Papers

Masked Diffusion Language Models (MDLMs) have emerged as a compelling non-autoregressive alternative to standard large language models; however, their application to morphologically rich languages remains limited. In this paper, we introduce $\textit{Diffutron}$, a masked diffusion language model specifically designed for Turkish. Our approach leverages a resource-efficient training pipeline, starting with LoRA-based continual pre-training of a multilingual encoder on a large-scale corpus. To enable generative capabilities, we employ a progressive instruction-tuning strategy, sequentially adapting the model on general and task-specific instruction sets. Experimental results across comprehensive benchmarks demonstrate that, despite its compact size, our model achieves competitive performance compared to existing multi-billion-parameter baselines. These findings validate the effectiveness of masked diffusion modeling combined with multi-stage tuning for non-autoregressive text generation in Turkish.

A number of scientific conferences and journals have recently enacted policies that prohibit LLM usage by peer reviewers, except for polishing, paraphrasing, and grammar correction of otherwise human-written reviews. But, are these policies enforceable? To answer this question, we assemble a dataset of peer reviews simulating multiple levels of human-AI collaboration, and evaluate five state-of-the-art detectors, including two commercial systems. Our analysis shows that all detectors misclassify a non-trivial fraction of LLM-polished reviews as AI-generated, thereby risking false accusations of academic misconduct. We further investigate whether peer-review-specific signals, including access to the paper manuscript and the constrained domain of scientific writing, can be leveraged to improve detection. While incorporating such signals yields measurable gains in some settings, we identify limitations in each approach and find that none meets the accuracy standards required for identifying AI use in peer reviews. Importantly, our results suggest that recent public estimates of AI use in peer reviews through the use of AI-text detectors should be interpreted with caution, as current detectors misclassify mixed reviews (collaborative human-AI outputs) as fully AI generated, potentially overstating the extent of policy violations.

Verification-guided self-improvement has recently emerged as a promising approach to improving the accuracy of large language model (LLM) outputs. However, existing approaches face a trade-off between inference efficiency and accuracy: iterative verification-rectification is computationally expensive and prone to being trapped in faulty reasoning, while best-of-N selection requires extensive sampling without addressing internal model flaws. We propose a training-free regeneration paradigm that leverages an offline-curated contrastive Reflection Memory (RM) to provide corrective guidance, while regenerating from scratch helps break out of faulty reasoning. At inference time, the method performs RM-guided self-verification followed by a single RM-guided regeneration, avoiding both iterative correction and multi-sample selection. We evaluated our method on nine benchmarks that span algorithmic, reasoning, symbolic, and domain-specific tasks in both small- and large-scale LLMs. Experiment results show that our method outperforms prior methods while maintaining low computational cost.

While Large Audio-Language Models (LALMs) have been shown to exhibit degraded instruction-following capabilities, their ability to infer task patterns from in-context examples with audio remains understudied. To address this gap, we design a three-stage evaluation pipeline that progressively reduces textual guidance to systematically evaluate LALMs' in-context learning ability in the audio modality. Evaluating six LALMs across four audio understanding tasks under two output constraint categories, we uncover a consistent asymmetry across LALMs: in-context demonstrations reliably improve format compliance but fail to improve the core task performance. This suggests that LALMs can glean surface-level formatting patterns from demonstrations but may struggle to leverage cross-modal semantic grounding to reliably infer task objectives from examples with audio, highlighting potential limitations in current cross-modal integration. We further probe how demonstrations are used through two complementary analyses, demonstration label shuffling and attention knockout on demonstration spans, both showing that LALMs leverage in-context examples primarily to establish the output label space and format rather than to learn a meaningful input-output correspondence.

Large Language Models (LLMs) have demonstrated remarkable progress in scaling to access massive contexts. However, the access is via the latent and uninterpretable attention mechanisms, and LLMs fail to effective process long context, exhibiting significant performance degradation as context length increases. In this work, we study whether long-context processing can be externalized from latent attention into explicit, executable interactions, by allowing coding agents to organize text in file systems and manipulate it using its native tools. We evaluate off-the-shelf frontier coding agents as the general interface for tasks that require processing long contexts, including long-context reasoning, retrieval-augmented generation, and open-domain question answering with large-scale corpus contains up to three trillion tokens. Across multiple benchmarks, these agents outperform published state-of-the-art by 17.3% on average. We attribute this efficacy to two key factors: native tool proficiency, which enables agents to leverage executable code and terminal commands rather than passive semantic queries, and file system familiarity, which allows them to navigate massive text corpora as directory structures. These findings suggest that delegating long-context processing to coding agents offers an effective alternative to semantic search or context window scaling, opening new directions for long-context processing in LLMs.

