The paper introduces Tail‑Corrected Top‑k On‑Policy Distillation (TT‑OPD), a method that improves on existing Top‑k OPD by combining the selected top‑k tokens with a sampled token from the student’s rollout. This hybrid approach recovers the probability mass discarded by limiting to top‑k, yielding an unbiased estimator of the reverse KL divergence gradient while maintaining low computational cost. Experiments show TT‑OPD outperforms other OPD variants in accuracy.
By Linjian Meng, Siyuan Gan, YuHan Li, Xiran Wang, Ziyang Ding, Ditang Gou, Yiming Wu, Zhen Zhao
The paper evaluates nine on‑device named‑entity recognition models ranging from classical taggers to large language models, measuring not only accuracy but also latency and output validity. Using a silver‑gold benchmark derived from an LLM judge panel and a human‑validated corpus, the study shows that encoder‑based models achieve comparable accuracy to a 4 B instruct LLM while being much smaller, faster, and producing no malformed output. Confidence calibration of GLiNER is analyzed, revealing over‑confidence but improved reliability after temperature scaling and thresholding.
By Vinay Kumar Chaganti
The study evaluated six language‑model configurations for generating six distinct random integers between 1 and 49, using 1,200 attempts across four English prompts. Valid tickets were produced in 1,184 responses, but the effective diversity of number frequencies was low (9.9 to 18.0) compared to the expected 46.6–46.7 from independent uniform sampling. The models generated 8–93 distinct unordered tickets, with modal tickets comprising 22.5–68.0% of valid outputs, indicating substantial concentration in the generated lottery numbers.
By Dmitrij \.Zatuchin
MOMAT (Mixture of Multiple Atlases) is a hardware‑enhanced safety framework designed to defend quantized large language models (qLLMs) on edge devices against jailbreak attacks. It uses a collection of semantic atlases—each containing harmful or benign sample clusters and policy templates—to perform domain‑localized Retrieval‑Augmented Generation. A lightweight Mixture of Experts detector evaluates top‑k similarity features retrieved by a Compute‑in‑Memory (CiM) accelerated engine, achieving a 4.69 × 10⁶‑fold speedup and a 2.5 × 10⁵‑fold energy reduction compared to DRAM‑based baselines while matching state‑of‑the‑art defense performance.
By Boyang Li, Bingyu Shen, Weihao Hong, Zhiyuan Jiang, Xinlei Guan, Yan Ma, Miles Q. Li, Yi Sheng, Ruiyang Qin
The paper introduces LOOM, a method for scaling looped mixture‑of‑experts (MoE) Transformers beyond the typical two‑loop limit. LOOM addresses two key obstacles: it stabilizes deep recurrence by bounding residual variance and re‑injecting the input embedding, and it prevents expert selection collapse by using per‑loop routers and a looping residual to maintain computational diversity. Experiments on 100 M–1.7 B parameter models show stable scaling to 9–12 loops, with significant perplexity reductions and zero‑shot accuracy gains under near‑iso‑FLOP conditions.
By Di He, Pengxiang Li, Da Chang, Qingyan Meng, Lu Yin, Shiwei Liu
The article surveys LLM-based agentic reasoning frameworks, presenting a unified formal language that categorizes them into single-agent, tool-based, and multi-agent methods. It reviews application scenarios in scientific discovery, healthcare, software engineering, society, economics, and general-purpose tasks, and compares the distinct features and evaluation strategies of each category. The survey highlights the rapid development of complex agentic systems in real-world contexts.
By Bingxi Zhao, Lin Geng Foo, Ping Hu, Christian Theobalt, Hossein Rahmani, Jun Liu
SONIC‑O1 is a new benchmark designed to evaluate multimodal large language models on audio‑video understanding. It contains 60 hours of 231 clips across 13 real‑world conversational domains, with 4,958 human‑verified annotations and demographic metadata. The benchmark tests open‑ended summarization, multiple‑choice question answering, and temporally grounded reasoning, revealing performance gaps between model families and across demographic groups.
By Ahmed Y. Radwan, Christos Emmanouilidis, Hina Tabassum, Deval Pandya, Shaina Raza
TensorCommitments (TCs) is a lightweight, tensor-native proof‑of‑inference scheme that enables verifiable inference for large language models (LLMs) without requiring the verifier to rerun the model or possess a powerful GPU. By binding each inference to a commitment stored in multivariate Terkle Trees, TCs detect tampering with only a 0.97% overhead for the prover and 0.12% for the verifier on LLaMA2. The approach improves robustness against tailored LLM attacks by up to 48% compared to previous methods that needed a verifier GPU.
