Automated detection of machine-generated text is applied where the verdict carries a cost for the author, yet origin is a property of the production process, not an observable property of the text. The conditions under which a text is produced push it in the same direction in which a generative model is presumed to push it. This study measures the contribution of these conditions while origin is held fixed. The Russian-language corpus of 1882 documents comprises a machine part, generated by four generation channels (each backed by its own model family) across 45 tasks in three task modes, and an archival human part annotated for source regimentation level before the texts were read. The primary contrast is computed within the machine corpus: both halves of a pair are generated by the same model for the same task and differ only in task wording. Machine-likeness is measured by four operationalisations: an interpretable index with weights fixed before data collection, a trained classifier, a zero-shot metric on an open model, a judge model. The main result is disagreement between procedures: changing the task mode shifted the scores of all four, but three gave a negative shift, the judge a positive one, and none of the four intervals covers zero. The comparison plan and the specifications of procedures 2-4 were fixed after the result of procedure 1, before procedures 2-4 ran; the disagreement survived correction of a corpus defect and of the nested validation. A baseline on the source identifier alone reaches a median AUROC of 1.0000 on the thirteen group splits where it is defined, and removing emphasis markers flipped a confident machine verdict to human without changing the full text by a single character. The scores are properties of a specific measurement system and corpus: they reflect data collection, repetition, formatting, and preprocessing decisions as well as linguistic style.
Banks validate language models, but they deploy systems: a model inside a harness that decides what it sees, which tools it can call and who checks its actions. We test whether that harness changes compliance with banking conduct rules when the model is held fixed. Four open-weight models each ran five harnesses (a pipeline that makes one structured decision, a ReAct-style tool loop, a loop in which the policy is reachable only through search, a loop with a long benign customer memory, and a loop with a same-model compliance reviewer) on 156 pre-registered cases in a simulated retail bank whose 11 rule families paraphrase EU consumer-credit, payments, investor-protection, data-protection and anti-money-laundering provisions. Every one of the 3120 two-turn episodes was scored deterministically from the executed tool trace, with violations and over-refusals counted separately. Pooled across models the harness moved the correct-case rate only between 87.3% and 92.1%, but within a model it mattered a great deal: for one model the spread between its best and worst harness was 11.5 percentage points, enough to move it from first to last place among the four models, while for another the harness made no detectable difference; the model x harness interaction itself was suggestive rather than conclusive. The same-model reviewer cut violations from 9.6% to 6.4% without raising over-refusals, contrary to our pre-registered prediction. Rules that require date arithmetic failed in every harness. In an exploratory analysis with two LLM judges and a blind human check, 26% of violations were accompanied by a customer message that denied the action the agent had just executed, which a transcript-only review would record as a correct refusal. For model-risk management the unit of validation should be the model-harness pair, and conduct testing should read the action log, not the conversation.
For humans, referential disambiguation is a collaborative process involving clarification questions asked when they will meaningfully change the outcome. However, large language models (LLMs) have been shown to be biased toward premature answering, preferring definite single-turn responses over more clarification exchanges. In this paper, we study the problem of "strategic clarification" -- asking clarification questions in a reference game when their answers will narrow the set of candidate referents. We construct a controlled test environment from the SIMMC 2.1 dataset that records how many questions each ambiguous reference needs to be successfully resolved, and examine how baseline LLM behavior and common post-training methods ask clarifying questions vs. attempt to answer the referent. We also propose BACE (Belief-Aware Critic with Epistemic tracking), a lightweight action chooser method on top of a frozen LM backbone, and show that BACE more consistently chooses the optimal number of clarifying questions than methods such as GRPO and ArCHer.
Search can help LLM agents correct mistakes, but their starting conditions may shape which evidence they find and use. Drawing on cognitive-bias paradigms, we develop a controlled search harness to examine result order, records of prior work, and initial hypotheses while holding each question and document collection fixed. Matched controls distinguish order effects from differences in document composition and the effects of failed work from context length. Across 10 models and 1,000 synthetic search worlds, we track queries, retrieved documents, document openings, and final answers. Our comparisons suggest that final answers are more sensitive to initial evidence order during active search than when all evidence is supplied in full. Dead-end histories can reduce strategy switching relative to matched neutral histories, indicating an effect beyond context length. Under wrong hypotheses, stage measurements distinguish reduced retrieval or opening from errors that persist after counter-evidence is opened. Exploratory interventions show that opening counter-evidence more often can coincide with lower accuracy, while accuracy can improve with little change in opening.
