The paper presents a systematic study of token consumption in AI agents performing coding tasks. It finds that agentic tasks are far more expensive—about 1000 times more tokens than code reasoning or chat—primarily due to input tokens, and that token usage varies wildly, with accuracy peaking at moderate costs. Models differ significantly in efficiency, and current frontier models cannot reliably predict their own token usage, often underestimating it.
By Longju Bai, Zhemin Huang, Xingyao Wang, Jiao Sun, Rada Mihalcea, Erik Brynjolfsson, Alex Pentland, Jiaxin Pei
The study investigates why large language models (LLMs) show different triage performance when answering clinician‑authored vignettes in multiple‑choice versus free‑text formats. Using sparse‑autoencoder features on Gemma 3 and Qwen3 models, the authors find that medical information is encoded similarly in both formats, but at the decision token the multiple‑choice scaffold dominates, with over 91% of attribution coming from scaffold‑peaking features. The effect varies by model, and shuffling option order eliminates simple positional bias, suggesting the format influence is tied to answer selection rather than earlier case processing.
By David Fraile Navarro, Berardino Como, Jialei Sheng, Soundariya Ananthan, Shlomo Berkovsky
The paper proposes a training‑free detector that uses sentence‑level context sensitivity to identify unsupported content in retrieval‑augmented generation (RAG) answers. By re‑scoring each sentence with full context, no context, and each chunk removed, the method flags the chunk whose removal most reduces a sentence’s likelihood as the likely source. Evaluated on RAGTruth, TofuEval, and RAGBench, the detector outperforms answer‑level faithfulness scores, achieving AUCs up to 0.745 and matching per‑chunk fact‑checkers while using only a fraction of the compute required by large‑language‑model judges.
By Mohamed Aly Bouke
Latent-MOPD is a new on‑policy distillation method that allows a single large language model student to learn from multiple specialist teachers by using both the teachers’ output distributions and their hidden state representations. The approach selects late‑layer targets based on teacher‑student relationships, bridges hidden width differences with a shared projection, and groups updates by domain, enabling gradual shift from representation to token supervision. Experiments show that Latent‑MOPD outperforms token‑only, representation‑only, and uniform‑averaging baselines across nine benchmarks in math, code, and logic, and even surpasses the best individual teacher on most tasks.
By Zhengyu Fang, Seoyeon Hong, Jie Yang, Muyang Li, Koyoshi Shindo, Brandon Joseph Lwowski, Jing Li
The paper introduces the Radio Alternating Operator Ansatz (RAOA), a recurrent computing architecture that uses programmable radio propagation as a computational depth. RAOA alternates energy‑derived problem updates with mixing updates over a persistent latent state, enabling repeated compositional passes even with the same learned controls. Experiments on discrete optimization, simulation of programmable propagation, and adaptation of pretrained language models show that RAOA can improve solution quality, handle higher‑order interactions, and adapt models with performance close to shallow MLPs, all within a single recurrent framework.
By Toshiaki Koike-Akino
The paper investigates how Mixture-of-Experts (MoE) Particle Transformers perform on the 188-class JetClass-II jet classification task. By varying expert count, routing capacity, top‑K, and auxiliary loss, the authors find that top‑1 MoE models can surpass dense baselines with similar nominal compute, but adding more experts yields diminishing accuracy gains. Activating multiple experts per token improves predictions at higher computational cost, and routing analyses reveal that expert assignments correlate with particle identity and kinematics, though this correlation does not consistently predict performance.
By Kaushik Pendiyala, Haris Zia, Trevin Lee, Timothy Legge, Alejandro J. De Leon, Zihan Zhao, Aaron Wang, Abhijith Gandrakota, Jennifer Ngadiuba, Richard Cavanaugh, Javier Duarte
The paper investigates how well a simple spatially shared linear map can predict changes in a model’s internal feature space caused by various image manipulations, including geometric, photometric, occlusion, and diffusion-generated semantic edits. Experiments across ConvNeXt, SwinV2, and DINOv3 show that this linear operator often performs nearly as well as more complex, higher‑capacity probes, especially in deeper layers of supervised backbones. The study concludes that a shared linear map is frequently sufficient to capture diverse image edits, with its leading singular components encoding semantic content and higher‑rank components refining details.
