The paper investigates how multilingual large language models perform when generating stories in Urdu, a low‑resource language. The authors created a corpus of 93 Urdu stories produced by GPT‑5.1, Qwen‑3‑Max, and DeepSeek‑3.1, and manually annotated errors across a nine‑label taxonomy covering linguistic, semantic, and cultural aspects. Findings reveal frequent grammatical and semantic mistakes, lack of coherence, unnatural repetition, and pervasive cultural shallowness, with few‑shot prompting failing to resolve many of these issues.
By Farah Adeeba, Abdul Rafae Khan, Rajesh Bhatt, Hassan Sajjad
The paper investigates using Vision Language Models (VLMs) to accelerate verification and validation (V&V) of classification models by automatically detecting systematic errors. It introduces a VLM-based error slice detection (ESD) method that groups and labels errors, demonstrating its ability to identify perturbations in a non-military dataset and to cluster images by surroundings in a military context. The study highlights challenges such as underrepresentation of defence data in VLM training and limited contextual diversity, and suggests that while fully automated V&V is not yet feasible, VLMs could speed up the process in the future.
By Dieuwertje Alblas, Alma M. Liezenga, Jan Erik van Woerden, Fedor Taggenbrock, Dalia Aljawaheri, Klamer Schutte
HALDETECT is a system developed for the English hallucination-detection track of ImageEval 2026, where the task is to identify the single visually grounded statement among three culturally plausible options. The approach treats the problem as a contrastive decision, outputs the answer before an explanation, and bases reasoning on colour/texture, shape/form, and context. The best model fine‑tunes Qwen2.5‑VL‑7B‑Instruct with 4‑bit QLoRA, freezes the vision encoder, and achieves a Contrastive Instability score of 0.035 on the test set, placing third among eight teams.
By Syed Mohaiminul Hoque, Md Sakhawat Hossain
The paper introduces LASP, a framework that uses large language models to generate structured B‑spline priors from procedural modeling histories. By translating design intent into rich textual descriptions, LASP provides semantic reasoning that guides conventional CAD solvers toward more accurate and coherent surface fitting. Experiments show that language‑driven priors outperform traditional machine learning approaches, establishing a new paradigm for language‑guided geometric optimization.
By Yunzhong Lou, Yusheng Luo, Jiahao Li, Yu Song, Xiangdong Zhou
SG-Blend introduces a per‑layer adaptive activation that interpolates between a bias‑corrected, parametric Swish variant (SSwish) and GELU, using a learnable blend coefficient, sharpness, and zero‑centering bias. The method adds only three scalars per feed‑forward block and, on BERT‑style IMDB classification, matches peak accuracy while reducing seed‑to‑seed variance by 42 %. It also achieves the lowest validation perplexity on WikiText103 and generalizes to computer vision and other domains.
By Gaurav Sarkar, Syed Affan Daimi, Jay Gala, Subarna Tripathi
The paper introduces a new method for measuring metacognitive abilities in large language models (LLMs) without relying on self-reports, instead testing how well models can use knowledge of their internal states. Using two experimental paradigms, the authors find that recent frontier LLMs can assess and use their own confidence when answering factual and reasoning questions, and can anticipate and appropriately employ the answers they would give. The study also shows that these abilities are limited in resolution, context-dependent, differ qualitatively from human metacognition, and vary across models with similar capabilities, suggesting post‑training processes influence metacognitive development.
By Christopher Ackerman
CHRONOBERG is a temporally structured corpus of English book texts covering 250 years, curated from Project Gutenberg and enriched with temporal annotations. It enables quantification of lexical semantic change via time‑sensitive Valence‑Arousal‑Dominance analysis and the creation of historically calibrated affective lexicons. Experiments show that language models trained sequentially on CHRONOBERG struggle to encode diachronic shifts, highlighting the need for temporally aware training and evaluation pipelines.
By Niharika Hegde, Subarnaduti Paul, Lars Joel-Frey, Manuel Brack, Kristian Kersting, Martin Mundt, Patrick Schramowski
The paper introduces a method for identifying the dominant knowledge source behind large language model (LLM) outputs, distinguishing between faithfulness violations (misuse of provided context) and factuality violations (errors in internal knowledge). A simple linear probe trained on hidden representations can reliably classify this source, and the authors present AttriWiki, a self‑supervised pipeline that generates labeled training data by prompting models to recall withheld entities or read them from context. Probes trained on AttriWiki achieve high Macro‑F1 scores across several models and datasets, generalize zero‑shot to a benchmark, and show that attribution mismatches can increase error rates by up to 70%.