Generative AI is currently fragmented, with separate subscriptions required for text generation, image synthesis, and code assistance, leading to high costs and workflow disruptions. AI Express is a SaaS platform that consolidates these capabilities into a unified pay-as-you-go service. The platform provides a range of AI tools, including a Core Code Reviewer that evaluates GitHub repositories, a Text-to-Website generator that converts text into fully functional websites, script generation, grammar checking, data analysis, and watermark removal, all accessible through a single interface without requiring technical expertise.
Small Language Models (SLMs) are increasingly deployed on consumer-grade hardware, yet their ability to preserve information across progressively longer contexts remains insufficiently understood. Existing benchmarks primarily evaluate retrieval-oriented tasks, providing limited insight into whether successful retrieval reflects comprehensive long-context memory preservation. To address this limitation, this paper introduces Simple Retrieval-Reconstruction Memory (SRRM), a lightweight and fully reproducible benchmark that jointly evaluates Targeted Retrieval and Global Information Reconstruction. This study also proposes the Memory Preservation Ratio (MPR), a benchmark-level metric that quantifies the relationship between retrieval and reconstruction performance. SRRM was evaluated using a synthetic dataset on three Qwen2.5 models (0.5B, 1.5B, and 3B parameters) across context loads ranging from 10 to 100 factual statements. Experimental results show that reconstruction performance deteriorated substantially faster than retrieval accuracy as contextual information increased, indicating that retrieval alone is insufficient to characterize long-context memory preservation. Among the evaluated models, Qwen2.5-1.5B achieved the highest retrieval accuracy (86.7%), reconstruction performance (45.7%), and MPR (0.53), outperforming the larger Qwen2.5-3B model (72.2%, 18.5%, and 0.26, respectively). These findings demonstrate that increasing parameter count does not necessarily improve long-context memory preservation. Overall, SRRM provides a reproducible and extensible benchmark for systematically evaluating long-context memory while revealing complementary memory behaviors that remain hidden under retrieval-only evaluation frameworks.
Generalisability is a central question for deep neural networks, including the cross-language generalisation of speaker embeddings. This study examined whether deep neural speaker embeddings generalise across languages and how such generalisation relates to human perception and acoustic structure. Using Cantonese--English bilingual speech presented in within- and cross-language conditions, we derived similarity from 18 state-of-the-art models and compared it with human ratings and acoustic features using linear mixed models and representational similarity analysis. Embeddings exhibited cross-language generalisation that broadly paralleled human behaviour, reflecting accurate identity discrimination and sensitivity to language switching. This generalisation relied on largely shared acoustic cues with language-sensitive weighting, mirroring human patterns. Together, these findings provide converging evidence for human--machine alignment in cross-language voice similarity processing.
Most memory systems for LLM agents are retrospective, preserving and retrieving past information to inform current decisions. Proactive agents, however, also require prospective memory: retaining future-directed intentions and realizing them when appropriate opportunities arise, even without explicit requests. Crucially, remembering the right intention does not guarantee acting at the right time: a related observation may call for waiting, while a later cue may make the same intention actionable. Inspired by cognitive theories of human prospective memory, we introduce Pro-Memory, a memory framework that separates maintaining future intentions from deciding when and how to realize them. Pro-Memory preserves pending objectives linked to their supporting evidence, assesses new observations against their original goals and activation conditions, and coordinates appropriate contributions with the agent's ongoing actions. It further updates intention states based on actions actually issued or evidence of resolution, while an experience memory distills feedback from completed episodes into guidance for future decisions. Experiments on two benchmarks spanning delayed-intention tasks and open-ended proactive assistance demonstrate consistent improvements in proactive task performance over representative memory approaches across multiple LLM backbones, while limiting unnecessary interventions.