By Elias Krey, Nils Neukirch, Nils Strodthoff
Power‑SMC is a training‑free sampling method that targets the sequence‑level power distribution used to sharpen large language model outputs for reasoning tasks. It maintains multiple candidate sequences, assigns importance weights based on how well each matches the power distribution, and periodically prunes low‑scoring candidates to keep inference latency close to standard decoding. Experiments on MATH500, GSM8K, GPQA, and HumanEval show that Power‑SMC matches or surpasses Metropolis–Hastings sampling in accuracy, preserves output diversity, and can accelerate inference speed by up to 17.6×.
By Seyedarmin Azizi, Erfan Baghaei Potraghloo, Minoo Ahmadi, Souvik Kundu, Massoud Pedram
The paper introduces <flowreader>, a method that casts evidence selection for long multimodal documents as a minimum‑cost flow problem over a multimodal content graph. It uses spectral decomposition to identify latent query‑relevant aspects and allocates a fixed evidence budget proportionally to their spectral energy, ensuring aspect coverage without a language‑model planning call. On the VisDoMBench benchmark with Qwen3‑VL‑32B, <flowreader> achieves the highest macro accuracy (68.9%) and outperforms prior systems by 2.7 points, while using fewer content nodes and reader tokens.
By Ambuj Mehrish, Sebastiano Vascon
The paper introduces XiangqiBench, an executable benchmark for evaluating large language model agents in Chinese chess. It tracks 8,568 multi‑turn trajectories from 12 frontier LLMs, revealing that metrics such as the Conversion Gap, Consistency Gap, and Simulation Gap overstate true closed‑loop success. The study shows that simply naming the correct move is insufficient; agents must reliably carry a plan through to a verified outcome against an opponent.
By Yekun Chai, Qiwei Peng, Haoyi Xiong
The paper investigates how post‑softmax attention weights behave across different language models by marginalizing over query positions to create a joint token‑head space. Across more than 60 diverse LLMs, token‑wise marginal attention reveals a text‑intrinsic signal that is conserved across models, while head‑wise marginal attention yields a model‑private signature that is consistent across documents. This insight enables a practical method for estimating per‑head budgets for key‑value cache eviction, achieving competitive performance on standard benchmarks without per‑document recomputation or training.
By Valentino Maiorca, Walter Nelson, Francesco Locatello
The paper introduces Collective Bias Mitigation (CBM), a framework that reduces bias in large language models by learning fine‑grained model behavior and enabling knowledge sharing among diverse LLMs. CBM systematically explores how to select and organize distinct models to produce fairer responses, outperforming standalone baselines. Experiments demonstrate significant bias reduction, with the Committee topology balancing effectiveness and inference cost.
By Mingzhe Du, Luu Anh Tuan, Xiaobao Wu, Yichong Huang, Yue Liu, Dong Huang, Huijun Liu, Bin Ji, Jie M. Zhang, See-Kiong Ng
StanceEval 2026 is the second edition of a shared task on stance detection in Arabic social media, where systems must classify a tweet’s stance toward a target as Favor, Against, or None. The event featured two tracks: Track 1 tests cross‑target transfer on thematically related topics (Women Driving vs. Women Empowerment), while Track 2 evaluates cross‑domain transfer to entirely unseen targets (E‑Cars and Trimester System). With 80 registered teams and 30 submissions, top systems achieved $F_{avg2}$ scores of 0.8994 (Track 1) and 0.9400 (Track 2), surpassing baseline performance and highlighting challenges such as target polarization and dialectal nuance.
By Rasha Albalawi, Nuha Albadi, Hamzah Luqman, Asma Yamani, Maram Kurdi, Saad Ezzini, Ahmed Ashraf, Maged Al-Shaibani, Nora Alturayeif
The paper investigates how large language model (LLM) agents exhibit source preference when selecting items in end‑to‑end search tasks across three domains. Across 12 agent models, each consistently favors certain sources and avoids others, even when items from disfavored sources better meet the user’s requirements. The study shows that revealing or hiding source identifiers influences selection, and that training rewards or missing information can create or mitigate these preferences.
By Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song, Yohan Jo
The paper introduces a single‑pass uncertainty head for detecting hallucinations in Persian medical language models. By adapting the LLM Uncertainty Head framework to Aya‑Expanse‑8B and two Gaokerena backbones, the authors build claim‑level hallucination datasets and train lightweight heads that achieve PR‑AUCs of 0.4820 and 0.4652 and ROC‑AUCs of 0.7852 and 0.7810. These heads operate without retrieval or repeated sampling during inference, offering a cost‑effective approach to hallucination detection in Persian medical LLMs.
By Mehrdad Ghassabi, Pedram Rostami, Hamidreza Baradaran Kashani, Sadra Hakim, Audrina Ebrahimi
The paper evaluates six structured decision models for hate‑speech moderation across four datasets, comparing them to specialized, zero‑shot, commercial, and supervised baselines. It investigates whether providing dataset definitions or decomposing them into multiple questions improves classification, and measures latency and cost. Results show that commercial LLMs outperform decision models on only one dataset, definitions alter up to 28% of predictions without consistent gains, and decomposition helps in only 20% of cases, yet the best hosted decision model approaches the best commercial LLM’s performance with roughly 97% lower inference cost.
By Demetris Paschalides, George Pallis, Marios D. Dikaiakos
The paper "Code2Math: Can Your Code Agent Evolve Math Problems Through Exploration?" explores how code agents can autonomously generate more complex variations of existing math problems. It introduces a multi‑agent framework that evolves problems while ensuring they remain solvable and increasingly difficult. Experiments show that with sufficient exploration, code agents can produce new, structurally distinct, and challenging problems, demonstrating a scalable method for creating high‑difficulty mathematical reasoning tasks.
By Dadi Guo, Yuejin Xie, Qingyu Liu, Weixian Huang, Jiayu Liu, Zhiyuan Fan, Qihan Ren, Shuai Shao, Tianyi Zhou, Jianjie Feng, Wenze Su, Yujiu Yang, Dongrui Liu, Yi R. Fung
The paper introduces Mawqif-XT, a new Arabic benchmark dataset comprising 996 manually annotated tweets from three public targets: Women Driving, E-Cars, and Trimester System. Each tweet is labeled for stance, sentiment, and sarcasm following the Mawqif annotation scheme, and the dataset is intended as a held‑out evaluation set to test cross‑target generalization. Baseline results are provided using Arabic and multilingual transformer models as well as zero‑shot large language models, enabling reproducible evaluation alongside the original Mawqif dataset.
By Rasha Albalawi, Nuha Albadi, Hamzah Luqman, Maram Kurdi, Saad Ezzini, Asma Yamani, Ahmed Ashraf
The paper introduces GAW-PO, a gradient‑aligned token reweighting technique for Direct Preference Optimization (DPO). It assigns weaker penalties to rejected tokens whose gradients align with preferred behavior, while maintaining stronger penalties for conflicting tokens. Across 11 benchmarks in mathematics, reasoning, coding, and question answering, GAW‑PO outperforms standard DPO and other baselines, and remains robust when the DPO regularization parameter is reduced.
By Andreea Dutulescu, Stefan Ruseti, Mihai Masala, Traian Rebedea, Mihai Dascalu
The paper introduces a simulation‑grounded vision‑language model (VLM) framework for wildfire monitoring that converts 2D wildfire simulations into labeled video episodes using a fixed Blender mapping to create low‑detail 3D proxies. These proxies, along with controllable video generation, provide a multimodal memory that a training‑free multi‑agent VLM system uses to retrieve reference episodes, reconcile visual and memory‑based predictions, and generate structured wildfire reports. The system achieves 77.3% accuracy on six simulator‑derived report fields, outperforming direct VLM querying and text‑only memory baselines.
By Duowen Chen, Yuchen Sun, Zhiqi Li, Yuxuan Liao, Sinan Wang, Bart van Bloemen Waanders, Bo Zhu