Peer Revenue Publish or profit

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Saturday 2 May

Computation and Language

MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents

Hung Pham Van, Nguyen Manh Hieu, Khang Pham Tran Tuan et al. (+4 more)

Large Language Models (LLMs) lack persistent memory for long-term personalized conversations. Existing graph-based memory systems suffer from information dilution, absent provenance tracking, and uniform retrieval that ignores query context. We introduce MemORAI (Memory Organization and Retrieval via Adaptive Graph Intelligence), a framework that integrates three innovations: selective memory filtering with dual-layer compression to retain user-persona-relevant content, a provenance-enriched multi-relational graph tracking factual origins at the turn level, and query-adaptive subgraph retrieval with Dynamic Weighted PageRank that applies query-conditioned edge weighting. Evaluated on LOCOMO and LongMemEval benchmarks, MemORAI achieves state-of-the-art performance in memory retrieval and personalized response generation, demonstrating that selective storage, enriched representation, and adaptive retrieval are essential for coherent, personalized LLM agents.
Read the full abstract

All authors Hung Pham Van, Nguyen Manh Hieu, Khang Pham Tran Tuan, Nam Le Hai, Linh Ngo Van, Nguyen Thi Ngoc Diep, Trung Le

Computation and Language

A framework for analyzing concept representations in neural models

Burin Naowarat, Hao Tang and Sharon Goldwater

Understanding how neural models represent human-interpretable concepts is challenging. Prior work has explored linear concept subspaces from diverse perspectives, such as probing and concept erasure. We introduce a unified framework to study these subspaces along two axes: containment, which tests if a concept is fully represented in a subspace but not outside it, and disentanglement, which tests for isolation from other concepts. In experiments on both text and speech models, we first highlight that concept subspaces may not be uniquely determined, and discuss the implications for concept subspace analysis. Then, we compare properties of concept subspaces estimated using five estimators, proposed in different communities. We find that (1) the choice of estimator impacts the containment and disentanglement properties; (2) the state-of-the-art concept erasure method, LEACE, performs well on both testing axes, but still struggles to generalize to unseen data; and (3) in HuBERT speech representations, phone information is both contained and disentangled from speaker information, while speaker information is hard to contain in a compact subspace, despite being disentangled from phones.
Read the full abstract

Computation and Language

MTA: Multi-Granular Trajectory Alignment for Large Language Model Distillation

Pham Khanh Chi, Quoc Phong Dao, Thuat Nguyen et al. (+3 more)

Knowledge distillation is a key technique for compressing large language models (LLMs), but most existing methods align representations at fixed layers or token-level outputs, ignoring how representations evolve across depth. As a result, the student is only weakly guided to capture the teacher's internal relational structure during distillation, which limits knowledge transfer. To address this limitation, we propose Multi-Granular Trajectory Alignment (MTA), a framework that aligns teacher and student representations along their layer-wise transformation trajectory. MTA adopts a layer-adaptive strategy: lower layers are aligned at the word level to preserve lexical information, while higher layers operate on phrase-level spans (e.g., noun and verb phrases) to capture compositional semantics. We instantiate this idea through a Dynamic Structural Alignment loss that matches the relative geometry among semantic units within each layer. This design is motivated by empirical findings that Transformer representations become increasingly abstract with depth, and is also consistent with linguistic views in which higher-level meaning emerges through the composition of lower-level lexical units. We further incorporate a Hidden Representation Alignment loss to directly align selected teacher-student layers. Experiments show that MTA consistently outperforms state-of-the-art baselines on standard benchmarks, with ablations confirming the contribution of each component.
Read the full abstract

All authors Pham Khanh Chi, Quoc Phong Dao, Thuat Nguyen, Linh Ngo Van, Trung Le, Thanh Hong Nguyen

Computation and Language

Focus on the Core: Empowering Diffusion Large Language Models by Self-Contrast

Jinyuan Feng, Xin Yu, Yiqun Chen et al. (+5 more)

The iterative denoising paradigm of Diffusion Large Language Models (DLMs) endows them with a distinct advantage in global context modeling. However, current decoding strategies fail to leverage this capability, typically exhibiting a local preference that overlooks the heterogeneous information density within the context, ultimately degrading generation quality. To address this limitation, we systematically investigate high-information-density (HD) tokens and present two key findings: (1) explicitly conditioning on HD tokens substantially improves output quality; and (2) HD tokens exhibit an early-decoding tendency, converging earlier than surrounding tokens. Motivated by these findings, we propose Focus on the Core (FoCore), a training-free decoding strategy that utilizes HD tokens in a self-contrast manner, wherein HD tokens are temporarily remasked as negative samples, to guide generation. We further introduce FoCore\_Accelerate (FoCore\_A), an efficient variant that, upon detecting HD token convergence, performs parallel decoding over stable candidates within a local context window, substantially accelerating generation. Extensive experiments on math, code and logical reasoning benchmarks demonstrate that FoCore consistently improves generation quality and efficiency across both LLaDA and Dream backbones. For instance, on HumanEval, FoCore improves pass@1 from 39.02 to 42.68 over standard Classifier-Free Guidance, while FoCore-A reduces the number of decoding steps by 2.07x and per-sample latency from 20.76s to 8.64s (-58.4%).
Read the full abstract

