The paper introduces information set emulation, a method that attaches detailed causal certificates—such as source evidence, timing, and proposed causal roles—to AI‑derived features extracted from electronic health records (EHRs). These certificates provide auditable evidence for causal roles and guide whether a feature can be used for causal inference or should be routed to compatible reporting or separate analyses. The framework integrates with a joint EHR observation map and offers identification, estimation, and diagnostic tools under standard causal assumptions, illustrated through synthetic simulations and a finite‑world example.
By Takes Fujita (VRI), Nobutaka Hattori (Department of Neurology, Juntendo University School of Medicine)
The paper introduces GYROval, a benchmark designed to assess cultural value orientation in large language models along the Inglehart‑Welzel axes across multiple domains and roles. It uses binary contrastive scenarios where both options are valid, and scores models by the proportion of responses aligning with each pole. The benchmark was tested on twenty models, with eleven also evaluated in Russian and at a different sampling temperature, and its stability was analyzed using tie‑corrected Kendall’s W against a permutation null.
By Alexander Didenko, Anna Shabanova, Vladislav Zapylikhin, Alexander Antipov, Ruslana Raemgulova
The paper proposes a split‑control architecture for adaptive security at the network edge, where an untrusted planner emits typed security intents that are vetted by a deterministic governor before being enacted. The governor checks each intent against safety, resource, temporal‑stability, and proportionality invariants, issuing signed receipts for admitted actions that are compiled into eBPF map updates. Experiments on a Raspberry Pi 5 connected to a university 5G test network show the governor can admit, reject, and bound intents at microsecond cost without disrupting protected‑flow regularity.
By Ijaz Ahmad, Ijaz Ahmad, Flavio Esposito, Erkki Harjula
The paper introduces AWED, a word‑level emission‑delay metric, and demonstrates that post‑training a delayed‑streaming model with a joint reward (GRPO) improves both accuracy and latency. Using a single operating point (τ=6 frames), the method outperforms both its supervised baseline and the Voxtral Realtime backbone across all lookahead budgets, reducing WER by up to 30.8% at 80 ms delay and lowering median AWED from 1.17 s to 1.04 s.
By Xiuwen Zheng
The paper reports the first systematic audit of open‑weight large language models (LLMs) in hiring contexts, examining how job‑posting language influences recruiter and job‑seeker simulations across six models. It finds that agentic language lowers recruiter scores for female candidates while communal language mitigates this effect, and that coded‑exclusion language sharply reduces recruiter scores for non‑White candidates and discourages non‑White personas from applying. The study also identifies the explicit demographic label as the main causal factor and proposes a concrete pre‑deployment audit protocol aligned with EU and U.S. regulatory requirements.
By Kosuke Kitahara, Nobuhiro Yamaguchi
The paper addresses the challenge of selecting demonstrations for long-context language model queries, where transformer inference costs grow quadratically with sequence length. It proposes two algorithms that distill transformer behavior into state space models (SSMs) with linear inference time, partitioning transformer layers into groups and estimating separate SSMs for each. The distilled SSMs achieve less than 0.7% approximation error, and in downstream tasks they reduce FLOPs by 14.2× while improving accuracy by 6.48% compared to baseline methods.
By Ziniu Zhang, Zhenshuo Zhang, Ruoxuan Xiong, Gene Cooperman, Hongyang R. Zhang
The paper investigates how procedural traces—detailed step-by-step accounts of a language model’s reasoning—affect the decision-making of LLM overseers tasked with auditing another model’s outputs. Using signal detection theory, the authors evaluated five overseers on 19 compliance tasks, finding that while error detection remains high when disconfirming evidence is always visible, more elaborate traces shift the decision criterion toward rejection, leading to increased false alarms. The study also shows that providing option labels reduces the stated inability to link evidence to options, yet some overseers still exhibit residual rejection of correct work that grows with trace detail.
By Zihan Chen, Di Zhu, Lei Zheng, Weiling Li
The paper investigates how a minority of biased agents in a multi‑agent system of large language models (LLMs) can amplify bias through textual interactions. Even a small percentage of persistently extreme agents causes significant opinion shifts among the non‑biased agents, with the effect occurring faster in the Llama 3.2 model than in a classical Friedkin‑Johnsen model. Semantic analysis shows that rhetorical consistency rises with biased exposure and that non‑biased agents adopt the biased vocabulary even when their numerical opinions change only modestly.
By Omran Berjawi, Giuseppe Fenza, Rida Khatoun
GroupKV is a lightweight hierarchical KV cache management system designed for long‑context diffusion large language model (dLLM) inference. It partitions the context into contiguous groups and uses coarse‑to‑fine sparse selection, cross‑layer consistency for predictive prefetching, and a staleness correction mechanism to keep the cache coherent amid dynamic KV updates. The approach also incorporates streaming prefill to lower peak memory usage, achieving up to 48× longer serviceable context, 3.73× faster inference in offload‑based settings, and competitive task accuracy.
By Jinhao Wang, Zhexin Hu, Kangjie Zhou, Xin Zhou, Fangfang Liu
The paper investigates how to balance model size and fine‑tuning strategy for UAV audio classification. Using a dataset of 3,100 clips across 31 drone classes, it compares transformer and convolutional backbones under full fine‑tuning, classifier‑only fine‑tuning, and four parameter‑efficient fine‑tuning methods. Results show that selective batch‑norm tuning of EfficientNet‑B7 yields the best accuracy (97.65%) while updating less than 0.5% of parameters, and that lightweight CNNs generally outperform transformers in both accuracy and efficiency.
