The paper introduces StratCBT, a new dataset of 9,688 psychological counseling sessions with 256K utterances, each counselor response aligned to one of eight Cognitive Behavioral Therapy (CBT) strategies. It was created by modeling clients’ negative thoughts and generating high‑quality conversations through self‑chat, using realistic sessions for guidance. Experiments show that strategy‑aligned generation improves professional and effective counseling when evaluated with large language model‑simulated clients.
By Zimu Wang, Yiwen Jiang, Xiangyu Zhao, Yaling Shen, Jiahe Liu, Stephanie Fong, Maxmartwell H Cheng, Guilherme C Oliveira, Anh Nguyen, Robert Desimone, Barnaby Nelson, Dominic Dwyer, Zongyuan Ge
The paper investigates why Word-in-Context (WiC) remains difficult for language models, suggesting that the lack of an explicit sense inventory contributes to the challenge. By evaluating open LLMs on both WiC and traditional Word Sense Disambiguation (WSD) tasks, the authors find that providing candidate senses—akin to WSD—consistently improves WiC performance. Human evaluation indicates that many WiC errors stem from label ambiguity or mismatched sense boundaries, with models often over‑discriminating senses and making overly fine‑grained distinctions.
By Yi Zhou, Kiamehr Rezaee, Danushka Bollegala, Mohammad Taher Pilehvar, Jose Camacho-Collados
The paper investigates how privileged information—such as a teacher’s full solution or reasoning trace—affects on‑policy self‑distillation (OPSD) in language models. Using the AMPLE‑Math benchmark, the authors compare distillation with and without extra teacher views, finding that reference‑free distillation explains most gains for Qwen3‑1.7B, while additional references provide modest benefits, especially for polished solutions. The study also shows that the impact of privileged data depends on the student’s training regime and that altering token‑level supervision can leave student behavior largely unchanged.
By XiuYu Zhang, Wei Chow, Junfeng Fang, Zhenkai Liang, Tat-Seng Chua
The paper introduces On‑Demand Attention (ODA), a decoding strategy that lets pretrained language models decide when to use global attention based on a lightweight recall head. ODA keeps the original model weights unchanged, only training the recall head, and can be implemented with GPU‑side conditional execution to reduce global reads. Experiments on Qwen, Gemma, and hybrid‑attention models show that ODA largely recovers performance lost by local attention while cutting the number of global attention operations.
By Haibo Feng, Ruiqi Liang, Hanyang Peng, Shiqi Yu
The study investigates how limited reading capacity and claim wording influence consensus outcomes in language‑model networks. By modeling message capacity as the number of messages an agent reads, the authors show that when agents read fewer than about 6.4 messages on average, a wrong consensus becomes unreachable. However, the wording of a claim—its inherent threshold—can override this effect, leading to incorrect consensus even when most agents start correct.
By Makoto Fukushima
GeLaCo is an evolutionary method for compressing large language models by collapsing layers through parametrized weight merging. It uses population-based search with a fitness function that balances similarity of residual updates and language modeling KL divergence, enabling both single and multi-objective compression. The approach yields Pareto-optimal trade-offs between compression and quality, outperforming existing methods in perplexity and generative evaluations.
By David Ponce, Thierry Etchegoyhen, Javier Del Ser
The paper reports the first large‑scale empirical comparison of AI‑agent and human online communities, analyzing 73,899 Moltbook and 189,838 Reddit posts across five matched communities. It finds that Moltbook shows extreme participation inequality (Gini = 0.84 vs. 0.47) and high cross‑community author overlap (33.8% vs. 0.5%). Linguistically, AI‑generated content is emotionally flattened, more assertive than exploratory, and socially detached, leading to community‑level homogenization that is largely a structural artifact of shared authorship. At the individual level, AI agents are more identifiable than human users due to outlier stylistic profiles amplified by their extreme posting volume.