All authors Jinyuan Feng, Xin Yu, Yiqun Chen, Xiaochi Wei, Yan Gao, Yi Wu, Yao Hu, Zhiqiang Pu

Computation and Language

Embedding-based In-Context Prompt Training for Enhancing LLMs as Text Encoders

Ailiang Lin, Zhuoyun Li, Keyu Mao et al. (+2 more)

Large language models (LLMs) have been widely explored for embedding generation. While recent studies show that in-context learning (ICL) effectively enhances the representational capability of LLMs by prepending a few task-related demonstrations, it causes substantial token overhead due to the increased sequence length. In this work, we propose EPIC, a novel embedding-based in-context prompt training strategy that leverages ICL to generate high-quality embeddings while reducing computational burden during both training and inference. This approach replaces discrete text demonstrations with their corresponding continuous embeddings, which not only encourages the LLM to align semantically-related text pairs during contrastive learning, but also requires the model to interpret demonstration embeddings as part of the in-context prompt. Consequently, EPIC-trained models achieve excellent embedding performance both with or without in-context prompts at inference time. Comprehensive experiments demonstrate that our method establishes new state-of-the-art results on the MTEB benchmark, surpassing frontier models trained solely on publicly available retrieval data. Extensive ablation studies further validate the effectiveness and necessity of our mechanism.
Read the full abstract

All authors Ailiang Lin, Zhuoyun Li, Keyu Mao, Kotaro Funakoshi, Manabu Okumura

Computation and Language

On Stable Long-Form Generation: Benchmarking and Mitigating Length Volatility

Zhitao He, Haolin Yang, Rui Min et al. (+2 more)

Large Language Models (LLMs) excel at long-context understanding but exhibit significant limitations in long-form generation. Existing studies primarily focus on single-generation quality, generally overlooking the volatility of the output. This volatility not only leads to significant computational costs but also severely impacts the models' reliable application. To address this gap, our work unfolds in three stages: benchmarking, probing, and mitigation. We first propose the VOlatility in Long-form Text Benchmark (VOLTBench), a novel heterogeneous-task benchmark designed to systematically quantify the length volatility of long-form generation. Subsequently, by analyzing attention traces, we conduct an in-depth probe to identify several common internal patterns that cause this volatility. Finally, to mitigate long-form output volatility, we propose Stable Generation via Logits Boosting (GLoBo), a lightweight decoding-stage optimization strategy, designed to significantly enhance both the length accuracy and stability of long-form generation without additional training. Extensive experiments on VOLTBench provide the first systematic confirmation of severe long-form output instability in mainstream models and validate that our proposed method successfully improves the mean output length of the base model by 148% and reduces the length volatility by 69%, while maintaining high generation quality.
Read the full abstract

All authors Zhitao He, Haolin Yang, Rui Min, Zeyu Qin, Yi R. Fung

Computation and Language

LLM Output Detectability and Task Performance Can be Jointly Optimized

Koshiro Saito, Ryuto Koike, Masahiro Kaneko et al. (+1 more)

Detecting machine-generated text is essential for transparency and accountability when deploying LLMs. Watermarking enables statistically reliable detection by biasing token distributions to embed detectable signals into LLM outputs. However, it has been reported that watermarked LLMs often perform worse on downstream tasks. We propose PUPPET, a framework that fine-tunes an LLM via DPO to generate text that is both more detectable by a target detector and better performing on downstream tasks. We use two rewards: a detector that outputs a machine-class likelihood and an evaluator that measures a task-specific metric. Just as a watermark is verified with its secret key, this detector-specific design lets an LLM provider track how its published model is used. Experiments on long-form QA, summarization, and essay writing show that LLMs trained with PUPPET achieve detectability competitive with watermarking methods—even at strict low FPRs—while outperforming them on downstream tasks. Moreover, this optimization requires only a few thousand samples and 1–2 GPU hours, and its gains hold across out-of-domain tasks, six detectors of diverse architectures, and different LLM families and sizes, and are even robust to paraphrasing attacks.
Read the full abstract

All authors Koshiro Saito, Ryuto Koike, Masahiro Kaneko, Naoaki Okazaki

Computation and Language

MAD-OPD: Breaking the Ceiling in On-Policy Distillation via Multi-Agent Debate

Jianze Wang, Ying Liu, Jinlong Chen et al. (+7 more)

