowledge Bases (PTKB) into LLM-based query reformulation. Their findings suggested that personalization from PTKBs could be detrimental and that human annotations were often noisy. However, these conclusions were based on single-run experiments using the GPT-3.5 Turbo model, raising concerns about output variability and repeatability. In this reproducibility study, we rigorously reproduce and extend their work, focusing on LLM output variability and model generalization. We apply the original methods to the new TREC iKAT 2024 dataset and evaluate a diverse range of models, including Llama (1B-70B), Qwen-7B, GPT-4o-mini. Our results show that human-selected PTKBs consistently enhance retrieval performance, while LLM-based selection methods do not reliably outperform manual choices. We further compare variance across datasets and observe higher variability on iKAT than on CAsT, highlighting the challenges of evaluating personalized CIR. Notably, recall-oriented metrics exhibit lower variance than precision-oriented ones, a critical insight for first-stage retrievers. Finally, we underscore the need for multi-run evaluations and variance reporting when assessing LLM-based CIR systems. By broadening evaluation across models, datasets, and metrics, our study contributes to more robust and generalizable practices for personalized CIR.

This paper presents the development of Rezwan, a large-scale AI-assisted Hadith corpus comprising over 1.2M narrations, extracted and structured through a fully automated pipeline. Building on digital repositories such as Maktabat Ahl al-Bayt, the pipeline employs Large Language Models (LLMs) for segmentation, chain–text separation, validation, and multi-layer enrichment. Each narration is enhanced with machine translation into twelve languages, intelligent diacritization, abstractive summarization, thematic tagging, and cross-text semantic analysis. This multi-step process transforms raw text into a richly annotated research-ready infrastructure for digital humanities and Islamic studies. A rigorous evaluation was conducted on 1,213 randomly sampled narrations, assessed by six domain experts. Results show near-human accuracy in structured tasks such as chain–text separation (9.33/10) and summarization (9.33/10), while highlighting ongoing challenges in diacritization and semantic similarity detection. Comparative analysis against the manually curated Noor Corpus demonstrates the superiority of Najm in both scale and quality, with a mean overall score of 8.46/10 versus 3.66/10. Furthermore, cost analysis confirms the economic feasibility of the AI approach: tasks requiring over 229,000 hours of expert labor were completed within months at a fraction of the cost. The work introduces a new paradigm in religious text processing by showing how AI can augment human expertise, enabling large-scale, multilingual, and semantically enriched access to Islamic heritage.

Recent advances in Large Language Models (LLMs) have significantly reshaped the landscape of Natural Language Processing (NLP). Among the various prompting techniques, few-shot prompting has gained considerable attention for its practicality and effectiveness. This study investigates how few-shot prompting strategies impact the Word Sense Disambiguation (WSD) task, particularly focusing on the biases introduced by imbalanced sample distributions. We use the GLOSSGPT prompting method, an advanced approach for English WSD, to test its effectiveness across five languages: English, German, Spanish, French, and Italian. Our results show that imbalanced few-shot examples can cause incorrect sense predictions in multilingual languages, but this issue does not appear in English. To assess model behavior, we evaluate both the GPT-4o and LLaMA-3.1-70B models and the results highlight the sensitivity of multilingual WSD to sample distribution in few-shot settings, emphasizing the need for balanced and representative prompting strategies.

Parkinson's Disease (PD) affects over 10 million people worldwide, with speech impairments in up to 89% of patients. Current speech-based detection systems analyze entire utterances, potentially overlooking the diagnostic value of specific phonetic elements. We developed a granularity-aware approach for multilingual PD detection using an automated pipeline that extracts time-aligned phonemes, syllables, and words from recordings. Using Italian, Spanish, and English datasets, we implemented a bidirectional LSTM with multi-head attention to compare diagnostic performance across the different granularity levels. Phoneme-level analysis achieved superior performance with AUROC of 93.78% +- 2.34% and accuracy of 92.17% +- 2.43%. This demonstrates enhanced diagnostic capability for cross-linguistic PD detection. Importantly, attention analysis revealed that the most informative speech features align with those used in established clinical protocols: sustained vowels (/a/, /e/, /o/, /i/) at phoneme level, diadochokinetic syllables (/ta/, /pa/, /la/, /ka/) at syllable level, and /pataka/ sequences at word level. Source code will be available at https://github.com/jetliqs/clearpd.