By Agam Goyal, Olivia Pal, Hari Sundaram, Eshwar Chandrasekharan, Koustuv Saha
The paper introduces Redemption Score (RS), a multi‑modal evaluation framework for image captioning that combines three complementary signals: Mutual Information Divergence for global image‑text alignment, DINO‑based perceptual similarity of cycle‑generated images for visual grounding, and LLM text embeddings for contextual similarity to human references. RS fuses these signals to provide a more holistic assessment, achieving a Kendall‑τ of 58.42 on Flickr8k and outperforming most prior methods. The framework demonstrates consistent performance across Conceptual Captions and MS COCO, offering a robust evaluation that captures both visual accuracy and text quality.
By Ashim Dahal, Ankit Ghimire, Saydul Akbar Murad, Nick Rahimi
GAPrompt++ is a multi-granular geometry-aware prompting method designed to adapt pre-trained 3D vision models to downstream tasks efficiently. It introduces a Point Shift Prompter for multi-scale geometric feature extraction, a Keypoint Prompter for local geometric saliency, and a Prompt Propagation mechanism to embed these cues throughout the model hierarchy. Experiments demonstrate that GAPrompt++ outperforms other prompting-based PEFT methods and even surpasses full fine-tuning while using less than 2% trainable parameters, and the authors provide two new challenging benchmarks for future research.
By Zixiang Ai, Zhenyu Cui, Yufei Guo, Wenwen Qiang, Lei Chen, Jiwen Lu, Jiahuan Zhou
PART is a transformer-based framework that performs 3D part retrieval and assembly by selecting appropriate parts from a library and predicting their 6‑DoF poses to reconstruct a target shape. It tackles the combinatorial search space, variable‑length outputs, and continuous pose estimation by formulating the task as set prediction and jointly training with a segmentation‑enhanced optimization module. The authors also introduce a large‑scale dataset of over 80,000 shapes and demonstrate that PART generalizes to scene layouts, image targets, and real‑world scans.
By Ruchao Bao, Wenzheng Wu, Chucheng Xiang, Zhongyuan Liu, Yuan Liu, Jinxin Dong, Ligang Liu, Ziqi Wang
The paper introduces the concept of xeno-interpretability, which studies internal distinctions in large language models that lack corresponding human concepts. It distinguishes between human‑interpretable and xeno‑semantic spaces, showing that LLMs possess a far larger internal representational space than can be captured by finite human descriptions. The authors propose an empirical program to identify and characterize these xeno‑representations, noting their potential to influence model behavior in ways that are not fully visible through human‑readable communication.
By F. Pierucci, M. Bracale Syrnikov, M. Prandi, M. Galisai, F. Giarrusso, P. Bisconti
Paint-Anything introduces a unified hex-prompt interface that allows users to specify any 24‑bit hex color for both image generation and editing. The method trains on a new Paint‑500K dataset created from real images with object grounding, perceptual color labeling, and editing‑pair synthesis, and supplements this with pure‑color anchors to address shadow‑induced color inaccuracies. Evaluated on the newly proposed Any Color Benchmark (ACBench), Paint‑Anything achieves significant improvements over the base FLUX.2‑4B model, boosting T2I and editing scores by 85.3 % and 28.3 % respectively, and outperforms competing methods on the CompColor metric.
By Ji Xie, Dewei Zhou, Xinyu Huang, Zhennan Chen, Xun Wang
Video DeltaNet (VDN) introduces a hybrid attention mechanism for livestream video generation, combining local Softmax attention with a bidirectional linear memory branch called Video Delta Attention (VDA). VDA updates memory once per frame, integrating spatial tokens, while separate output projections and learnable gates balance the two branches. Applied to MiniMax H3, VDN achieves a 14.5× speedup over the dense baseline, completing 14.3‑second, 768p video denoising in 6.70 seconds on eight NVIDIA B200 GPUs.
By Haocheng Xi, Yiming Xie, Hexu Zhao, Yiwen Zhang, Michael Liu, Thomas Creavin, Kurt Keutzer, Xiuyu Li, Zhaoyang Lv, Chenfeng Xu, Haiwen Feng
The paper introduces a six‑dimension prompt‑side structural‑complexity index to assess code‑generation reliability before a model generates output. Using 5,000 Python prompts and 21 large language models, the authors find that pass rates exhibit a non‑monotonic breakpoint around a composite score of 13.75, with task‑type and construction‑frame adjustments shifting this threshold. The study also reports high inter‑rater reliability (ICC = 0.872) and demonstrates that the index can predict failure likelihood without relying on output correctness.