On-policy distillation (OPD) trains a student on its own trajectories under token-level teacher supervision, but existing methods are capped by a single-teacher capability ceiling: when the teacher errs, the student inherits the error. OPD also remains largely unexplored in agentic tasks, where per-step errors compound across long trajectories and destabilize training. We propose MAD-OPD (Multi-Agent Debate-driven On-Policy Distillation), which breaks this ceiling by recasting the distillation teacher as a deliberative collective of teachers that debate over the student's on-policy state; the debate produces an emergent collective intelligence that supplies token-level supervision, with each teacher's contribution weighted by its post-debate confidence. To extend OPD to agentic tasks, we also introduce On-Policy Agentic Distillation (OPAD), which adds step-level sampling to stabilize training under multi-step error compounding. We additionally derive a task-adaptive divergence principle, selecting JSD (Jensen-Shannon divergence) for agentic stability and reverse KL (Kullback-Leibler) divergence for code generation, and verify it both theoretically and empirically. Across six teacher-student configurations (Qwen3 and Qwen3.5; 1.7B-14B students, 8B-32B teachers) and five agentic and code benchmarks, MAD-OPD ranks first across all six configurations; on the 14B+8B$\to$4B setting it lifts the agentic average by $+2.4\%$ and the code average by $+3.7\%$ over the stronger single-teacher OPD.
Read the full abstract

All authors Jianze Wang, Ying Liu, Jinlong Chen, Xuchun Hu, Qilong Zhang, Yu Cao, Jun Wang, Hua Yang, Yong Xie, Qianglong Chen

Computation and Language

A Multi-View Media Profiling Suite: Resources, Evaluation, and Analysis

Muhammad Arslan Manzoor, Dilshod Azizov, Daniil Orel et al. (+4 more)

News outlets shape public opinion at a scale that makes automated detection of political bias and factuality essential. However, the field still lacks unified resources, comprehensive evaluations across diverse approaches, and systematic analyses of the representations and fusion strategies that matter most, especially under label sparsity and dataset diversity. In addition, there is little empirical work reporting broad, observation-driven findings about what consistently works, what fails, and why. We address these gaps through four main contributions. First, we introduce MBFC-2025, a large-scale label set covering approximately 2,600 outlets from Media Bias/Fact Check (MBFC). Second, we construct multiview representations for ACL-2020 (Panayotov et al., 2022), which includes around 900 outlets, as well as for MBFC-2025. These representations span Alexa graphs, hyperlink graphs, LLM-derived graphs, articles, and Wikipedia descriptions. Third, we provide a systematic evaluation and analysis of embedding views and fusion strategies, including a reinforcement learning-based fusion variant. Fourth, we conduct extensive experiments that achieve state-of-the-art results on ACL-2020 and establish strong benchmarks on MBFC-2025.
Read the full abstract

All authors Muhammad Arslan Manzoor, Dilshod Azizov, Daniil Orel, Umer Siddique, Zain Muhammad Mujahid, Yufang Hou, Preslav Nakov

Computation and Language

OralMLLM-Bench: Evaluating Cognitive Capabilities of Multimodal Large Language Models in Dental Practice

Rongyang Wang, Shuang Zhou, Jiashuo Wang et al. (+2 more)

Multimodal large language models (MLLMs) have emerged as a promising paradigm for dental image analysis. However, their ability to capture the multi-level cognitive processes required for radiographic analysis remains unclear. Here, we present a comprehensive benchmark to evaluate the cognitive capabilities of MLLMs in dental radiographic analysis. It spans three critical imaging modalities, i.e., periapical, panoramic, and lateral cephalometric radiographs, and defines four cognitive categories: perception, comprehension, prediction, and decision-making. The benchmark comprises 27 clinically grounded tasks derived from public datasets, with manually curated annotations and 3,820 clinician assessments for evaluation. Six frontier MLLMs, including GPT-5.2 and GLM-4.6, are evaluated. We demonstrate the performance gap between MLLMs and clinicians in dental practice, delineate model strengths and limitations, characterize failure patterns, and provide recommendations for improvement. This data resource will facilitate the development of next-generation artificial intelligence systems aligned with clinical cognition, safety requirements, and workflow complexity in dental practice.
Read the full abstract

All authors Rongyang Wang, Shuang Zhou, Jiashuo Wang, Wenya Xie, Xiaoxia Che

Computation and Language

SiNFluD: Creating and Evaluating Figurative Language Dataset for Sindhi

Wazir Ali, Adeeb Noor and Saifullah Tumrani

In this article, we introduce SiNFluD, a novel benchmark dataset for Sindhi figurative language classification. We first collect raw text from various blogs, social media platforms, and literary sources, and subsequently prepare the corpus for annotation. Two native annotators label the data using the Doccano text annotation tool, achieving an inter-annotator agreement of 0.81. We then establish baseline results using 5-fold and 10-fold cross-validation. Finally, we evaluate mBERT, XLM-RoBERTa, and XLM-RoBERTa-XL models, along with SetFit for few-shot fine-tuning of sentence transformers. Among these, the pretrained XLM-RoBERTa-XL achieves the best performance.
Read the full abstract

Computation and Language

Benchmarking LightGBM and BiLSTM for Sentiment Analysis on Indonesian E-Commerce Reviews

Lidia Natasyah Marpaung, Vania Claresta, Iqfina Haula Halika et al. (+3 more)