By Oguzhan Baser, Elahe Sadeghi, Eric Wang, Nico Vergauwen, Sam Kazemian, Hong Kang, Sandeep P. Chinchali, Sriram Vishwanath
The paper introduces TRACE, a method that removes duplicated text—known as note bloat—from clinical notes by leveraging EHR metadata and frequency-based de‑duplication. Across 5.3 million notes from diverse patient cohorts, TRACE eliminated 47.3 % of chart text while preserving information extraction and prediction performance, with only 0.3–6.6 % of removed content being author‑generated. The authors project that applying TRACE could yield net savings of $1.00 M to $13.58 M over three years at a large academic center, depending on model pricing schemes.
By Jordan L. Cahoon, Chloe Stanwyck, Asad Aali, Rachel Madding, Sulaiman S. Somani, Emma Sun, Yixing Jiang, Renumathy Dhanasekaran, Emily Alsentzer
GENSCRIPT is an inference‑only pipeline that generates synthetic data without training a generative model. It creates a deterministic statistical profile of the source data, uses a language model to infer field semantics and cross‑column constraints, and then compiles these into an executable sampler that works for single‑table, temporal, and relational data. The method builds generators in minutes, samples large datasets quickly, and achieves fidelity comparable to leading methods while preserving key data relationships such as 1‑to‑1 mappings and primary‑foreign key constraints.
By Zilong Zhao, Abdul Raheem, Jiayu Li, Sohei Arisaka, Darius Lim Hong Yi, Milad Abdollahzadeh, Uzair Javaid, Biplab Sikdar
The paper introduces a formal framework for measuring AI propensities—tendencies of models to exhibit particular behaviours—using a bilogistic formulation that identifies an "ideal band" of success probability. It estimates the limits of this band with task‑agnostic rubrics and applies the method to six families of LLMs, showing how shifts in propensity affect task performance. The study finds that propensity estimates from one benchmark predict behaviour on held‑out tasks and that combining propensity with capability metrics yields stronger predictive power than either alone.
By Daniel Romero-Alvarado, Fernando Mart\'inez-Plumed, Lorenzo Pacchiardi, Hugo Save, Siddhesh Milind Pawar, Behzad Mehrbakhsh, Pablo Antonio Moreno Casares, Ben Slater, Paolo Bova, Peter Romero, Zachary R. Tidler, Jonathan Prunty, Luning Sun, Jose Hernandez-Orallo
Autoregressive transformers trained on limited trajectories of nonlinear dynamical systems can extrapolate to unseen parameter regimes, reproducing period-doubling cascades, chaotic dynamics, and attractor structures with high fidelity. In the logistic map, the model captures successive period doublings up to period 128, achieving a scaling ratio within $5 imes10^{-4}$ of the Feigenbaum constant. The study also shows how control‑parameter information is processed via attention, shaping the closed‑loop dynamics during training.
By Yilun Liu, Yi Zhang, Ganyu Wu, Sikuan Yan, Mengyue Wang, Alois Knoll, Volker Tresp, Yunpu Ma
The paper introduces RATIO, a framework for improving quantized reasoning models by identifying overthinking tokens and applying token-specific penalties. It uses Quantization-aware Reasoning Behavior Analysis to detect problematic tokens and Token-Specific Penalty Determination to assign penalties without extra training. Experiments show RATIO outperforms existing methods, boosting accuracy by up to 9.8 points and shortening chain-of-thought length by up to 51.3%.
By Chengzhu Bao, Xianglong Yan, Tianao Zhang, Jiaqi Chen, Shaoqiu Zhang, Yulun Zhang
The paper reviews one‑pixel attacks (OPAs), highlighting their extreme adversarial fragility across domains such as medical diagnosis, autonomous driving, biometrics, and quantum communication. It presents a PRISMA‑guided synthesis of studies from 2017 to 2026, offering a unified taxonomy that covers algorithmic foundations, black‑box evolutionary optimization, hybrid and program‑synthesis attacks, defence mechanisms, interpretability tools, and domain‑specific vulnerabilities. The review identifies dominant Differential Evolution strategies, emerging efficiency‑optimized and saliency‑guided methods, and gaps in dataset diversity, transferability, and evaluation standards, while proposing future research directions and a regulatory framework for robustness testing and AI security governance.