Legislative hearings are distinct communication events with their own human recognizable structure and logic, yet they have been analyzed as distinct phenomena beyond their transcripts. Understanding legislative hearings at a deeper level is a major step in analyzing the work and power relationships within democratic bodies. In this work, we present a novel end-to-end pipeline for parsing rhetorical structures from legislative committee hearings with the goal of improving legislative discourse analysis and democratic accountability. We also propose a rhetorical structure grammar and find that 55.21% of California state bill discussions in our dataset can be parsed with this grammar. We evaluate our pipeline with two utterance section classifiers on a random subset of over 2,000 annotated utterances corresponding to 100 hearings. We find that a logistic regression-based classifier with grammar-based output masking and utterance tag smoothing outperforms more complex models, achieving an F1 score of 52.8% against our ground truth discourse section labels.
When a language model chooses a number between 1 and 30, every integer in the range is valid, yet the model can favor particular answers. We use numerical choice to study how preference is processed across layers and tokens. Perturbation probing and gradient attribution nominate neurons for activation interventions. In a selected Qwen3-4B case, a pre-answer edit changes 17 to 15 while both runs first emit 1. After the edit ends, selective cache and attention-output replacements trace part of its continuing influence on the second digit. Replacing one attention head's output or the other heads' outputs jointly produces similar immediate readout contributions; both interventions restore 17 but have different final log-odds effects on 7 versus 5. Holding later multilayer perceptron (MLP) outputs at their values from the edited run without either replacement reduces this effect gap from 0.119 to 0.0025 nat. Later computation therefore changes how similar local contributions affect the continuation. Retrospective estimates across eighteen contexts capture much of the probability movement but miss the selected answer switch. Prompt tests also separate influence from answer changes: one held-out prompt retains influence without switching, and only one of twelve fresh wordings qualifies for the reader/MLP test; there, joint MLP clamping again reduces the gap between the two reader effects while the answer remains 17. These results use numerical choice as a controlled case for studying persistent state and conditional downstream processing; they do not establish the origin or a complete circuit of natural numerical preference.
Forecasting emotion in conversation is increasingly framed as predicting the next emotional state's value. We show this targets a hard quantity in a state-observed testbed and propose a learnable, deployable reformulation instead. On valence-arousal-dominance trajectories from a dyadic corpus, repeating a speaker's own previous state sets a persistence baseline that no frozen-encoder correction meaningfully improves upon and a state-space filter exceeds; the floor requires the historical state to be observed, since an estimated anchor collapses persistence below a direct forecaster. What survives is the timing of a departure from persistence, learnable on three of four corpora and near chance on the highest-base-rate corpus, a negative result we report. We recast the task as selective departure forecasting: a lead-time-weighted metric with a leakage protocol, and commitment policies whose best choice switches with the departure rate. A dialogue-level conformal risk control bounds the expected loss of committed predictions, meeting the target in all twelve configurations tested, though the guarantee is bought through abstention, with coverage and alarm recall falling sharply on the hardest corpus. One tested language model's confidence fails as a ranking signal where another succeeds, a model-dependent failure not yet attributable to family versus scale. Code and result logs are available from the corresponding author upon request.
Speech transcription errors can be difficult to recognize when competing word hypotheses are unfamiliar or visually similar. We describe an interface that combines alternative transcripts, replay and contextually checked pictograms for selected uncertain concepts. The system will initially support French and Spanish functional utterances. A counterbalanced study will assess whether pictorial feedback helps listeners recover the speaker's intended action beyond text alternatives alone. We will also quantify semantic mismatch and additional interaction time, since an inappropriate image can reinforce the wrong interpretation rather than repair it.
Comparisons in conversational emotion-shift forecasting are easily distorted when a learned model receives rich features but its transition baseline is given a noisy or unavailable current-emotion label, when hyperparameter budgets differ, or when target-turn information enters the representation. We present an information-matched, leakage-safe auditing protocol that closes these gaps with an explicit commitment point, executable causality tests, equal four-configuration validation searches, and seed-dialogue-aware inference. Applying it to six causal model families on four text corpora and two feature-source sensitivity configurations shows that a deployable train-only current-emotion classifier feeding the transition baseline yields a different verdict than an oracle diagnostic where both sides receive gold current emotion. Under the deployable condition, learned models improve over the predicted-label baseline on all four primary corpora; DailyDialog gains are large and survive a 72-comparison Holm correction (best ROC-AUC gain = +0.167, adjusted p=.0144). Under matched gold information, the baseline remains competitive on IEMOCAP and stronger on MELD, while every model retains a large DailyDialog advantage. We additionally report precision-recall AUC, calibration, a strictly utterance-local feature control, same-person self-shift forecasting, and a future-access dose-response test. An eight-item leakage checklist accompanies the protocol. Code, raw seed predictions, and reporting artifacts are available from the corresponding author upon request.