This study presents a comparative analysis between two primary approaches in Natural Language Processing (NLP): Machine Learning (ML) utilizing the PyCaret AutoML framework, and Deep Learning (DL). The evaluation is conducted on a sentiment analysis task using an Indonesian e-commerce review dataset sourced from Hugging Face. The dataset, consisting of 15,000 samples, is partitioned into training, validation, and testing sets. The ML experiments compare LightGBM, Logistic Regression, and Support Vector Machine (SVM) algorithms, whereas the DL experiment implements a Bidirectional Long Short-Term Memory (BiLSTM) architecture. The experimental results demonstrate that the BiLSTM model outperforms all ML models, achieving an accuracy of 98.87% and an F1-Score of 98.87%. Meanwhile, LightGBM emerges as the best-performing ML model with an accuracy of 98.23% in a highly efficient training time. This research proves that the BiLSTM architecture is highly capable of capturing the sequential context of Indonesian review texts, making it the superior model for this specific classification task.
Read the full abstract

All authors Lidia Natasyah Marpaung, Vania Claresta, Iqfina Haula Halika, Luluk Muthoharoh, Ardika Satria, Martin Clinton Tosima Manullang

Computation and Language

Sentiment Analysis of Mobile Legends App Reviews Using Machine Learning and LSTM-Based Deep Learning Models

Vira Putri Maharani, Kharisa Harvanny, Daris Samudra et al. (+3 more)

This paper compares Machine Learning and LSTM-based Deep Learning methods for sentiment analysis of Mobile Legends app reviews. Using a dataset of 10,000 reviews labeled as positive, negative, and neutral, the study evaluates traditional models with TF-IDF and PyCaret AutoML and compares them against an LSTM model designed to capture sequential text dependencies. The results show that the LSTM model outperforms the classical Machine Learning baselines, achieving 92% accuracy and a weighted F1-score of 91%. The findings indicate that deep learning is more effective for handling informal and context-dependent user review text.
Read the full abstract

All authors Vira Putri Maharani, Kharisa Harvanny, Daris Samudra, Luluk Muthoharoh, Ardika Satria, Martin Clinton Tosima Manullang

2605.01317 · 2 May · 8 pages, 3 figures, includes comparative evaluation of Machine Learning and LSTM models for sentiment analysis on Indonesian Mobile Legends app reviews, with dataset description, methodology, model architecture, results, discussion, acknowledgments, and references

Abstract PDF $DATA ×1

Computation and Language

Enhancing Game Review Sentiment Classification on Steam Platform with Attention-Based BiLSTM

Abit Ahmad Oktarian, Fadhil Fitra Wijaya, Dhafin Razaqa Luthfi et al. (+3 more)

This paper investigates sentiment classification of Steam game reviews using an attention-based Bidirectional Long Short-Term Memory (BiLSTM) model. Using a dataset of 50,000 reviews sampled from a larger Steam review corpus, the authors compare a traditional machine learning baseline based on TF-IDF and PyCaret AutoML with a deep learning approach implemented in PyTorch. The proposed BiLSTM+Attention model is trained with class-weighted cross-entropy to address class imbalance and achieves 83% accuracy and 85% weighted F1-score on the test set, with 90% recall for negative reviews. The paper also presents attention visualizations to show interpretability by highlighting sentiment-bearing words. The study concludes that the BiLSTM+Attention model is effective for analyzing user sentiment in Steam reviews and useful for helping developers understand player feedback.
Read the full abstract

All authors Abit Ahmad Oktarian, Fadhil Fitra Wijaya, Dhafin Razaqa Luthfi, Luluk Muthoharoh, Ardika Satria, Martin Clinton Tosima Manullang

Computation and Language

Beyond Semantic Relevance: Counterfactual Risk Minimization for Robust Retrieval-Augmented Generation

Peiyang Liu, Qiang Yan, Ziqiang Cui et al. (+3 more)

Standard Retrieval-Augmented Generation (RAG) systems predominantly rely on semantic relevance as a proxy for utility. However, this assumption collapses in realistic decision-making scenarios where user queries are laden with cognitive biases, such as false premises or confirmation bias. In such cases, maximizing relevance paradoxically promotes the retrieval of sycophantic evidence that reinforces hallucinations, a critical failure we term the “Relevance-Robustness Gap”. To bridge this gap, we propose CoRM-RAG (Counterfactual Risk Minimization for RAG), a framework that aligns retrieval with decision safety rather than mere similarity. Grounded in causal intervention, we introduce a Cognitive Perturbation Protocol to simulate user biases during training, which is then distilled into a lightweight Evidence Critic. This scoring module learns to identify documents that possess sufficient evidential strength to steer the model toward correctness despite adversarial query perturbations. Extensive experiments on decision-making benchmarks demonstrate that CoRM-RAG significantly outperforms strong dense retrievers and LLM-based rerankers in adversarial settings, while enabling effective risk-aware abstention through reliable robustness scoring. Our code is available at https://github.com/PeiYangLiu/CoRM-RAG.git.
Read the full abstract

All authors Peiyang Liu, Qiang Yan, Ziqiang Cui, Di Liang, Xi Wang, Wei Ye

Computation and Language

Addressing Data Scarcity in Bangla Fake News Detection: An LLM-Based Dataset Augmentation Approach