We report on an experiment in which Claude Opus~4.6, equipped with a suite of Model Context Protocol (MCP) tools for the Rocq proof assistant, autonomously proved 10 of 12 problems from the 2025 Putnam Mathematical Competition. The MCP tools, designed with Claude by analyzing logs from a prior experiment on miniF2F-Rocq, encode a "compile-first, interactive-fallback" strategy. Running on an isolated VM with no internet access, the agent deployed 141 subagents over 17.7 hours of active compute (51.6h wall-clock), consuming approximately 1.9 billion tokens. All proofs are publicly available.

Understanding the fundamental mechanisms governing the production of meaning in the processing of natural language is critical for designing safe, thoughtful, engaging, and empowering human-agent interactions. If meaning is constituted rather than retrieved, then the search for context-independent features or circuits in the pursuit of mechanistic interpretability may be fundamentally limited. Experiments in cognitive science and social psychology have demonstrated that human semantic processing exhibits contextuality more consistent with quantum logical mechanisms than classical Boolean theories, and recent works have found similar results in large language models—in particular, clear violations of the Bell inequality in experiments of contextuality during interpretation of ambiguous expressions. In this work, we explore the CHSH $|S|$ parameter—the metric associated with the inequality—across the inference parameter space of models spanning four orders of magnitude in scale and cross-reference our findings with MMLU, hallucination rate, and nonsense detection benchmarks. We find that the interquartile range of the $|S|$ distribution is completely orthogonal to all external benchmarks, while overall violation rate shows weak anticorrelation with all three benchmarks. We investigate how $|S|$ varies with sampling parameters and word order, and discuss the information-theoretic constraints that genuine contextuality imposes on prompt injection defenses and its human analogue, whereby careful construction and maintenance of social contextuality can be carried out at scale, shaping the space of possible interpretations before any particular one is reached. We consider the implications for mechanistic interpretability and how genuine contextuality sets an information-theoretic bound on the decomposability of semantic processing.

Video agentic models have advanced challenging video-language tasks. However, most agentic approaches still heavily rely on greedy parsing over densely sampled video frames, resulting in high computational cost. We present VideoSeek, a long-horizon video agent that leverages video logic flow to actively seek answer-critical evidence instead of exhaustively parsing the full video. This insight allows the model to use far fewer frames while maintaining, or even improving, its video understanding capability. VideoSeek operates in a think-act-observe loop with a well-designed toolkit for collecting multi-granular video observations. This design enables query-aware exploration over accumulated observations and supports practical video understanding and reasoning. Experiments on four challenging video understanding and reasoning benchmarks demonstrate that VideoSeek achieves strong accuracy while using far fewer frames than prior video agents and standalone LMMs. Notably, VideoSeek achieves a 10.2 absolute points improvement on LVBench over its base model, GPT-5, while using 93% fewer frames. Further analysis highlights the significance of leveraging video logic flow, strong reasoning capability, and the complementary roles of toolkit design.

Large vision–language models (VLMs) are increasingly applied to long-video question answering, yet inference is often bottlenecked by the number of input frames and resulting visual tokens. Naive sparse sampling can miss decisive moments, while purely relevance-driven selection frequently collapses onto near-duplicate frames and sacrifices coverage of temporally distant evidence. We propose a question-adaptive greedy frame selection method that jointly optimizes query relevance and semantic representativeness under a fixed frame budget. Our approach constructs a 1~FPS candidate pool (capped at 1000) with exact timestamp alignment, embeds candidates in two complementary spaces (SigLIP for question relevance and DINOv2 for semantic similarity), and selects frames by greedily maximizing a weighted sum of a modular relevance term and a facility-location coverage term. This objective is normalized, monotone, and submodular, yielding a standard (1-1/e) greedy approximation guarantee. To account for question-dependent trade-offs between relevance and coverage, we introduce four preset strategies and a lightweight text-only question-type classifier that routes each query to its best-performing preset. Experiments on MLVU show consistent accuracy gains over uniform sampling and a strong recent baseline across frame budgets, with the largest improvements under tight budgets.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper provides evidence that it is not. Three classifiers (a regex-only detector, a regex-plus-LLM pipeline, and a Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce faithfulness rates of 74.4%, 82.6%, and 69.7%. Per-model gaps range from 2.6 to 30.6 percentage points; all pairwise McNemar tests are significant (p < 0.001). The disagreements are systematic: Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under Sonnet; OLMo-3.1-32B moves from 9th to 3rd. Different classifiers operationalize faithfulness at different levels of stringency (lexical mention versus epistemic dependence), yielding divergent measurements on the same behavior. These results indicate that published faithfulness numbers cannot be meaningfully compared across studies using different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies.