Large language model (LLM) shopping assistants now recommend products in multi-turn conversations. However, reliable evaluation faces Data scarcity and Difficulty Standardizing Criteria. To address these obstacles, we introduce CARTBench for CART (Complex and Ambiguous Requests with Timelines). For Data scarcity, we curate cases from real commercial shopping needs across 17 retail domains, pairing timestamped dialogue with 1,603 product records. For Difficulty Standardizing Criteria, we develop case-specific rubrics guided by Ideal Response Principles, distilled from excellent replies manually selected from successfully completed orders. These principles also guide Principle-Guided CART Prompting, our structured recommendation strategy. The experimental benchmark contains 1,328 expert-annotated criteria across Core, Excellence, and Safety. To answer the CART question, we evaluate 7 flagship commercial API models, including GPT-6.1-Sol and Claude-Opus-5, and 35B+ parameters open-weight models. In the first full-suite evaluation under the frozen rubrics, 3 of 10 models exceed 90% on Core, while none exceeds 45% on Excellence. Therefore, we identify an understanding-to-execution gap between required behavior and high-quality purchase advice. To examine these limitations, we diagnose shared failures in product evidence and unresolved needs, linked to changing requirements and decisive unknowns. Code and data are available at https://github.com/Da1yuqin/CARTBench.
Streaming speech recognition is increasingly consumed by machines, such as translators and voice agents, that act on each token as it appears. A token that is later revised costs wasted work or a wrong action. In this work, we present Confucius4-R2T2 (short for Real Real-Time Transcription), a stable-prefix streaming ASR model that never modifies its earlier output while maintaining high accuracy and low latency. R2T2 is obtained by post-training Qwen3-ASR-1.7B on 1,500 hours of Chinese and English audio while freezing its pre-trained audio encoder and projector. The adaptation is based on the concept of Longest Stable Prefix (LSP), a new notion that we introduce in this work. We prove that LSP is monotone under the prefix order, consistent with the full-audio output, and maximal among labels that equal a whole earlier output. Learning from LSP labels converts a full-context model into a stable-prefix streaming one without any architectural change. Extensive evaluations show that R2T2 brings streaming recognition close to the base model's full-context quality while delivering stable text with low emission lag. Its gains extend beyond the post-training languages, demonstrating a practical path from full-context ASR to stable streaming. We open-source the model weights and code at https://huggingface.co/netease-youdao/Confucius4-R2T2 and https://github.com/netease-youdao/Confucius4-R2T2.
Different retrievers exist for retrieval-augmented generation, yet none performs best for all queries. Therefore, the key question is: given a query and retrieved passages from different retrievers, which one is most suited for that query? Existing routers usually make predictions from query-specific features and a fixed pool of retrievers; such routers have no output for an unseen retriever and are incapable by design on a new corpus. To fill these gaps, RAGCompass reads the retrieved passage sets themselves: passages from an unfamiliar retriever or corpus stay directly comparable to other passages, so routing stays well designed where a query-only map would fail. We contrastively fine-tune a reranker against the answer each evidence set produces, which accounts for most of the gain; a small comparison head adds a moderate improvement, and neither component has retriever-specific parameters. RAGCompass improves answer F1 by 28% relative to the best fixed retriever, and outperforms fusion baselines, three query-only routers, and a prompted 8B language model. Generalization experiments show that RAGCompass keeps most of its gain under an unseen corpus, retriever, or reader, while query-only baselines fall to the best fixed retriever or below.