Large Language Models (LLMs) have consistently demonstrated strong performance in machine translation, especially when guided by high-quality prompts. Few-shot prompting is an effective technique to improve translation quality; however, most existing example selection methods focus solely on query-to-example similarity and do not account for the quality of the examples. In this work, we propose TreePrompt, a novel example selection approach that learns LLM preferences to identify high-quality, contextually relevant examples within a tree-structured framework. To further explore the balance between similarity and quality, we combine TreePrompt with K-Nearest Neighbors (K-NN) and Adaptive Few-Shot Prompting (AFSP). Evaluations on two language pairs - English-Persian (MIZAN) and English-German (WMT19) - show that integrating TreePrompt with AFSP or Random selection leads to improved translation performance.

The rapid development of parameter-efficient fine-tuning methods has noticeably improved the efficiency of adapting large language models. Among these, LoRA has gained widespread popularity due to its strong balance of effectiveness and parameter efficiency. However, LoRA relies on initializing two low-rank matrices whose product is zero, which limits its ability to effectively activate and leverage the original model weights-creating a potential bottleneck for optimal performance. To address this limitation, we propose IniLoRA, a novel initialization strategy that initializes the low-rank matrices to closely approximate the original model weights. Experimental results indicate that IniLoRA achieves better performance than LoRA across a range of models and tasks. Additionally, we introduce two variants, IniLoRA-$\alpha$ and IniLoRA-$\beta$, both leveraging distinct initialization methods to enhance performance further.

Humans understand the world through the integration of multiple sensory modalities, enabling them to perceive, reason about, and imagine dynamic physical processes. Inspired by this capability, multimodal foundation models (MFMs) have emerged as powerful tools for multimodal understanding and generation. However, today's MFMs fall short of serving as effective world models. They lack the essential ability such as perform counterfactual reasoning, simulate dynamics, understand the spatiotemporal information, control generated visual outcomes, and perform multifaceted reasoning. We investigates what it takes to bridge the gap between multimodal foundation models and world models. We begin by improving the reasoning capabilities of MFMs through discriminative tasks and equipping MFMs with structured reasoning skills, such as causal inference, counterfactual thinking, and spatiotemporal reasoning, enabling them to go beyond surface correlations and understand deeper relationships within visual and textual data. Next, we explore generative capabilities of multimodal foundation models across both image and video modalities, introducing new frameworks for structured and controllable generation. Our approaches incorporate scene graphs, multimodal conditioning, and multimodal alignment strategies to guide the generation process, ensuring consistency with high-level semantics and fine-grained user intent. We further extend these techniques to controllable 4D generation, enabling interactive, editable, and morphable object synthesis over time and space.

We propose a speaker-attributed (SA) Whisper-based model for multi-talker speech recognition that combines target-speaker modeling with serialized output training (SOT). Our approach leverages a Diarization-Conditioned Whisper (DiCoW) encoder to extract target-speaker embeddings, which are concatenated into a single representation and passed to a shared decoder. This enables the model to transcribe overlapping speech as a serialized output stream with speaker tags and timestamps. In contrast to target-speaker ASR systems such as DiCoW, which decode each speaker separately, our approach performs joint decoding, allowing the decoder to condition on the context of all speakers simultaneously. Experiments show that the model outperforms existing SOT-based approaches and surpasses DiCoW on multi-talker mixtures (e.g., LibriMix).