In contested domains, instruction-tuned language models must balance user-alignment pressures against faithfulness to the in-context evidence. To evaluate this tension, we introduce a controlled epistemic-conflict framework grounded in the U.S. National Climate Assessment. We conduct fine-grained ablations over evidence composition and uncertainty cues across 19 instruction-tuned models spanning 0.27B to 32B parameters. Across neutral prompts, richer evidence generally improves evidence-consistent accuracy and ordinal scoring performance. Under user pressure, however, evidence does not reliably prevent user-aligned reversals in this controlled fixed-evidence setting. We report three primary failure modes. First, we identify a negative partial-evidence interaction, where adding epistemic nuance, specifically research gaps, is associated with increased susceptibility to sycophancy in families like Llama-3 and Gemma-3. Second, robustness scales non-monotonically: within some families, certain low-to-mid scale models are especially sensitive to adversarial user pressure. Third, models differ in distributional concentration under conflict: some instruction-tuned models maintain sharply peaked ordinal distributions under pressure, while others are substantially more dispersed; in scale-matched Qwen comparisons, reasoning-distilled variants (DeepSeek-R1-Qwen) exhibit consistently higher dispersion than their instruction-tuned counterparts. These findings suggest that, in a controlled fixed-evidence setting, providing richer in-context evidence alone offers no guarantee against user pressure without explicit training for epistemic integrity.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, the truthfulness of their outputs is not guaranteed, and their tendency toward overconfidence further limits reliability. Uncertainty quantification offers a promising way to identify potentially unreliable outputs, but most existing methods rely on repeated sampling or auxiliary models, introducing substantial computational overhead. To address these limitations, we propose Semantic Token Clustering (STC), an efficient uncertainty quantification method that leverages the semantic information inherently encoded in LLMs. Specifically, we group tokens into semantically consistent clusters using embedding clustering and prefix matching, and quantify uncertainty based on the probability mass aggregated over the corresponding semantic cluster. Our approach requires only a single generation and does not depend on auxiliary models. Experimental results show that STC achieves performance comparable to state-of-the-art baselines while substantially reducing computational overhead.

The Hyperspace Analogue to Language (HAL) model relies on global word co-occurrence matrices to construct distributional semantic representations. While these representations capture lexical relationships effectively, aggregating them into sentence-level embeddings via standard mean pooling often results in information loss. Mean pooling assigns equal weight to all tokens, thereby diluting the impact of contextually salient words with uninformative structural tokens. In this paper, we address this limitation by integrating a learnable, temperature-scaled additive attention mechanism into the HAL representation pipeline. To mitigate the sparsity and high dimensionality of the raw co-occurrence matrices, we apply Truncated Singular Value Decomposition (SVD) to project the vectors into a dense latent space prior to the attention layer. We evaluate the proposed architecture on the IMDB sentiment analysis dataset. Empirical results demonstrate that the attention-based pooling approach achieves a test accuracy of 82.38%, yielding an absolute improvement of 6.74 percentage points over the traditional mean pooling baseline (75.64%). Furthermore, qualitative analysis of the attention weights indicates that the mechanism successfully suppresses stop-words and selectively attends to sentiment-bearing tokens, improving both classification performance and model interpretability.

Large Language Models (LLMs) achieve strong performance on many reasoning benchmarks, yet these evaluations typically focus on isolated tasks that differ from real-world usage in task-oriented dialogue (TOD). In this setting, LLMs must perform reasoning inherently while generating text and adhering to instructions on role, format, and style. This mismatch raises concerns about whether benchmark performance accurately reflects models' reasoning robustness in TOD setting. We investigate how framing reasoning tasks within TOD affects LLM performance by introducing BOULDER, a new dynamic benchmark covering eight travel-related tasks that require arithmetic, spatial, and temporal reasoning with both commonsense and formal aspects. Each problem is presented in both isolated and dialogue-based variants, enabling controlled comparison while mitigating data contamination. Experiments on eight LLMs reveal a substantial and consistent performance gap between isolated and dialogue settings. Through ablations and qualitative analysis, we show that this gap is largely driven by the multi-turn nature of dialogue, with additional effects from role conditioning and tool-use requirements. Our results highlight the need to evaluate LLM reasoning in realistic interactive scenarios.