Cognitive Behavioral Therapy (CBT) demands more than empathic validation, requiring instead a structured formulation of the interplay among situations, automatic thoughts, emotions, and behaviors. Therapists formalize this interplay through a Cognitive Conceptualization Diagram, where maintaining longitudinal coherence necessitates continuous evidential adjudication to separate tentative hypotheses from established findings. Current dialogue systems track conversational history via retrieval or static profiles but lack the representational capacity to encode epistemic status. Consequently, provisional interpretations are often prematurely reified as established facts, inadvertently biasing downstream interventions. Addressing this limitation, CCDWiki is introduced as a plug-and-play advisor conceptualizing longitudinal therapy as evidence-grounded hypothesis adjudication. The approach encodes each clinical hypothesis as a mutable entry bound to corroborating and contradicting evidence. Across sessions, accumulating data strengthens supported entries, preserves provisional ones, and retracts contradicted assumptions. The resulting formulation distills into concise guidance for unmodified therapeutic language models. Experiments demonstrate substantial improvements in both conceptualization and intervention quality without requiring backbone retraining.
This paper presents an overview of \emph{SciClaimEval}, a pilot task at NTCIR-19 on verifying scientific claims against the tables and figures of the papers in which they appear. Our dataset pairs authentic claims from 180 papers in machine learning, natural language processing, and medicine with contrastive versions of their tables and figures, one supporting the claim and one minimally modified so that it does not. The task comprises two subtasks: deciding whether a piece of evidence supports a claim, and selecting which of two evidence versions does. Twenty-two teams registered, the highest number among NTCIR-19 tasks, and 14 submitted 177 runs. Most systems built on multimodal large language models, relying on prompting and multi-stage inference rather than task-specific fine-tuning. On the first subtask, the best systems reached 98.1\% pair accuracy, but all six top-ranked systems relied on the contrastive pairs; the best system that judged each claim-evidence pair on its own scored 24.3 points lower. Open-weight models, embedded in carefully designed pipelines, matched proprietary ones. Yet verifying a single claim against its evidence, the setting that matters in practice, remains an open problem.
In natural human-AI dialogue, users' corrections, requests, and self-descriptions carry reusable signals about how they want work to be done. These signals reveal individual process preferences, user-specific standards for evidence, depth, analytical focus, procedure, and delivery. We ask whether a preference revealed in one dialogue can shape an assistant's later answer when the task does not restate it. ProcessPrefBench uses three months of natural dialogues from two knowledge workers to build 66 source-target cases separately judged for source evidence and target applicability. Sources are preselected, so the benchmark tests application rather than retrieval. We test three models with target-only, full-history, and distilled-preference prompts. Target-specific 0-2 rubrics score preference realization, and a task gate flags severe failures. Compared with target-only prompting, full-history and imperative distilled-preference prompting raise preference realization by 0.204 and 0.244 points on the 0-2 scale, respectively, with no reliable advantage from distillation. Fabricated-evidence failures rise from 5.8% under target-only prompting to 10.1% under full-history and 10.4% under imperative prompting. ProcessPrefBench highlights a tension in learning from dialogue, where drawing on prior interactions can improve preference realization yet also invite fabricated evidence. We hope it helps build assistants that honor a user's process preferences without sacrificing factual reliability.
Expert deactivation at inference time changes both token routing and expert weights, and its effect on performance depends on how these changes are implemented. We compare removal, freezing, remaining-expert renormalization, and unnormalized re-routing on identical disabled sets across five instruction-tuned checkpoints from four model families, with a base OLMoE checkpoint as an additional comparison. We develop a weight-sum model that relates the loss of expert contributions and changes in branch scale to the weight rule used during training. With half of the experts deactivated according to activation-based rankings, remaining-expert renormalization yields the highest perplexity in models trained with raw router probabilities, whereas unnormalized re-routing yields the highest perplexity in models trained with normalized weights. Perplexity ranges from 1.4 to 73 times the intact baseline when the training rule is retained, and from 14 to 28,740 times the baseline under a rule mismatch. Task outcomes vary across evaluation types: on OLMoE, GSM8K strict-match accuracy falls to zero under all four implementations. Extending the model to changes in the number of active experts yields a further prediction, supported by the OLMoE experiments: the perplexity penalty from additional renormalization increases as fewer experts are activated.