"whyItMatters":"The study demonstrates that knowing the source of an LLM’s answer is crucial for effective mitigation of hallucinations, as attribution mismatches significantly raise error rates."
By Ivo Brink, Alexander Boer, Dennis Ulmer
The paper announces VectraYX‑Vision‑1B, a Spanish/LATAM cybersecurity vision‑language model that uses a natively‑trained visual tower (Qwen2‑VL) and a frozen decoder. The authors report that transplanting the visual tower improves the failing 8‑nibble address field from 0.00 to 0.81 exact, using a coarser token budget than 2×2 tiling, and that a second pre‑registered field shows 63% contamination and is demoted. The model’s B6/B7 tool‑id remains at floor, and code/checkpoints are available on Hugging Face.
By Juan S. Santillana
EGGROLL replaces dense Gaussian perturbations in evolution strategies with low‑rank Gaussian products, enabling practical optimization of large language models while maintaining exactness on quadratic objectives. The paper analyzes the mean update field, error bounds, and shows that rank‑one perturbations add only a small variance penalty compared to dense ES. A new leave‑one‑out estimator, LOO‑ROLL, further reduces estimator MSE and improves post‑training performance on transformer blocks and GSM8K accuracy.
By Ege C. Kaya, Abolfazl Hashemi
The paper introduces a new benchmark for vision‑language models that tests their ability to decide whether to answer a physics question immediately or to request additional experimental evidence. Each problem presents one measurement image and four possible physical worlds defined by two masses and two values of another property; the model must either stop and answer or choose the cheapest experiment that resolves the question. Across six open models and 144 parameter sets, the models almost always repeat the same action even when the optimal choice changes, and only a single model gets both decisions correct on 5.9% of cases.
By Sourajit Saha, Shubhashis Roy Dipta, Nobin Sarwar, Shaswati Saha, Yuxuan Jiang, Siyuan Li, Qiheng Wang
The paper introduces T1, a 122‑billion‑parameter Mixture‑of‑Experts model trained with reinforcement learning to perform long‑horizon terminal tasks such as coding and scientific discovery. T1 operates a real shell in a cloud sandbox, making over 300 tool‑call turns per task and receiving rewards from task‑specific verifiers. The authors detail a training recipe that includes aggressive warm‑starting, TITO construction with drift repair, and rollout‑routing replay, achieving significant performance gains on Terminal‑Bench 2.1 and surpassing GPT‑5.4 and GLM‑5.1 on the Long‑Horizon Terminal Bench.
By Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
E-CONAN introduces Arabic textual entailment and natural inference benchmarks comprising two datasets: E-CONAN-2 (2-way RTE) and E-CONAN-3 (3-way NLI). The datasets are built from automatically-translated pairs, human-validated machine translations, hand-crafted pairs from Arabic teaching books, and rumor-containing news headlines. The authors evaluated nine multilingual pretrained models and five large language models on these benchmarks, demonstrating that E-CONAN offers a more diverse and robust assessment than existing datasets like XNLI and ArNLI.
By Khloud AL Jallad, Nada Ghneim, Ghaida Rebdawi
LOCUS is a post‑training technique that selects a task‑aware low‑rank adaptation subspace to reduce token generation cost while preserving a model’s preference alignment. By updating only a tiny fraction of parameters (0.24–0.28 %) on 3 B‑parameter backbones, LOCUS cuts continuation length by up to 39.84 % on Pythia‑2.8B and 14.87–17.58 % on Qwen2.5‑3B, without materially affecting internal preference diagnostics. The method demonstrates that careful subspace selection can make large language models more token‑efficient without sacrificing utility.