We aim to examine the extent to which Large Language Models (LLMs) can 'talk much' about grammar modules, providing evidence from syntax core properties translated by ChatGPT into Arabic. We collected 44 terms from generative syntax previous works, including books and journal articles, as well as from our experience in the field. These terms were translated by humans, and then by ChatGPT-5. We then analyzed and compared both translations. We used an analytical and comparative approach in our analysis. Findings unveil that LLMs still cannot 'talk much' about the core syntax properties embedded in the terms under study involving several syntactic and semantic challenges: only 25% of ChatGPT translations were accurate, while 38.6% were inaccurate, and 36.4.% were partially correct, which we consider appropriate. Based on these findings, a set of actionable strategies were proposed, the most notable of which is a close collaboration between AI specialists and linguists to better LLMs' working mechanism for accurate or at least appropriate translation.

Direct Preference Optimization (DPO) is widely used after supervised fine-tuning (SFT) to align language models, yet empirical behavior under small backbones and modest data is under-specified. We systematically compare SFT-only, DPO-only, and staged SFT-to-DPO training alongside full fine-tuning (FFT) versus LoRA on a GPT-2-scale decoder, evaluating paraphrase detection and Shakespearean sonnet continuation. DPO yields small, task-dependent gains over strong SFT and can match competitive SFT accuracy without a warm start when the preference construction closely parallels the supervised objective. In contrast, parameterization dominates: FFT consistently outperforms LoRA at matched training depth, and LoRA does not reduce wall-clock time on our hardware. These findings indicate that, in this small-scale regime, supervised full-parameter adaptation remains the primary performance lever, while preference optimization and low-rank adaptation provide limited marginal returns.

We investigate how verbal and nonverbal linguistic features, exhibited by speakers and listeners in dialogue, can contribute to predicting the listener's state of understanding in explanatory interactions on a moment-by-moment basis. Specifically, we examine three linguistic cues related to cognitive load and hypothesised to correlate with listener understanding: the information value (operationalised with surprisal) and syntactic complexity of the speaker's utterances, and the variation in the listener's interactive gaze behaviour. Based on statistical analyses of the MUNDEX corpus of face-to-face dialogic board game explanations, we find that individual cues vary with the listener's level of understanding. Listener states ('Understanding', 'Partial Understanding', 'Non-Understanding' and 'Misunderstanding') were self-annotated by the listeners using a retrospective video-recall method. The results of a subsequent classification experiment, involving two off-the-shelf classifiers and a fine-tuned German BERT-based multimodal classifier, demonstrate that prediction of these four states of understanding is generally possible and improves when the three linguistic cues are considered alongside textual features.

Large language models (LLMs) have driven substantial advances in speech language models (SpeechLMs), yielding strong performance in automatic speech recognition (ASR) under high-resource conditions. However, existing benchmarks predominantly focus on high-resource languages, leaving the ASR behavior of SpeechLMs in low-resource languages insufficiently understood. This gap is critical, as practical ASR systems must reliably support low-resource languages and generalize across diverse language families, and it directly hinders the deployment of SpeechLM-based ASR in real-world multilingual scenarios. As a result, it is essential to evaluate SpeechLMs on low-resource languages to ensure their generalizability across different language families. To address this problem, we propose LoASR-Bench, a comprehensive benchmark designed to evaluate low-resource automatic speech recognition (ASR) of the latest SpeechLMs across diverse language families. LoASR-Bench comprises 25 languages from 9 language families, featuring both Latin and non-Latin scripts, enabling cross-linguistic and cross-script assessment of ASR performance of current SpeechLMs. Experimental results highlight the limitations of the latest SpeechLMs in handling real-world low-resource languages.