Confidence estimates guide human review in LLM-built knowledge graphs, but better error prediction can yield little improvement in an audit. We study this gap through repairability, selection headroom and label validity. Modelling non-adaptive auditing as budgeted covering, we characterise the minimal decision-relevant error statistics and bound the regret of scalar-preserving first-order planning by a structural mass $\Delta^{\mathrm{cl}}$. We simulate ideal repair against benchmark-aligned labels for five extraction configurations and two cross-document graphs on Re-DocRED, followed by frozen confirmation on test and replication on 500 hash-selected training documents. On the held-out cross-document graphs, an error-rate $\times$ fan-out queue leaves 3.98-4.74 percentage points more errors than a joint model with the same actions. By comparison, opening shared actions for that joint planner reduces residual errors by 0.89-1.43 points relative to its own fact-only plans, in point estimates. Equivalence between joint and independent-block models remains inconclusive. Generic LLM verification improves three of four model-budget configurations on each held-out set. LLM adjudication identifies possible benchmark omissions.
点云分割是3D数据分析中的一项基础任务，应用遍及机器人、增强现实与虚拟现实以及遥感等领域。然而，点云固有的稀疏性、噪声和无序特性带来了显著挑战，尤其是在实现可靠分类方面。尽管最先进的方法依赖训练模型来显式建模不确定性，但在探索使用生成式方法估计不确定性所能扮演的不同角色方面仍存在空白。本工作通过显式建模认知不确定性并将其用于不同角色（即作为特征和用于损失函数）来研究确定性不确定性在点云分割中的影响。我们在四个大规模LiDAR基准（涵盖机载和移动采集，即ISPRS Vaihingen、AHN3、Paris-Lille-3D和Toronto-3D）上评估了一个基线人工神经网络（ANN）和两个不确定性感知变体（不确定性作为额外输入特征，以及作为训练损失中的逐样本置信度权重），并纳入了PointNet++和RandLA-Net参考模型及其不确定性增强对应模型，采用规范的空间不相交训练/测试划分。在该协议下，不确定性感知建模提升了平均交并比（mean Intersection over Union），找回了基线ANN完全无法预测的少数类，并在半数配置中改善了预测误差与估计不确定性之间的一致性（以Patch衡量）。
LLMs can repair noisy OCR transcripts of historical documents, but they also change text that was already correct. We propose EditVeto, a veto gate placed after a full-context LLM rewrite. The rewrite is character-aligned to the OCR input, its differences are split into blocks, and a 0.6B-parameter gate keeps or reverts each block. An oracle analysis motivates the design: given the true error locations, per-span generation is worse than no correction, whereas vetoing a full rewrite's edits helps. Trained on rewrites by two Qwen3 models and evaluated, frozen, on the eight HIPE-OCRepair-2026 test sets, the gate improves all five rewriters tested, two of them unseen in training: weighted cMER falls by 0.0026-0.0044 and the preference score rises by 0.15-0.25 (paired bootstrap, p ~ 0.0002 each). On published competition outputs it lowers cMER for three of four teams but raises it for the winning run (0.0050 to 0.0065). In a second, disclosed test look, a retrained gate still improves LoRA-tuned rewriters (cMER -0.0011 to -0.0021), and an ensemble of their gated edits reaches a cMER of 0.0069 and a preference score of 0.928, mostly through fine-tuning and pooling rather than gating. Built only on public models and data, it is behind the winner (0.0050) and tied with the second-ranked run (0.0071), both from the only team using domain pre-training, and ahead of all other published runs. A third, secondary look shows the unchanged gate improving zero-shot Qwen3.8-27B (cMER 0.0114 to 0.0108) but only the preference score of its LoRA-tuned version.
Small LoRA adapters trained on Tajik news and books lowered Qwen3's Tajik text loss at every size we tried, from 0.6B to 8B, by 0.60 to 1.45 nats per token, but none of them improved Tajik reading comprehension where it can be measured. On Belebele the 1.7B and 8B adapters left accuracy unchanged and the 4B adapter lowered it by 3.7 points; at 0.6B the score rose by 4.2 points but sits near chance and coincides with a change in which answer letter the model prefers, so we do not read it as understanding. At 8B, continued pretraining on about 375 million tokens of raw Tajik text was followed by a Belebele gain of 9.9 points. After a further supervised fine-tuning stage, our main model, e3, is 11.4 points above Qwen3-8B on Belebele; its English MMLU changed by -1.2 points, and a loss larger than 3.3 points is excluded. The same heavy recipe applied to Qwen3-0.6B-Base brought its text loss close to that of the 8B model after books-only pretraining, while its multiple-choice answers stayed near chance and its English MMLU fell by 7.6 points. We report 15 trained models and their 5 starting models from June to October 2025, scored in 2026 on one harness with paired intervals, and we report what went wrong: lost training records, a thinking-mode dataset with almost no reasoning in it, later branches whose training data was never documented, and a 4-bit release that contains no transformer layers.