By Dongfang Zhao
The paper introduces a multi‑signal pipeline for detecting hallucinations in large language models, combining fine‑tuned DeBERTa‑v3 classification, Monte Carlo Dropout uncertainty, and temperature‑scaled calibration. On the HaluEval benchmark it achieves high performance (F1 = 0.915, AUROC = 0.977) across QA, summarization, and dialogue, and shows that 25 % of training data yields 77 % of full‑data performance. The authors also demonstrate that applying Direct Preference Optimization to a Qwen2.5‑0.5B generator cuts hallucination rates from 85.5 % to 37.7 %, and that domain‑specific fine‑tuning (PubMedBERT on SciFact) outperforms general‑domain models for biomedical text.
By Varun Teja Chundru, Debasmita Biswas
The paper introduces LOGIC (Logit‑Space Integration for Contextual Biasing), a new framework that injects contextual entity information directly into the decoding layer of Speech Large Language Models, bypassing the limitations of prompt‑based methods. LOGIC operates with constant‑time complexity regardless of the size of the entity list, and experiments with the Phi‑4‑MM model across 11 multilingual locales show an average 9% relative reduction in Entity WER while adding only a 0.30% increase in False Alarm Rate.
By Peidong Wang, Jian Xue, Jinyu Li
OpenResearcher is a fully open, reproducible pipeline for generating long‑horizon deep research trajectories that interleave search, evidence aggregation, and multi‑step reasoning. It decouples corpus bootstrapping from trajectory synthesis and runs the search‑and‑browse loop offline using three browser primitives over a 15M‑document corpus. Using GPT‑OSS‑120B as a teacher, the pipeline produced over 97K trajectories, enabling a 30B‑A3B model to achieve 54.8% accuracy on BrowseComp‑Plus and providing insights into pipeline design through controlled analysis.
By Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, Wenhu Chen
The paper introduces a training‑free, alignment‑free method for corporate intelligence that uses deterministic sparse seed vectors to hash word strings into a fixed high‑dimensional basis. By accumulating these seed vectors across sentence contexts, the authors create corpus‑specific semantic signatures that enable rapid document comparison, issuer fingerprinting, vocabulary shift tracking, and thematic sentence extraction—all on standard CPU hardware. Applied to a multi‑year set of SEC filings, the approach reveals distinct semantic profiles for major corporate events such as Boeing’s 737 MAX crisis, Intel’s supply‑chain disruptions, and Bunge’s acquisition of Viterra, with each profile traceable to its source sentences without any domain‑specific training or LLM inference.
By Jean-Fran\c{c}ois Delpech
The paper discusses recursive self‑improvement (RSI) for AI, describing how systems can use experience and feedback to make lasting enhancements to both their abilities and their future improvement processes. It introduces the Headroom‑Closed Index (HCI) to expose limitations in current large language models and outlines a development roadmap for RSI, progressing from autonomy in execution to full recursive meta‑improvement. The authors analyze RSI in various contexts such as scientific discovery, embodied intelligence, and software engineering, noting differing requirements and development speeds, and they connect RSI research to practical industry applications while highlighting key challenges to achieving genuine RSI.
By Yi Duan, Ying Liu, Zirui Tang, Haodong Chen, Jun Zhou, Yumou Liu, Bangrui Xu, Yukai Wu, Sidi Chen, Yuhan Zhou, Haoyu Wang, Xiaoyou Yu, Shaokun Han, Xuzhou Zhu, Le Zhou, Bolin Lu, Wei Zhou, Jiachen Liu, Nuozhou Fang, Jiaxin Tian, Ruoyu Chen, Yuxuan Li, Kai Zuo, Kaiyan Zhang, Jiantao Qiu, Conghui He, Guoliang Li, Bowen Zhou, Zhiyuan Liu, Zhoufutu Wen, Jihua Kang, Xuanhe Zhou, Fan Wu
The paper introduces LLM-Anchored Paralinguistic Enrichment (LAPE), a method that enhances language model representations with speech-based paralinguistic cues such as pauses and word elongations. LAPE incorporates three innovations: prosodic event textualization, lexico-prosodic unitization and chunking, and text-anchored paralinguistic fusion using NormGate. Evaluations on the ADReSS and ADReSSo datasets show that LAPE achieves state‑of‑the‑art performance in detecting Alzheimer’s disease from speech.
By Xiao Wei, Yuqin Lin, Yaru Cao, Jinyu Li, Bin Wen, Kai Li, Yueying Chen, Longbiao Wang, Jianwu Dang