Knowledge graph question answering (KGQA) is a promising approach for mitigating LLM hallucination by grounding reasoning in structured and verifiable knowledge graphs. Existing approaches fall into two paradigms: retrieval-based methods utilize small specialized models, which are efficient but often produce unreachable paths and miss implicit constraints, while agent-based methods utilize large general models, which achieve stronger structural grounding at substantially higher cost. We propose RouterKGQA, a framework for specialized–general model collaboration, in which a specialized model generates reasoning paths and a general model performs KG-guided repair only when needed, improving performance at minimal cost. We further equip the specialized with constraint-aware answer filtering, which reduces redundant answers. In addition, we design a more efficient general agent workflow, further lowering inference cost. Experimental results show that RouterKGQA outperforms the previous best by 3.57 points in F1 and 0.49 points in Hits@1 on average across benchmarks, while requiring only 1.15 average LLM calls per question. Codes and models are available at https://github.com/Oldcircle/RouterKGQA.

Translating natural language questions to SQL queries (Text-to-SQL) is a long-standing problem in database research. Recent efforts have focused on improving accuracy by building increasingly complex multi-stage large LLM pipelines, layering task decomposition, schema linking, and model-based query selection on top of an LLM. Despite this growing complexity, a substantial gap (>10%) between such systems and human experts persists on benchmarks, suggesting that pipeline engineering alone has hit a ceiling. We show that human-level Text-to-SQL performance is achievable by fine-tuning an LLM using RLVR on clean data, without pipeline components. In this paper, we identified the dominant bottleneck for RLVR on Text-to-SQL: existing training data contains pervasive annotation errors that mislead optimization. To address this, we developed a multi-round, expert-driven verification pipeline and used it to curate BIRD-Platinum, a dataset of 2.5k verified instances sampled from BIRD Train, with errors corrected in 61% of instances. We show that fine-tuning Qwen3-235B on BIRD-Platinum yields consistent improvements (11-16%) over BIRD Train on Arcwise-Plat (an expert-verified version of BIRD) and Spider2, outperforming SOTA open-source systems by 0.6-16%. Furthermore, we diagnosed two failure modes of standard RLVR on Text-to-SQL. We find that (1) result-based rewards have non-trivial false positive rates, and (2) models systematically ignore the external knowledge in BIRD-style problems. To address them, we propose ReViSQL-BIRD, a specialized reward shaping method that combines result-based rewards with SQL equivalence verification and leverages process rewards for incentivizing external-knowledge use. We fine-tuned Kimi-K2.6 with ReViSQL-BIRD. On Arcwise-Plat, ReViSQL-BIRD-K2.6 is the first method to achieve human-level accuracy (92.96%), outperforming top five open-source systems by 10-22%.

Explainable AI (XAI) research has experienced substantial growth in recent years. Existing XAI methods, however, have been criticized for being technical and expert-oriented, motivating the development of more interpretable and accessible explanations. In response, large language model (LLM)-generated XAI narratives have been proposed as a promising approach for translating post-hoc explanations into more accessible, natural-language explanations. In this work, we propose a multi-agent framework for XAI narrative generation and refinement. The framework comprises the Narrator, which generates and revises narratives based on feedback from multiple Critic Agents on faithfulness and coherence metrics, thereby enabling narrative improvement through iteration. We design five agentic systems (Basic Design, Critic Design, Critic-Rule Design, Coherent Design, and Coherent-Rule Design) and systematically evaluate their effectiveness across five LLMs on five tabular datasets. Results validate that the Basic Design, the Critic Design, and the Critic-Rule Design are effective in improving the faithfulness of narratives across all LLMs. Claude-4.5-Sonnet on Basic Design performs best, reducing the number of unfaithful narratives by 90% after three rounds of iteration. To address recurrent issues, we further introduce an ensemble strategy based on majority voting. This approach consistently enhances performance for four LLMs, except for DeepSeek-V3.2-Exp. These findings highlight the potential of agentic systems to produce faithful and coherent XAI narratives.

We investigate the separation of literal interpretation from contextual inference in a collaborative block-building tasks, where an agent must resolve underspecified instructions using context. We adapt an existing two-speaker psycholinguistic paradigm into an interactive benchmark called Build What I Mean (BWIM). This setup contrasts a pragmatically cooperative speaker with one who is only literally reliable. In BWIM, models face underspecified instructions and must choose between making a contextual inference or requesting clarification at a small communication cost. Evaluating several state-of-the-art LLMs, we find a clear dissociation between judgment and action. Although models successfully detect speaker unreliability in explicit confidence ratings, they fail to leverage this awareness when taking action. Instead of deploying efficient clarification strategies, models default to suboptimal behaviors. These include partner-blind over-clarification and question-averse guessing under uncertainty. BWIM provides a controlled environment to evaluate online partner adaptation and contextual reasoning in interactive settings.