Ahmed Alfey Sani, Kazi Akib Zaoad, Shefayat E Shams Adib et al. (+2 more)

The growing spread of misinformation in digital media highlights the need for reliable fake news detection systems, yet progress in under-resourced languages such as Bangla is limited by small and imbalanced datasets. This study investigates whether Large Language Model (LLM) based augmentation can effectively address this limitation and improve Bangla fake news classification. Existing datasets remain valuable but highly imbalanced, limiting model performance, and LLM based augmentation for Bangla has been scarcely explored. To fill this gap, we propose a systematic augmentation framework that generates synthetic Bangla news articles using the instruction tuned Gemma 3 27B IT model, supported by semantic filtering and controlled subsampling to preserve label consistency and diversity. We compare zero shot and few shot prompting, evaluate multiple augmentation rates, and examine random versus similarity-based selection strategies. Our experiments show that augmenting only the minority class with a high augmentation rate and random subsampling yields the strongest gains, raising the Fake News F1 score from 0.85 to 0.88. To support reproducibility and further research in this low-resource domain, we publicly release 4,545 synthetically generated Bangla fake news samples along with our full implementation. These findings demonstrate that well-designed LLM-driven augmentation can significantly improve fake news detection in low resource settings and provide a practical foundation for advancing multilingual misinformation research.
Read the full abstract

All authors Ahmed Alfey Sani, Kazi Akib Zaoad, Shefayat E Shams Adib, Md Abdul Muqtadir, Ajwad Abrar

Computation and Language · Cross-listed from cs.CV

Chain of Evidence: Pixel-Level Visual Attribution for Iterative Retrieval-Augmented Generation

Peiyang Liu, Ziqiang Cui, Xi Wang et al. (+2 more)

Iterative Retrieval-Augmented Generation (iRAG) has emerged as a powerful paradigm for answering complex multi-hop questions by progressively retrieving and reasoning over external documents. However, current systems predominantly operate on parsed text, which creates two critical bottlenecks: (1) Coarse-grained attribution, where users are burdened with manually locating evidence within lengthy documents based on vague text-level citations; and (2) Visual semantic loss, where the conversion of visually rich documents (e.g., slides, PDFs with charts) into text discards spatial logic and layout cues essential for reasoning. To bridge this gap, we present Chain of Evidence (CoE), a retriever-agnostic visual attribution framework that leverages Vision-Language Models to reason directly over screenshots of retrieved document candidates. CoE eliminates format-specific parsing and outputs precise bounding boxes, visualizing the complete reasoning chain within the retrieved candidate set. We evaluate CoE on two distinct benchmarks: Wiki-CoE, a large-scale dataset of structured web pages derived from 2WikiMultiHopQA, and SlideVQA, a challenging dataset of presentation slides featuring complex diagrams and free-form layouts. Experiments demonstrate that fine-tuned Qwen3-VL-8B-Instruct achieves robust performance, significantly outperforming text-based baselines in scenarios requiring visual layout understanding, while establishing a retriever-agnostic solution for pixel-level interpretable iRAG. Our code is available at https://github.com/PeiYangLiu/CoE.git.
Read the full abstract

All authors Peiyang Liu, Ziqiang Cui, Xi Wang, Di Liang, Wei Ye

Computation and Language

GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models

Zhiwen Ruan, Yichao Du, Jianjie Zheng et al. (+6 more)

A promising paradigm for adapting instruction-tuned language models is to learn task-specific updates on a pretrained base model and subsequently merge them into the instruction-tuned model. However, existing approaches typically treat the instruction-tuned model as a passive target that is only involved at the final merging stage, without guiding the training process. We propose GIFT (Guided Fine-Tuning and Transfer), a simple and efficient framework that incorporates guidance from the instruction model into task adaptation. GIFT fine-tunes a low-rank adapter on the pretrained base model using confidence signals derived from the instruction-tuned model. The learned adapter is then merged into the instruction-tuned model, yielding task-specialized models that preserve general instruction-following behavior. We evaluate GIFT on mathematical and knowledge-intensive benchmarks across multiple model families and scales. Results show that GIFT consistently outperforms direct fine-tuning and representative transfer-based baselines, while maintaining robust generalization and favorable test-time scaling behavior.
Read the full abstract

All authors Zhiwen Ruan, Yichao Du, Jianjie Zheng, Longyue Wang, Yun Chen, Peng Li, Jinsong Su, Yang Liu, Guanhua Chen

Computation and Language · Cross-listed from cs.LG

Attention Sinks in Massively Multilingual Neural Machine Translation:Discovery, Analysis, and Mitigation