By Michael Hernandez, Tian Zhao
The paper "What Do We Expect from LLMs? Mapping the Design of LLM Benchmarks" analyzes 14,767 arXiv submissions from 2022 to 2026 that introduce or update evaluation resources for large language models. It systematically maps changes in target systems, domains, evaluation materials, conditions, and scoring mechanisms, revealing a growing emphasis on action, interaction, and professional applications. The study also notes uneven development in model participation, with LLM-based scoring increasing in both agent and non-agent groups, while model-generated materials do not show a comparable rise.
By Chao Wang (Independent Researcher)
The study examines how large language model (LLM)–based graphical user interface (GUI) agents respond to digital nudges. Using Dual‑Process Theory, researchers tested 3,600 agents across six frontier models in an online shopping experiment and found that the agents were susceptible to both automatic (Type 1) and reflective (Type 2) nudges. The agents’ reasoning configuration moderated these effects in opposite directions: extensive reasoning reduced susceptibility to automatic default nudges but increased susceptibility to reflective social‑influence nudges, with the effect systematically varying by model scale.
By Haya Halimeh, Sascha Kaltenpoth, Kevin B\"osch, Oliver M\"uller
The paper presents a three‑phase pipeline that shifts semantic user profiling from per‑user to per‑transaction‑pattern inference, dramatically reducing LLM inference costs. The Resolve phase maps item names to abstracts, Profile generates categorical labels, free‑text attributes, and prevalence estimates per pattern, and Tag clusters these attributes into a searchable database. Evaluated on a public e‑commerce corpus, the pattern‑based approach matches the performance of a per‑user LLM while enabling deployment at a major Japanese bank with tens of millions of users and a three‑order‑of‑magnitude reduction in inference targets.
By Ryota Mitsuhashi, Tetsuro Morimura, Hirotake Ito
The paper presents a new calibrated confidence framework for straight‑through processing (STP) of financial documents, decomposing confidence into perception, layout, and validation channels and applying conformal risk control. Using Vision‑Language Models (Qwen3.6‑27B and Gemini‑3.1‑Flash‑Lite) on three public datasets, the method raises AUROC from 0.54‑0.74 to 0.90‑0.99 and increases auto‑approval rates to 49‑72% while keeping empirical error below the target. This approach enables reliable, large‑scale STP without human review.
By Yichao Jin, Yushuo Wang, Yuxuan Han, Kwan Ching Yee Sonia, Weiyang Song, Chiu Jin-Chun Kent, Wong Chong Hwee, Wong Tiong Kiat, Kenneth Zhu Ke, Jingyuan Zhao
GAVEL is a framework that uses an explicit graph world model to verify and repair long‑horizon plans generated by large language models (LLMs). The graph encodes object relations, action pre‑conditions and effects, and probabilistic beliefs about unobserved object locations, allowing the system to predict action outcomes, detect violations, and repair them before execution. In experiments on BEHAVIOR‑1K, GAVEL boosts single‑task success from 41.2 % to 91.8 % and multi‑task success from 19.9 % to 92.6 %, while also reducing travel distance by about 5.4 % compared with a static variant.
By Ruiyang Wang, Hao-Lun Hsu, Swarajh Mehta, Jiwoo Kim, Zhihao Dou, Miroslav Pajic
CliniCIRCA is a modular large‑language‑model framework that reconstructs longitudinal mental‑health patient journeys from raw electronic health record narratives. It temporally classifies clinical events in unstructured discharge summaries without explicit timestamps, producing 15,891 tagged events from 52 summaries and correcting 629 errors to create verified gold‑standard timelines. The framework then generates temporally grounded summaries, compressing each source by 1.52×, and scales to produce 1,000 silver‑standard timelines for training, showing that instruction tuning improves event extraction, temporal tagging, and summarization across models.
By Aiwei Ivy Zhang, Nimra Ishfaq, Mohit Chandra, Santiago Alvarez Lesmes, Adam Coscia, Khatiya Chelidze Moon, Xiaohan Ding, Munmun De Choudhury