Misinformation spreads faster than professional fact-checkers can manually verify it and write articles explaining their verdicts. This creates a need to automate the production of fact-checking articles. This paper details work done by the DS@GT ARC team for CLEF 2026 CheckThat! Lab Task 3 Fact-Checking Article Generation. We seek to generate a full fact-checking article from a claim, its original verdict, and a set of scraped source documents. Our team built a four-stage pipeline that cleans noisy evidence, retrieves and clusters relevant sources, generates the article by conditioning \texttt{Llama-3.1-8B-Instruct} on the original fact-checker verdict, and recovers inline citations through deterministic post-processing. We achieved a mean score of 0.428 and ranked 3rd out of 9 teams for the CLEF 2026 CheckThat! Lab Task 3, compared to a baseline mean score of 0.272. Our code is available on GitHub at https://github.com/dsgt-arc/checkthat-2026-task3.
透明性已成为AI治理的核心原则，但关于透明性意味着什么、为谁服务以及在何种情境下适用，仍缺乏共识。本文借鉴技术的社会建构（Social Construction of Technology）中的解释灵活性，考察不同利益相关者如何理解和体验生成式AI中的透明性。本研究基于一项覆盖400名最终用户、AI开发者和政策制定者的全球调查，并辅以与最终用户开展的结构化焦点小组，将定量发现扎根于当前透明性实践中最未被充分服务群体的定性见解。我们呈现三项关键发现。第一，三组群体对透明性的含义及其应提供什么有着相似理解，核心在于确保准确性和隐私。第二，最终用户在透明性特征的重要性与其感知可用性之间的差距显著大于开发者，揭示了权力如何塑造所构建的内容及其服务对象。仅28%的最终用户报告当前透明性实践非常好地反映了其需求，而开发者和政策制定者分别为53%和49%。第三，透明性是情境性建构的，优先级随情境和任务结构而变化，角色特定框架在情境需要时被激活，在不需要时则退隐。我们认为，有意义的AI透明性需要超越一刀切方案，包括当前对AI披露和静态文档的强调。
Hundreds of languages have written text but almost no translations, and whether a model learning one from untranslated text builds on a related language or familiar letters is unclear: they usually coincide. Here we show that familiar letters, not a related language, let an unseen language into a shared space of meaning. GlyphMind, a tokenizer-free vision transformer that reads text as pixels and lets any sentence be shown in any script, was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, a language and script it had never seen, it retrieved English translations seven times as often as the strongest simple baseline. Rewritten in unfamiliar letters, Lao did not enter although its close relative Thai was in pretraining; Santali, which had none, entered in Devanagari, familiar from Hindi, but not in its own script. Transliteration into familiar scripts may thus open multilingual models to languages without translations.
Expressiveness is central to speech-to-speech translation (S2ST), yet existing datasets rarely align source and target speech jointly in content, prosody, timing, and speaker identity. Consequently, current systems often rely on indirect conditioning or transfer mechanisms. To obtain this supervision, we introduce a training-data construction strategy that combines professional dubbing, which provides human-performed correspondence in content, prosody, and timing, with multilingual voice conversion to reduce speaker mismatch. We evaluate this supervision using existing expressive S2ST systems and DirectS2ST, an end-to-end method developed in this work that predicts the complete target-speech representation directly from source speech. Experiments on English-German and English-Spanish show that holistic parallel supervision improves expressive realization across all evaluated systems. DirectS2ST exhibits the largest gains in content, prosody, timing, speaker similarity, and human-rated expressiveness. To our knowledge, this work provides the first systematic investigation of holistic parallel supervision for expressive S2ST.