Hillary Mutisya and John Mugane

Cross-attention patterns in neural machine translation (NMT) are widely used to study how multilingual models align linguistic structure. We report a systematic artifact in cross-attention analysis of NLLB-200 (600M): non-content tokens - primarily end-of-sequence tokens, language tags, and punctuation - capture 83 percent to 91 percent of total cross-attention mass. We term these "attention sinks," extending findings from LLMs [Xiao et al., 2023] to NMT cross-attention and identifying a causal mechanism rooted in vocabulary design rather than position bias. This artifact causes raw metrics to underestimate content-level similarity by nearly half (36.7 percent raw vs. 70.7 percent filtered), rendering uncorrected analyses unreliable. To address this, we validate a content-only filtering methodology that removes non-content tokens and renormalizes the distribution. Applying this to 1,000 parallel sentences across African languages (Swahili, Kikuyu, Somali, Luo) and non-African benchmarks (German, Turkish, Chinese, Hindi), we confirm the artifact is universal and recover masked linguistic signals: a 16.9 percentage-point gap between teacher-forcing and generation modes, clear language-family clustering in attention entropy, and a hidden Somali paradox linking SOV word order to monotonic alignment. We release our filtering toolkit and corrected datasets to support reproducible interpretability research on multilingual NMT.
Read the full abstract

Computation and Language

Lost in the Tower of Babel: The Adverse Effects of Incidental Multilingualism in LLMs

Anjishnu Mukherjee, Chutong Meng and Antonios Anastasopoulos

This paper argues that contemporary multilingual NLP has converged on a fragile and misleading paradigm of incidental multilingualism. Today's LLMs appear multilingual largely because they are trained on massive, uneven web corpora, not because multilingual or multicultural competence has been treated as a core design objective. We contend that this paradigm systematically produces unequal, brittle, and opaque behavior across languages, with severe consequences in real-world and agentic deployments where models must reason, plan, and act across multiple linguistic contexts. We report a focused empirical study of two practical questions: which languages models self-report as supported and which languages they actually respond in across multilingual prompts. We additionally demonstrate how even a simple language-change attack can surface these failures and expose hidden assumptions about language in LLM-based systems. To address this, we call for a shift toward multilingualism by design: a research agenda that treats equitable multilingual performance, cultural grounding, and cross-lingual behavioral understanding as first-class goals in all aspects of the model pipeline.
Read the full abstract

Computation and Language

SRA: Span Representation Alignment for Large Language Model Distillation

Quoc Phong Dao, Hoang Son Nguyen, Pham Khanh Chi et al. (+4 more)

Cross-Tokenizer Knowledge Distillation (CTKD) enables knowledge transfer between a large language model and a smaller student, even when they employ different tokenizers. While existing approaches mainly focus on token-level alignment strategies, which are often brittle and sensitive to discrepancies between tokenizers, we argue that the method of aggregating tokens into more robust representations before distillation is of equal importance. In this paper, we introduce SRA (Span Representation Alignment for Large Language Model Distillation), a novel framework that reframes CTKD through the physical lens of Multi-Particle Dynamical Systems. SRA shifts the fundamental unit of alignment from tokens to robust, tokenizer-agnostic spans. We model each span as a cluster of particles and represent its state by its Center of Mass (CoM) - an attention-weighted average that captures rich semantic information. We leverage the concept of span centers of mass with attention-derived weighting to prioritize the most salient spans. In addition, we employ a geometric regularizer to preserve the structural integrity of the representation space and introduce aligned span logit distillation to enhance knowledge transfer across models. In challenging cross-architecture distillation experiments, SRA consistently and significantly outperforms state-of-the-art CTKD baselines, validating our physically-grounded approach.
Read the full abstract

All authors Quoc Phong Dao, Hoang Son Nguyen, Pham Khanh Chi, Tung Nguyen, Linh Ngo Van, Nguyen Thi Ngoc Diep, Trung Le

Computation and Language · Cross-listed from cs.AI

LSR-Ben: A Logical and Scientific Reasoning Benchmark for Evaluating Process Reward Models

Zhouhao Sun, Xuan Zhang, Xiao Ding et al. (+10 more)

Currently, process reward models (PRMs) have exhibited remarkable potential for test-time scaling. Since large language models (LLMs) regularly generate flawed intermediate reasoning steps when tackling a broad spectrum of reasoning and decision-making tasks, PRMs are required to possess capabilities for detecting process-level errors in real-world scenarios. However, existing benchmarks primarily focus on mathematical reasoning, thereby failing to comprehensively evaluate the error detection ability of PRMs across diverse reasoning scenarios. To mitigate this gap, we introduce LSR-Ben, a process-level benchmark specifically designed for assessing PRM's performance across two primary reasoning domains (scientific and logical reasoning) and nine subdomains. We conduct extensive experiments on a diverse set of 22 models, encompassing both PRMs and LLMs, and derive two key findings: (1) In domains beyond mathematical reasoning, the error-detection ability of existing PRMs and LLMs is found to be markedly weaker by comparison. (2) In general, LLMs exhibit a tendency toward over-identification of errors compared to PRMs, whereas PRMs exhibit an inherent tendency to overlook errors compared to LLMs. We hope LSR-Ben can foster future researches on PRMs for broader domains, thereby enhancing the reasoning capabilities of LLMs.
Read the full abstract

All authors Zhouhao Sun, Xuan Zhang, Xiao Ding, Bibo Cai, Li Du, Kai Xiong, Xinran Dai, Fei Zhang, weidi tang, Zhiyuan Kan, Yang Zhao, Bing Qin, Ting Liu