By Mirza Niaz Morshed, Md. Masudul Islam, Galib Muhammad Shahriar Himel, Md. Aslam Uddin, Hui Liu, Md. Shafiqul Islam
4DMulti is a physics‑guided learning framework that automates multicomponent identification from large‑scale four‑dimensional scanning transmission electron microscopy data. It leverages a 6‑million‑pattern diffraction database, a retrieval‑conditioned latent diffusion transformer (Sim2real) for realistic pattern generation, and a rotation‑invariant convolutional network for phase classification, achieving 98.82% accuracy on a five‑phase benchmark. The method introduces a diffraction‑inferred structural complexity metric and produces single‑nanometer‑resolution structural maps of complex material interfaces such as superconducting heterostructures, corroded alloys, and degraded solid‑state battery interfaces.
By Haoran Zhang, Zian Mao, Shufen Chu, Xiaoya He, Yuyan Guan, Antong Yang, Mingze Li, Xiaoqin Zeng, Yujun Xie
DynGhost is a transformer-based architecture designed for dynamic ghost imaging, addressing two key limitations of prior methods: the lack of temporal coherence exploitation and the assumption of Gaussian noise. It alternates spatial and temporal attention blocks and incorporates a quantum-aware training framework that uses realistic detector simulations (SNSPDs, SPADs, SiPMs) and Anscombe variance-stabilizing normalization to handle Poissonian noise. Experiments on multiple benchmarks show that DynGhost surpasses traditional reconstruction methods and existing deep learning models, especially in dynamic and photon-starved scenarios.
By Vittorio Palladino, Ahmet Enis Cetin
The paper investigates PRM‑Pruned Fragment Grafting (PPFG), an inference‑time technique that extracts high‑reward prefixes from a pruned chain‑of‑thought and grafts them into a sibling decoding process. Experiments on Qwen2.5‑7B‑Instruct with Math‑Shepherd across 500 MATH problems and multiple seeds show that PPFG performs statistically indistinguishable from a parallel‑CoT baseline, with only 14% of grafts targeting genuinely struggling chains. The study extends across three language models, six benchmarks, and multiple PRM configurations, concluding that PPFG’s inertness is not due to heuristic specifics and providing an equivalence‑testing framework for mechanism nulls.
By Khawaja Murad ul Hassan, Mehran Ebrahimi
The paper introduces FedCausalCompose, a causal world‑model framework designed for modular large‑language‑model agents that interact with distinct services such as order, payment, inventory, and shipment. It demonstrates that standard observational world models suffer from irreducible interventional errors when unblocked back‑door paths exist, whereas incorporating intervention‑response evidence improves interface recovery and can outperform non‑causal baselines when coverage and local mechanism errors are controlled. Experiments show that causal interfaces are most beneficial in structured tool environments with clear API signatures, while they provide little advantage in dialogue or narrative settings unless the causal information becomes directly relevant to the agent’s decision making.
By Xinyuan Song, Zekun Cai
The study evaluates whether detailed, profession‑specific system prompts improve performance on scientific tasks. Using an open‑source corpus of 503 agent profiles and Gemini 3.8 Flash, the authors compared matched profiles to four control prompts across nine text‑based science benchmarks and a tool‑using bioinformatics benchmark. Results show no consistent accuracy gains; matched profiles actually increased token usage and cost, and in some cases reduced success rates, with only a minor advantage in one benchmark likely due to prompt length rather than domain expertise.
By Timothy Kassis
The paper introduces Rules to Tools (R2T), a system that provides executable checks for scientific coding agents to verify compliance with public scientific requirements. In experiments across multiple task cohorts, agents using R2T’s prepared checks achieved high repair success rates—26/30 with text and 29/30 with checks—while also demonstrating varying task preferences and cost trade‑offs. The study quantifies how tool‑enabled checks influence repair outcomes and agent‑side resource usage in scientific computing contexts.
By Jingjie Ning, Guojiang Zhao, Chen Xu, Shanshan Zhong, Xiaochuan Li, Ji Zeng, Guolin Ke