Reinforcement learning (RL) has become a standard paradigm for post-training and aligning Large Language Models (LLMs), yet recent evidence suggests it faces a persistent "capability ceiling": unlike classical RL systems that discover novel strategies, RL for LLMs often acts as a mere refiner of patterns already latent in pre-trained weights. In this work, we identify a fundamental structural bottleneck: while classical RL relies on compact, informative Markov states, current LLM post-training formulations are tethered to an ever-expanding history of actions. We revisit a classical principle long central to RL yet absent from LLM post-training: explicit Markov states. Theoretically, we provide rigorous guarantees demonstrating that leveraging estimated Markov states can significantly reduce sample complexity. Empirically, we show that introducing Markov states consistently breaks the performance boundaries of standard RL post-training across a suite of complex logic puzzles. Our findings suggest that moving beyond "history-as-state" modeling in favor of structured Markovian representations is essential for unlocking open-ended discovery and genuinely new reasoning capabilities in Generative AI.

Transformers have shown inconsistent success in AI planning tasks, and theoretical understanding of when generalization should be expected has been limited. We take important steps towards addressing this gap by analyzing the ability of decoder-only models to verify whether a given plan correctly solves a given planning instance. To analyse the general setting where the number of objects – and thus the effective input alphabet – grows at test time, we introduce C*-RASP, an extension of C-RASP designed to establish length generalization guarantees for transformers under the simultaneous growth in sequence length and vocabulary size. Our results identify a large class of classical planning domains for which transformers can provably learn to verify long plans, and structural properties that significantly affects the learnability of length generalizable solutions. Empirical experiments corroborate our theory.

This study presents a hybrid topic modelling framework for computational literary analysis that integrates Latent Dirichlet Allocation (LDA) with sparse Partial Least Squares Discriminant Analysis (sPLS-DA) to model thematic structure and longitudinal dynamics in narrative poetry. As a case study, we analyse Evgenij Onegin-Aleksandr S. Pushkin's novel in verse-using an Italian translation, testing whether unsupervised and supervised lexical structures converge in a small-corpus setting. The poetic text is segmented into thirty-five documents of lemmatised content words, from which five stable and interpretable topics emerge. To address small-corpus instability, a multi-seed consensus protocol is adopted. Using sPLS-DA as a supervised probe enhances interpretability by identifying lexical markers that refine each theme. Narrative hubs-groups of contiguous stanzas marking key episodes-extend the bag-of-words approach to the narrative level, revealing how thematic mixtures align with the poem's emotional and structural arc. Rather than replacing traditional literary interpretation, the proposed framework offers a computational form of close reading, illustrating how lightweight probabilistic models can yield reproducible thematic maps of complex poetic narratives, even when stylistic features such as metre, phonology, or native morphology are abstracted away. Despite relying on a single lemmatised translation, the approach provides a transparent methodological template applicable to other high-density literary texts in comparative studies.

The vision of an inclusive World Wide Web is impeded by a severe linguistic divide, particularly for communities in low-resource regions of Southeast Asia. While large language models (LLMs) offer a potential solution for translation, their deployment in data-poor contexts faces a dual challenge: the scarcity of high-quality, culturally relevant data and the prohibitive energy costs of training on massive, noisy web corpora. To resolve the tension between digital inclusion and environmental sustainability, we introduce Sustainable Agent-Guided Expert-tuning (SAGE). This framework pioneers an energy-aware paradigm that prioritizes the "right data" over "big data". Instead of carbon-intensive training on unfiltered datasets, SAGE employs a reinforcement learning (RL) agent, optimized via Group Relative Policy Optimization (GRPO), to autonomously curate a compact training set. The agent utilizes a semantic reward signal derived from a small, expert-constructed set of community dialogues to filter out noise and cultural misalignment. We then efficiently fine-tune open-source LLMs on this curated data using Low-Rank Adaptation (LoRA). We applied SAGE to translation tasks between English and seven low-resource languages (LRLs) in Southeast Asia. Our approach establishes new state-of-the-art performance on BLEU-4 and COMET-22 metrics, effectively capturing local linguistic nuances. Crucially, SAGE surpasses baselines trained on full datasets while reducing data usage by 97.1% and training energy consumption by 95.2%. By delivering high-performance models with a minimal environmental footprint, SAGE offers a scalable and responsible pathway to bridge the digital divide in the Global South.