Computation and Language

Compute Optimal Tokenization

Tomasz Limisiewicz, Artidoro Pagnoni, Srini Iyer et al. (+6 more)

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.
Read the full abstract

All authors Tomasz Limisiewicz, Artidoro Pagnoni, Srini Iyer, Mike Lewis, Sachin Mehta, Alisa Liu, Margaret Li, Gargi Ghosh, Luke Zettlemoyer

Friday 1 May

Computation and Language

Just Keep Prompting: Evaluating Repetitive Socratic Prompting in VLMs

Shayda Moezzi, Bishoy Galoaa, Lorena Genua et al. (+2 more)

Deploying Vision-Language Models (VLMs) in real-world settings requires not only strong visual reasoning but also stability under sustained conversational pressure. We introduce Just Keep Prompting (JKP), a multi-turn evaluation framework that measures VLM epistemic stability when users repeatedly challenge, question, or contradict a model's answer. JKP probes models for up to 10 follow-up turns using three strategies: Adversarial Negation (repeated rejection), Pure Socratic Interrogation (repeated calls to reassess certainty), and Context-Aware Socratic Summarization (reflecting the model's prior rationale back before asking for reconsideration). We evaluate GPT-4o, Gemini 2.5 Pro, and Qwen3-VL-30B on a subset of the STAR benchmark across 720 multi-turn runs. Aggregate accuracy changes modestly from Turn 0 to Turn 10, but trajectory-level analysis reveals substantial instability: correct answers regress, wrong answers recover, and many runs exhibit repeated answer flipping. Repeated prompting has bounded upside and often acts as a destabilizer rather than a reasoning aid. The effect is strongly model-dependent: Qwen3-VL-30B achieves the highest final accuracy but becomes confidently wrong under direct contradiction; Gemini 2.5 Pro is comparatively stable but token-expensive; GPT-4o is the most brittle and oscillatory. These findings reveal that multi-turn VLM evaluation captures not just additional reasoning but pressure-response profiles: how models trade off visual grounding, calibration, and conversational compliance under repeated challenge.
Read the full abstract

All authors Shayda Moezzi, Bishoy Galoaa, Lorena Genua, Taskin Padir, Sarah Ostadabbas

Computation and Language

Breaking Safety at the Token Boundary: How BPE Tokenization Creates Exploitable Gaps in LLM Alignment

Tung-Ling Li, Hongliang Liu and Yuhao Wu

Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B). An optimization targeting safety-token fragmentation flips the first-token refusal trigger on 80-100% of refused HarmBench prompts, with 48% of those flips producing genuinely harmful outputs (per-model 29-65%; gap-vs-behavior ROC-AUC 0.66-0.98, pooled 0.84). Activation patching localizes the disrupted signal to the last ${\sim}30\%$ of layers; an alignment-data scan finds zero fragmented prompts among 30,000 examples (positive-control recall $\geq 99\%$ at attack-relevant intensities); and targeted-mutation experiments isolate safety words as the disruption locus. On the defense side, a 68-cell grid (55 trained checkpoints) shows that no DPO configuration achieves seed- and pool-stable ASR closure on the three families with closed pool-size confounds. SFT trained on fragmented prompts closes ASR on 3/5 families but only via global collapse that raises refusal on benign prompts as well, indicating the missing distribution is necessary but not sufficient under the LoRA-16 recipe we tested. To distinguish selective repair from global collapse, we introduce Conv-Benign, a candidate paired diagnostic. All ASR claims are 3-judge-calibrated (cell rankings stable across judges; absolute levels $\pm$18pp; see App.~B.13).
Read the full abstract

Computation and Language

SPARCLE: SPeaker-aware Aligned Representations via Contrastive Language Embeddings

Priyam Mazumdar, Yurii Halychanskyi, Steven Guo et al. (+2 more)

Recent advances in speech synthesis have shifted from phoneme representations to direct grapheme modeling. While phonemes address the one-to-many mapping between text and acoustics, they rely on grapheme-to-phoneme (G2P) systems that fail to capture speaker-specific acoustic variation. Prior work demonstrates that grapheme-based models outperform phoneme-based systems at scale, but not in low-resource settings. In this paper, we propose SPARCLE, a speaker-aware grapheme representation model that enriches characters with their precise acoustic realizations. SPARCLE is trained with a contrastive objective to align graphemes with corresponding Wav2Vec2 acoustic representations while conditioned on speaker identity. The resulting model serves as a replacement to G2P systems for downstream text-to-speech (TTS) tasks. We demonstrate that SPARCLE improves generation quality, reducing word error rates by half in extreme low-resource settings compared to standard grapheme-based models.
Read the full abstract

All authors Priyam Mazumdar, Yurii Halychanskyi, Steven Guo, Mark Hasegawa-Johnson, Volodymyr Kindratenko

Computation and Language

KARA: Efficient Reasoning LLM Serving via Sliding-Window KV Cache Compression

Shen Han, Yuyang Wu, Junpu Yu et al. (+1 more)

