The study evaluates six frontier language models on a two‑agent <log(N)>‑Questions game using Wikipedia lead paragraphs. In each game a questioner must identify a target paragraph with exactly <log2 N> yes/no questions, while an answerer only sees the target and the question and replies with a single word. Across 408 games, the models perform similarly, with Claude Opus 5 winning 28 of 68 games and the top five models showing only marginal differences; win rates decline sharply with larger document sets, following a reliability parameter of 0.928 per question.
"whyItMatters":"The results reveal how well language models can communicate under information asymmetry, highlighting that even top models struggle to extract a full bit per question and that reasoning token usage does not strongly predict success."
By Peter Potash
PANORAMA introduces a new panoptic grounded captioning framework that jointly generates detailed image captions and associates each phrase with precise pixel-level masks. The authors create PanoCaps, a human‑annotated benchmark with dense captions and near‑complete pixel coverage, and propose a phrase‑mask matching protocol with a generalized Panoptic Quality metric. PANORAMA conditions a pretrained segmenter on contextualized phrase representations, learns to select appropriate masks, and achieves state‑of‑the‑art grounding performance on PanoCaps and other pixel‑level tasks.
By Sara Pieri, Evangelos Kazakos, Shizhe Chen, Josef Sivic, Cordelia Schmid
The paper introduces JMLLM, a multimodal jailbreaking approach that targets text, visual, and auditory inputs to expose vulnerabilities in large language models. It also presents TriJail, a new dataset containing jailbreak prompts across all three modalities. Experiments on TriJail and AdvBench show higher attack success rates and lower time overhead compared to existing methods.
By Yanxu Mao, Peipei Liu, Tiehan Cui, Zhaoteng Yan, Congying Liu, Datao You
The paper investigates how adjectival modifiers affect the semantic plausibility of events, using the Adept benchmark of 16,000 English sentence pairs that differ by a single adjective. Experiments show that sentence transformers, despite being conceptually suited to the task, underperform compared to models like RoBERTa. The authors provide an error analysis and discuss the implications of their findings for future work on balancing training and test data.
By Anna Golub, Beate Zywietz, Annerose Eichel
The paper introduces the first reproducible taxonomy for programming languages based on resource availability, categorizing 646 languages into four tiers. It finds that a small fraction (1.9%) of high-resource languages (Tier 3) generate the majority (74.6%) of tokens in major corpora, while the majority of languages (71.7%) are scarce and contribute only 1.0% of tokens. Statistical analysis confirms the extreme and systematic imbalance across tiers.
By Nishat Raihan, Christian Newman, Marcos Zampieri
The paper introduces CROP, a method for selective on‑policy distillation that prioritizes token‑level supervision based on task relevance. CROP uses paraphrase‑calibrated counterfactual sensitivity to measure how much each response token depends on the semantic content of the input, constructing validated original‑paraphrase‑counterfactual triplets for each prompt. Experiments in two teacher‑student settings show that CROP outperforms other selectors, improving aggregate performance by 1.92 and 2.96 points.
By Enhan Li, Junhao He, Hongyang Du
RT-SEMamba is a fully causal speech enhancement model that uses causal time‑frequency Mamba blocks instead of Transformer‑based architectures, allowing efficient long‑form inference with a fixed‑size recurrent state. The authors introduce a progressive knowledge distillation strategy that compresses an 8‑layer teacher into a single‑layer student by jointly distilling spectral outputs and intermediate representations. On the Voicebank‑DEMAND benchmark, the 8‑layer model achieves 3.32 PESQ under a 25 ms latency constraint, while the distilled 1‑layer student improves from 3.06 to 3.18 PESQ, maintains the same steady‑state real‑time factor, and runs 2.64× faster than the teacher.
By Rong Chao, Sung-Feng Huang, Moreno La Quatra, Sabato Marco Siniscalchi, Wen-Huang Cheng, Szu-Wei Fu, Yu Tsao
Position Anchor Tuning (PAT) is a parameter‑efficient fine‑tuning method for pre‑trained point cloud transformers that improves inference efficiency. PAT reduces the computational cost of multi‑head attention and feed‑forward networks by using token aggregation‑expansion pairs: a token aggregation module (TAM) selects representative tokens based on 3D position anchors, and a token expansion module (TEM) propagates the learned representations back to the full token set. Combined with base‑sharing low‑rank adaptation (BSLoRA) for the TAMs, PAT achieves performance comparable to state‑of‑the‑art methods while requiring fewer trainable parameters and lower computational overhead.