Decision models choose among predefined options in tasks such as routing and action selection. In multi-agent collaboration, a decision determines how members divide work, exchange resources, and act in parallel. We introduce TandemBench, a benchmark of 1,200 questions in multi-agent cooking scenarios where coordination choices affect goal completion. Models receive complete global information and candidate team plans, with reference answers established through execution and exact search. Models select an optimal team plan and, in paired variants that retain or remove it, judge whether better arrangements exist beyond the supplied options. We evaluate 40 native decision models and compare them with 10 general-purpose language models. The best native decision model achieves 62.50% decision accuracy. Native decision models remain close to chance in optimality verification. Paired tests reveal that some report better alternatives regardless of whether an optimal plan is already present. These findings show that team-plan decision accuracy alone can conceal systematic errors in judging whether the supplied options contain an optimal arrangement. The dataset and evaluation code are publicly available at https://github.com/lililibaohang/TandemBench.
Large language models increasingly automate complex tasks, but human decision capacity does not scale with their capabilities. In many applications, humans remain essential for providing private context, authority, or judgments unavailable to the agent. We formulate human decision capacity allocation: given multiple consultation opportunities and a limited budget, at which decisions should an LLM agent consult a human? The central challenge is estimating consultation value, which is executor-dependent because the same human input may benefit different downstream agents differently. We introduce Consultation Value Distillation (CVD), which estimates consultation values offline using Monte Carlo executor rollouts and distills these values into lightweight predictors for budgeted allocation, requiring no rollout-based value estimation at deployment. Across five batch and three sequential settings with two frontier executors, CVD improves allocation through joint and continuation-aware value learning, while independent consultation values remain competitive in several settings. CVD also broadly outperforms direct prompting and Value of Information where applicable, remains robust with substantially less rollout supervision, and enables nearly lossless approximate allocation at low search cost.
This study investigates the authorship impersonation capacity of six widely used large language models (LLMs): Gemini 3 Pro, Claude Opus, GPT 5.2, GPT 5.2 Pro, DeepSeek 3.2, and Grok 4.1 Fast. The evaluation was conducted using short product review texts selected to ensure high within author consistency in individual writing style. Two types of in context prompt were designed: a simple prompt with minimal instruction and a complex prompt incorporating explicit stylometric guidance for mimicry. Authorship mimicry was evaluated using a cosine distance based distance to training metric, supplemented by visual analyses of embedding distributions. The results indicate that authorship impersonation remained challenging for most LLMs, with Gemini 3 Pro showing the strongest apparent mimicry performance. However, subsequent overlap analysis revealed that Gemini 3 Pro extensively copied and reused sentences and phrases from the training texts, at levels far exceeding those observed among human authors. Follow up factor analyses showed that training set consistency was the most robust predictor of mimicry success across models, whereas author distinctiveness did not generalise as a reliable predictor. Global embedding visualisations further revealed a consistent macro level pattern: for most models, generated texts occupied regions of the embedding space that were largely separated from the corresponding human training clusters. This indicates limited convergence towards target author style and suggests the presence of model specific lexical and phraseological preferences. The findings highlight the importance of distinguishing genuine stylistic imitation from direct textual reuse and have implications for the forensic analysis of textual evidence involving potential LLM enabled impersonation.
How much phonological knowledge do language models (LMs) possess about Chinese characters, and how is this knowledge represented and used? We introduce PinyinBench, a controlled benchmark covering homophones, phonological features (initials, finals, and tones), and polyphonic characters. We investigate 17 text-only and multimodal LMs through three complementary approaches: behavioral evaluation, layer-wise probing, and activation steering. Behaviorally, models demonstrate substantial knowledge of initials and finals, with a frontier model (DS-v4-flash) scoring 90%+ in homophone judgment, when two characters share no phonological features. However, its accuracy drops sharply to 72% when two characters differ only in lexical tone. Probing reveals that tone is highly decodable from embeddings and improves modestly across layers, whereas segmental information starts less decodable and improves substantially with depth. Activation steering shows that segmental representations can systematically influence generation, while tone exhibits much weaker effects. Furthermore, audio-bearing training improves phonological decodability, including tone, without corresponding improvements in steering effectiveness. Together, these findings reveal distinct representational and functional profiles for segmental and tonal information, with tonal information much harder to access and manipulate, demonstrating the necessity for combined evaluation, probing, and causal intervention for understanding linguistic knowledge in LMs.