Reasoning language models often generate long chain-of-thought (CoT), which accumulates a massive KV cache during the decoding phase and incurs high decoding latency and limited throughput. To address these issues, KV cache compression has emerged as a promising technique for reducing memory overhead by selectively removing unimportant KV pairs while preserving useful ones for subsequent decoding. Nevertheless, we identify two key limitations in existing KV cache compression methods: 1) their threshold-triggered compression policy may provide limited throughput improvement or even reduce throughput, and may fully eliminate KV pairs from certain blocks of the sequence, potentially worsening information loss. 2) they typically retain either isolated KV pairs or fixed-size chunks with rigid boundaries, failing to preserve important flexible-sized chunks at arbitrary token positions. To overcome these limitations, we propose Kara, a sliding-window KV cache compression method that performs decoding-time compression by operating only on the recently generated context. Kara leverages bidirectional attention to score and select informative KV pairs in the window. To enable flexible preservation of important semantic information, we design a Token2Chunk module to expand a subset of selected KV pairs into chunks. Furthermore, we adapt Kara to PagedAttention and develop KvLLM, an inference framework built upon vLLM, which reduces KV cache memory usage and effectively improves output throughput. Extensive experiments demonstrate consistent performance improvements of proposed Kara and KvLLM.
Read the full abstract

All authors Shen Han, Yuyang Wu, Junpu Yu, Olexandr Isayev

Computation and Language

Safeguarding LLM Agents from Misalignment through Provenance Analysis

Yining She, Yiliang Liang and Eunsuk Kang

As LLM agents gain increasing access to powerful tools, ensuring that their actions align with the user's intent becomes critical. When an agent's proposed action deviates from that intent—a phenomenon called misalignment—it may cause harm that is difficult to undo. Existing runtime guardrails rely on an LLM-as-a-judge paradigm that lacks a systematic framework for reasoning about alignment, often producing inconsistent or difficult-to-audit judgments. Motivated by provenance analysis, we propose a conceptual framework that formalizes misalignment detection as determining whether a proposed tool call is supported by traceable evidence in the agent's context. Based on this framework, we build ProvenanceGuard, a multi-stage pipeline that analyzes the agent's action for three types of misalignment before its execution and only allows aligned actions. We evaluated ProvenanceGuard on AgentSafetyBench and WorkBench, across 11 backbone LLMs. Compared to the LLM-as-a-judge baseline, ProvenanceGuard reduces error rate on misaligned traces from 44.3% to 2.1% on Agent-SafetyBench and from 32.4% to 18.7% on WorkBench, while reducing interventions on task-successful traces from 31.2% to 13.0% and introducing no statistically significant increase in unnecessary interventions on aligned traces. These results demonstrate that structured, provenance-based reasoning provides an effective and practical foundation for safeguarding LLM agents from misalignment.
Read the full abstract

Computation and Language

Persona Without Substrate: Regime-Dependence and the LLM Individuation Problem

Shuaizhi Cheng

Beckmann & Butlin's (2026) ontological framework for the LLM individuation problem inherits an unargued cross-regime co-reference assumption from the persona-vectors literature: that the same direction picks out the same content under prompt-conditioning, gradient-descent fine-tuning, and inference-time steering. We present four empirical wedges from persona-topology experiments on Qwen3-4B-Instruct and Mistral-7B-Instruct-v0.2 - non-collinearity of prompt-extracted vectors and fine-tune basins; fictional personas displacing the model along real-anchor directions more strongly than real anchors do; contradictory-valenced mixtures biased toward a training-history-determined attractor; and asymmetric compositional algebra under inference-time arithmetic versus fine-tune-time chimera training - that jointly undermine the assumption. We propose regime-indexed individuation: the identity unit for representational content is a (vehicle, regime) pair, not a vehicle alone. Under this framework, Beckmann & Butlin's three candidate positions describe three different regime-internal objects rather than competing for the same referent; the same diagnosis applies to Mollo & Millière, Chalmers, and Cerullo.
Read the full abstract

Computation and Language

Reducing Conversational Escalation in Large Language Model Dialogue with Nonviolent Communication Constraints

Zhixing Sun, Shenghe Xu and Tao Li

Large language models (LLMs) are increasingly used in emotionally charged situations involving interpersonal conflict, frustration, and distress. While prior safety research has focused on preventing explicit harms such as toxic or policy-violating content, less attention has been paid to conversational behaviors that may unintentionally escalate conflict. In this paper, we investigate whether LLMs can be guided toward more de-escalating dialogue behavior through lightweight prompt-level constraints derived from Nonviolent Communication (NVC). We reformulate NVC principles as process-oriented guidelines that discourage blame attribution, emphasize attention to users' emotional experiences, and encourage clarification before advice. Using a dual-agent simulation framework across multiple instruction-tuned models and user resistance levels, we show that NVC-constrained prompting consistently reduces conversational escalation and stabilizes interactions with highly resistant users. These results suggest that simple communication constraints can meaningfully improve the trustworthiness of LLM dialogue in conflict-prone settings.
Read the full abstract