By Zheng Liu, Xin Gao, Jinchao Zhu, Gao Huang
The paper introduces Vocabulary‑Guided Gait Recognition (Gait‑World), a new paradigm that leverages human vocabularies and Vision‑Language Models (VLMs) to better understand and recognize gait. It proposes the α‑Gait model, which uses a Vocabulary Relation Mapper and a Gait Fine‑grained Detector to align vocabulary prompts with gait features, addressing the challenge of heterogeneous embedding spaces. Experiments on multiple gait datasets demonstrate the potential of incorporating VLM‑derived vocabulary information into gait recognition.
By Panjian Huang, Saihui Hou, Chunshui Cao, Xu Liu, Yongzhen Huang
STRADAViT is a self‑supervised continued‑pretraining framework that adapts Vision Transformer (ViT) backbones for radio‑astronomy image analysis. It curates mixed‑survey data, generates radio‑astronomy‑aware training views, and initializes encoders with ViT‑MAE, optionally adding register tokens. Evaluations on three morphology benchmarks (MiraBest, LoTSS DR2, and Radio Galaxy Zoo) show that a register‑based two‑stage checkpoint improves linear‑probe Macro‑F1 scores over the ViT‑MAE baseline and enhances fine‑tuning on MiraBest and RGZ DR1, though performance on LoTSS DR2 fine‑tuning declines; these differences are statistically significant.
By Andrea DeMarco, Ian Fenech Conti, Hayley Camilleri, Ardiana Bushi, Simone Riggi
The paper introduces agentic-eCAL, an extension of the Energy Cost of AI Lifecycle metric to evaluate multi‑agent AI workflows across the edge‑cloud continuum. By combining a two‑rate energy model with OSI‑layer transport analysis, the authors quantify that inter‑agent text transfer accounts for only 0.25% of total workflow energy, highlighting that the main energy cost lies in additional inference and context processing triggered by communication. The study uses extensive GPU benchmarks on NVIDIA A100/H100 with 16 open‑weight models and 8 orchestration topologies to validate the metric and explore placement implications.
By Carolina Fortuna, Vid Han\v{z}el, Tim Strnad, Bla\v{z} Bertalani\v{c}
The paper introduces Mahalanobis-Ensemble Decoding (ME-Decoding), a new framework for Large Language Model decoding that treats candidate token selection as an ensemble pruning problem. It uses a Mahalanobis distance-driven objective to promote semantic diversity while maintaining high probabilities, employing a token similarity matrix built with an adaptive-bandwidth kernel over token embeddings. An efficient greedy algorithm with near-linear complexity and theoretical guarantees makes ME-Decoding a plug‑and‑play module with negligible inference overhead, and experiments show strong performance across reasoning and generation tasks.
By Dunyao Xue, Chengshuo Du, Zhengbo Wang, Wenlin Dai, Cheng Meng
PentestChain is a ten‑phase automated penetration testing framework that uses a cost‑aware AI cascade, starting with a local 7B‑parameter Ollama model (qwen2.5‑7b) and then free‑tier OpenRouter and Cerebras models, with a rule‑based fallback. It exposes the entire pipeline through a Model Context Protocol (MCP) server that includes eleven tools. The authors evaluate the framework using standard testbeds (AutoPenBench, Cybench subset, PentestGPT 182‑sub‑task benchmark) and report that the local model keeps paid‑API cost at zero while detecting 26 services and enriching 34 CVEs on legacy targets.
By Rushabh Vipulkumar Patel, Dipo Dunsin, Mohammed Almaiah, Mohamed Chahine Ghanem
The paper introduces HOPE, a second‑order pruning method for Mixture‑of‑Experts language models that accounts for cooperative interactions between experts. Unlike first‑order methods such as REAP, HOPE derives an objective that provably bounds pruning error and is shown to outperform baselines across three large MoE models, multiple calibration sets, and diverse benchmarks, especially at high pruning rates and on agentic tasks. The results demonstrate that preserving expert interactions allows aggressive compression with minimal performance loss on complex workloads.