Efficient communication requires balancing informativity and simplicity when encoding meanings. The Information Bottleneck (IB) framework captures this trade-off formally, predicting that natural language systems cluster near an optimal accuracy-complexity frontier. While supported in visual domains such as colour and motion, linguistic stimuli such as words in sentential context remain unexplored. We address this gap by framing translation as an IB optimisation problem, treating source sentences as stimuli and target sentences as compressed meanings. This allows IB analyses to be performed directly on bitexts rather than controlled naming experiments. We applied this to spatial prepositions across English, German and Serbian translations of a French novel. To estimate informativity, we conducted a pile-sorting pilot-study (N=35) and obtained similarity judgements of pairs of prepositions. We trained a low-rank projection model (D=5) that predicts these judgements (Spearman correlation: 0.78). Attested translations of prepositions lie closer to the IB optimal frontier than counterfactual alternatives, offering preliminary evidence that human translators exhibit communicative efficiency pressure in the spatial domain. More broadly, this work suggests that translation can serve as a window into the cognitive efficiency pressures shaping cross-linguistic semantic systems.

Machine Translation (MT) and automatic MT evaluation have improved dramatically in recent years, enabling numerous novel applications. Automatic evaluation techniques have evolved from producing scalar quality scores to precisely locating translation errors and assigning them error categories and severity levels. However, it remains unclear how to reliably measure the evaluation capabilities of auto-evaluators that do error detection, as no established technique exists in the literature. This work investigates different implementations of span-level precision, recall, and F-score, showing that seemingly similar approaches can yield substantially different rankings, and that certain widely-used techniques are unsuitable for evaluating MT error detection. We propose "match with partial overlap and partial credit" (MPP) with micro-averaging as a robust meta-evaluation strategy and release code for its use publicly. Finally, we use MPP to assess the state of the art in MT error detection.

Do LLMs talk like us? This question intrigues a multitude of scholar and it is relevant in many fields, from education to academia. This work presents an interpretable statistical feature for distinguishing human written and LLMs generated dialogue. We introduce a lightweight metric derived from semantic categories distribution. Using the Empath lexical analysis framework, each text is mapped to a set of thematic intensity scores. We define semantic delta as the difference between the two most dominant category intensities within a dialogue, hypothesizing that LLM outputs exhibit stronger thematic concentration than human discourse. To evaluate this hypothesis, conversational data were generated from multiple LLM configurations and compared against heterogeneous human corpora, including scripted dialogue, literary works, and online discussions. A Welch t-test was applied to the resulting distributions of semantic delta values. Results show that AI-generated texts consistently produce higher deltas than human texts, indicating a more rigid topics structure, whereas human dialogue displays a broader and more balanced semantic spread. Rather than replacing existing detection techniques, the proposed zero-shot metric provides a computationally inexpensive complementary signal that can be integrated into ensemble detection systems. These finding also contribute to the broader empirical understanding of LLM behavioural mimicry and suggest that thematic distribution constitutes a quantifiable dimension along which current models fall short of human conversational dynamics.

The ubiquity of multimedia content is reshaping online information spaces, particularly in social media environments. At the same time, search is being rapidly transformed by generative AI, with large language models (LLMs) routinely deployed as intermediaries between users and multimedia content to retrieve and summarize information. Despite their growing influence, the impact of LLM inaccuracies and potential vulnerabilities on multimedia information-seeking tasks remains largely unexplored. We investigate how generative AI affects accuracy, efficiency, and confidence in information retrieval from videos. We conduct an experiment with around 900 participants on 8,000+ video-based information-seeking tasks, comparing behavior across three conditions: (1) access to videos only, (2) access to videos with LLM-based AI assistance, and (3) access to videos with a deceiving AI assistant designed to provide false answers. We find that AI assistance increases accuracy by 3-7% when participants viewed the relevant video segment, and by 27-35% when they did not. Efficiency increases by 10% for short videos and 25% for longer ones. However, participants tend to over-rely on AI outputs, resulting in accuracy drops of up to 32% when interacting with the deceiving AI. Alarmingly, self-reported confidence in answers remains stable across all three conditions. Our findings expose fundamental safety risks in AI-mediated video information retrieval.