Vision-language models trained on large-scale multimodal datasets show strong demographic biases, but the role of training data in producing these biases remains unclear. A major barrier has been the lack of demographic annotations in web-scale datasets such as LAION-400M. We address this gap by creating person-centric annotations for the full dataset, including over 276 million bounding boxes, perceived gender and race/ethnicity labels, and automatically generated captions. These annotations are produced through validated automatic labeling pipelines combining object detection, multimodal captioning, and finetuned classifiers. Using them, we uncover demographic imbalances and harmful associations, such as the disproportionate linking of men and individuals perceived as Black or Middle Eastern with crime-related and negative content. We also show that a linear fit predicts 60-70% of gender bias in CLIP and Stable Diffusion from direct co-occurrences in the data. Our resources establish the first large-scale empirical link between dataset composition and downstream model bias. Code is available at https://github.com/ExplainableML/LAION-400M-Person-Centric-Annotations.

Evaluating the quality of multi-turn chatbot interactions remains challenging, as most existing methods assess interactions at the turn level without addressing whether a user's overarching goal was fulfilled. A “goal” here refers to an information need or task, such as asking for policy information or applying for leave. We propose a comprehensive framework for goal-oriented evaluation of multi-agent systems (MAS), introducing the Goal Success Rate (GSR) to measure the percentage of fulfilled goals, and a Root Cause of Failure (RCOF) taxonomy to identify reasons for failure in multi-agent chatbots. Our method segments conversations by user goals and evaluates success using all relevant turns. We present a model-based evaluation system combining teacher LLMs, where domain experts define goals, set quality standards serving as a guidance for the LLMs. The LLMs use “thinking tokens” to produce interpretable rationales, enabling explainable, data-efficient evaluations. In an enterprise setting, we apply our framework to evaluate AIDA, a zero-to-one employee conversational agent system built as a ground-up multi-agent conversational agent, and observe GSR improvement from 63% to 79% over six months since its inception. Our framework is generic and offers actionable insights through a detailed defect taxonomy based on analysis of failure points in multi-agent chatbots, diagnosing overall success, identifying key failure modes, and informing system improvements.

Medical problem-solving demands expert knowledge and intricate reasoning. Recent studies of large language models (LLMs) attempt to ease this complexity by introducing external knowledge verification through retrieval-augmented generation or by training on reasoning datasets. However, these approaches suffer from drawbacks such as retrieval overhead and high annotation costs, and they heavily rely on substituted external assistants to reach limited performance in medical field. In this paper, we introduce MedReflect, a generalizable framework designed to inspire LLMs with a physician-like reflective thinking mode. MedReflect generates a single-pass reflection chain that includes initial hypothesis generation, self-questioning, self-answering and decision refinement. This self-verified and self-reflective nature releases large language model's latent capability in medical problem-solving without external retrieval or heavy annotation. We demonstrate that MedReflect enables cost-efficient medical dataset construction. With only a minimal subset of randomly sampled training examples and lightweight fine-tuning, this approach achieves notable absolute accuracy improvements across a series of medical benchmarks while significantly cutting annotation requirements. Our results provide evidence that LLMs can learn to solve specialized medical problems via self-reflection and self-improvement, reducing reliance on external supervision and extensive task-specific fine-tuning data.

The use of derogatory terms in languages that employ code mixing, such as Roman Urdu, presents challenges for Natural Language Processing systems due to unstated grammar, inconsistent spelling, and a scarcity of labeled data. In this work, we propose a QLoRA based fine tuning framework to improve offensive language detection in Roman Urdu-English text. We translated the Roman Urdu-English code mixed dataset into English using Google Translate to leverage English LLMs, while acknowledging that this translation reduces direct engagement with code mixing features. Our focus is on classification performance using English translated low resource inputs. We fine tuned several transformers and large language models, including Meta LLaMA 3 8B, Mistral 7B v0.1, LLaMA 2 7B, ModernBERT, and RoBERTa, with QLoRA for memory efficient adaptation. Models were trained and evaluated on a manually annotated Roman Urdu dataset for offensive vs non offensive content. Of all tested models, the highest F1 score of 91.45 was attained by Meta LLaMA 3 8B, followed by Mistral 7B at 89.66, surpassing traditional transformer baselines. These results demonstrate the efficacy of QLoRA in fine tuning high performing models for low resource environments such as code mixed offensive language detection, and confirm the potential of LLMs for this task. This work advances a scalable approach to Roman Urdu moderation and paves the way for future multilingual offensive detection systems based on LLMs.

Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token's influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO's learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.

Multimodal retrieval-augmented Generation (MM-RAG) is a key approach for applying large language models (LLMs) and agents to real-world knowledge bases, yet current evaluations are fragmented – focusing on either text or images in isolation, or simplified multimodal setup, failing to capture document-centric multimodal use cases. In this paper, we introduce UniDoc-Bench, the first large-scale, realistic benchmark for MM-RAG built from $k$ real-world PDF pages across domains. Our pipeline extracts and links evidence from text, tables, and figures, then generates multimodal QA pairs spanning factual retrieval, comparison, summarization, and logical reasoning queries. To ensure reliability, all of QA pairs are validated by multiple human annotators and expert adjudication. UniDoc-Bench supports apples-to-apples comparison across four paradigms: 1) text-only, 2) image-only, 3) multimodal text-image fusion and 4) multimodal joint retrieval – under a unified protocol with standardized candidate pools, prompts, and evaluation metrics. UniDoc-Bench can also be used to evaluate Visual Question Answering (VQA) tasks. Our experiments show that multimodal text-image fusion RAG systems consistently outperform both unimodal and jointly multimodal embedding-based retrieval, indicating that neither text nor images alone are sufficient and that current multimodal embeddings remain inadequate. Beyond benchmarking, our analysis reveals when and how visual context complements textual evidence, uncovers systematic failure modes, and offers actionable guidance for developing more robust MM-RAG pipelines.

Sparse Autoencoders (SAEs) are widely used to steer large language models (LLMs), based on the assumption that their interpretable features naturally enable effective model behavior steering. Yet, a fundamental question remains unanswered: does higher interpretability indeed imply better steering utility? To answer this question, we train 90 SAEs across three LLMs (Gemma-2-2B, Qwen-2.5-3B, Gemma-2-9B), spanning five architectures and six sparsity levels, and evaluate their interpretability and steering utility based on SAEBench (arXiv:2501.12345) and AxBench (arXiv:2502.23456) respectively, and perform a rank-agreement analysis via Kendall's rank coefficients (tau b). Our analysis reveals only a relatively weak positive association (tau b approx 0.298), indicating that interpretability is an insufficient proxy for steering performance. We conjecture the interpretability utility gap may stem from the selection of SAE features, as not all of them are equally effective for steering. To further find features that truly steer the behavior of LLMs, we propose a novel selection criterion called Delta Token Confidence, which measures how much amplifying a feature changes the next token distribution. We show that our method improves the steering performance of three LLMs by 52.52 percent compared to the current best output score based criterion (arXiv:2503.34567). Strikingly, after selecting features with high Delta Token Confidence, the correlation between interpretability and utility vanishes (tau b approx 0), and can even become negative. This further highlights the divergence between interpretability and utility for the most effective steering features.

Training speech recognizers with unpaired speech and text – known as unsupervised speech recognition (UASR) – is a crucial step toward extending ASR to low-resource languages in the long-tail distribution and enabling multimodal learning from non-parallel data. However, existing approaches based on phones often rely on costly resources such as grapheme-to-phoneme converters (G2Ps) and struggle to generalize to languages with ambiguous phoneme boundaries due to training instability. In this paper, we address both challenges by introducing a syllable-level UASR framework based on masked language modeling, which avoids the need for G2P and the instability of GAN-based methods. Our approach achieves up to a 40% relative reduction in character error rate (CER) on LibriSpeech and generalizes effectively to Mandarin, a language that has remained particularly difficult for prior methods. Code will be released upon acceptance.