By Alex M. Tseng, Prannay Kaul, Luca Zancato, Wei Xia, Stefano Soatto
This paper reports a fully localized, low‑resource framework that runs a trillion‑parameter biomedical large language model on a single consumer‑grade RTX 4060 laptop (32 GB system memory, 8 GB VRAM) and routine clinical workstations. The system completes an entire tumor‑paired whole‑genome sequencing workflow—from raw FASTQ input to a clinical‑grade full‑variation‑spectrum report—in 18 hours at 30× depth, achieving a 99.62 % F1 score for somatic variant detection and over 99.9 % concordance with an industrial‑standard A100 cluster pipeline. The study demonstrates that adaptive heterogeneous memory scheduling accounts for 71 % of execution time and that model optimization adds less than 9 % of detection error, establishing a low‑cost, high‑accuracy pathway for global primary medical institutions to adopt precision oncology without expensive GPU clusters.
By Rui Xiao, Yili Xu
The paper proposes a Mixture-of-Bottleneck (MoB) framework for video-based multimodal sentiment analysis that treats sentiment as an ordinal regression problem, splitting it into polarity recognition and intensity prediction. MoB assigns modality‑specific latent experts to each sub‑task, learns compact, task‑relevant representations via an information bottleneck, and fuses these experts with a multimodal bottleneck routing module and hard mining strategy. Experiments on four datasets and language models demonstrate that MoB captures fine‑grained intra‑ and inter‑modal dynamics, improving performance and enabling more trustworthy localization of nuanced sentiment signals.
By Ronghao Lin, Qiaolin He, Zefeng Lu, Yichu Liu, Li Huang, Sijie Mai, Haifeng Hu, Yap-peng Tan
Lumen is a pathology vision‑language model that aligns frozen unimodal foundation models (Virchow2 and BioMedBERT) using rank‑4 adapters and projection heads, training only 0.40% of the total parameters on the QUILT‑1M corpus. It achieves the highest mean chance‑corrected balanced accuracy (0.546) across nine zero‑shot patch benchmarks and demonstrates strong performance on lymph‑node metastasis detection, with AUROC scores of 0.964 internally and 0.955 externally. While it ranks third in cross‑modal retrieval, Lumen’s low‑parameter training yields competitive results at both patch and slide levels.
By Kiarash Tajbakhsh, Abdelrahman Faqieh, Michael Jopiti, Javier Garcia-Baroja, Philipp Zens, Branislav Zagrapan, Yuri Tolkach, Martin D. Berger, Aurel Perren, Bastian Dislich, Inti Zlobec, Amjad Khan
The paper studies Graph-Guided Token Merging (G2TM), a module that reduces token count in Vision Transformers. It evaluates G2TM across multiple segmentation frameworks and decoder types, finding that its performance gains are tied to the encoder rather than the decoder. The authors report consistent reductions in GFLOPs (22‑47%) and throughput improvements (up to 74%) on ADE20K, with optimal hyperparameters depending mainly on backbone pre‑training and target dataset.
By Victor Bercy, Martyna Poreba, Michal Szczepanski, Samia Bouchafa
The paper presents a decoupled framework for sim-to-real traffic scene understanding, separating semantic fact extraction from caption generation. It uses a frozen V-JEPA encoder for predictive scene representations and a lightweight Llama-based predictor for VQA, followed by a training-free structured refinement that leverages statistical priors, inter-question relationships, and temporal consistency. The refined facts are then fed to Qwen3-VL-8B to produce pedestrian and vehicle descriptions, achieving top performance on the 2026 AI City Challenge Track 2 benchmark with 87.09% VQA accuracy and an overall S2 score of 60.0853.
By Nguyen Hoai Thuong Bui, Thanh Nguyen Vo, Trinh Tra Giang Nguyen, Ha Duc Bui
GVD (Governed Versioning and Deduplication) is a framework that unifies cross‑document version linking with rule‑level conflict resolution under an auditable update policy. It assigns incoming documents to version families via bidirectional rule alignment, then compares their rules against family memory to detect duplicates, contradictions, asymmetric refinements, and new knowledge, using Counterfactual Span Probing (CSP) to correct neutral misclassifications. In a test on 120 enterprise documents across 59 version families, GVD achieved an F1 of 0.97 for version‑family construction and 0.94 for rule‑level consistency, with CSP improving rule consistency from 0.90 to 0.94.
By Mohammadreza Sediqin, Shivali Dalmia, Sumukha Thoppanahalli, Abhishek Mukherji