By Andrew P. Berg, Qian Zhang, Mia Y. Wang
The paper extends Mixed-Integer Differentiable Predictive Control (MI‑DPC) to handle multi‑modal discrete decisions and nonconvex polynomial dynamics in Underground Pumped Hydro Energy Storage Systems (UPHES). It introduces a neural policy that maps problem parameters to continuous setpoints and integer mode selections using a Gumbel‑Softmax layer, trained via self‑supervised differentiation of the finite‑horizon control objective. Key methodological advances include a parallel differentiable simulator, a Transformer encoder for long‑range temporal dependencies, and a Gumbel‑Softmax temperature annealing schedule, enabling the framework to achieve only 1.6% suboptimality compared to a mixed‑integer quadratic programming baseline while providing a five‑order‑of‑magnitude speedup in online scheduling.
By Honghui Zheng, J\'an Boldock\'y, Yury Dvorkin, J\'an Drgo\v{n}a
The paper introduces PrefDT, a preference-conditioned Decision Transformer designed for multi‑objective scheduling of UAV mobile edge computing fleets. PrefDT accepts a desired energy‑delay trade‑off as input, enabling a single offline‑trained model to generate any point on the Pareto front during runtime. The authors employ attention pooling with a per‑user bypass to maintain scheduler operation when user reports are lost, and a distillation pipeline to create a preference‑labeled flight corpus, achieving superior trade‑off curves and tight energy budget adherence in simulations.
By Qiao Liao, Zhiyong Feng, Bin Wu, Guodong Fan
The paper introduces a calibrated instrument for rigorously measuring how inference optimizations—such as quantization, early‑exit, and speculative decoding—affect the output quality of large language models. It uses a formally calibrated LLM judge that verifies no systematic bias between statistically equivalent outputs and includes a null condition to ensure measured differences are zero. Applying this method, the authors find that a 4‑bit model is indistinguishable from its 16‑bit counterpart, while 3‑bit quantization and early‑exit techniques incur measurable quality losses that vary by language and task, and that token‑certainty‑based acceptance rules cannot reliably identify impactful errors.
By Jerry Kaplan
The paper introduces FairInference, a system that guarantees token-level latency isolation for well-behaved clients in multi-tenant LLM serving. It provides a δ-token fairness guarantee, ensuring that a token generated in isolation within time d will be produced within d + δ in a shared environment. The approach enforces per-token deadlines, bounds GPU compute sharing delays, and accounts for shared KV cache overhead, leading to reduced latency spikes and higher overall throughput compared to existing LLM serving systems.
By Dev Bali, Soujanya Ponnapalli, Yichuan Wang, Natacha Crooks, Scott Shenker, Matei Zaharia
The paper introduces a compact, NISQ‑compatible quantum transformer for synthetic QNLP sequence modelling. It replaces classical attention and feed‑forward layers with variational quantum encoder blocks, connector circuits, decoder blocks, and a two‑qubit measurement readout, while preserving the autoregressive next‑token interface. The authors evaluate several variants on deterministic and lexicographic grammar‑generation tasks, finding that the quantum models can learn nontrivial grammar structure but are less accurate and stable than a compact classical transformer baseline.
By Julian Hager, Michael K\"olle, Gerhard Stenzel, Tobias Rohe, Jonas Stein, Claudia Linnhoff-Popien
The paper introduces a reinforcement‑learning based rewriting agent that rewrites training data to reduce the distribution mismatch between supervised fine‑tuning (SFT) and a language model’s generation distribution. By formulating data rewriting as a policy‑learning problem, the authors train a lightweight LoRA rewriting policy that optimizes alignment with question‑answering style, maintains semantic diversity, and enforces task consistency. Experiments on three instruction‑tuned backbones show that models fine‑tuned with the rewritten data achieve downstream performance comparable to standard SFT while mitigating degradation on non‑downstream benchmarks, and preliminary tests suggest the policy can transfer across domains such as logical reasoning and medical question answering.
By Jiacheng Wang, Zhijie Liu, Ping Jian, Zirong Chen, Ke Ren Liao, Zhen Yang, Zhongbin Guo
FRAUDSkill is a structured frozen‑weight adaptation framework for audio anti‑fraud detection that keeps the underlying audio‑language model unchanged while optimizing external skill programs, route‑specific policies, and decision rules. It combines structured output control with validation‑guided multi‑path inference to produce protocol‑compliant predictions. On the TeleAntiFraud benchmark, FRAUDSkill achieves a 73.50% Macro‑F1 score, outperforming the shared frozen‑model baseline by 31.96% and reducing invalid outputs to 1.94%.
By Chengxian Hu, Zhiming Ma, Mingjun Pan, Yifan Wang, Shun Zhang, Qifan Wang, Zhilei Zhao, Yijin Zhou, Yuxi Zhao, Huiyuan Liu, Peidong Wang, Peng Chen
The paper introduces a multitask large reasoning model for molecular science that incorporates chemical knowledge via a multispecialist architecture, chain-of-thought supervision, and molecule-informed reinforcement learning. It coordinates prediction and inference specialists across ten molecular tasks—including description, generation, nomenclature translation, property prediction, and reaction prediction—using task-conditioned routing. The model surpasses more than 20 general-purpose and molecular large language models, improving aggregate performance by 50.3% and outperforming leading multitask baselines on most tasks, while maintaining interpretable chemical inference and demonstrating a workflow for CNS candidate generation and retrosynthetic planning.
By Pengfei Liu, Shuang Ge, Xiaobo Wang, Xin Liu, Jun Tao, Yan Li, Chao Liu, Ling Chen, Zhixiang Ren
The paper investigates how transformer models learn latent structure by training a small decoder-only transformer on three variants of the Alchemy benchmark. It finds that the model acquires different components of latent structure in discrete stages, with a notable asymmetry: it robustly composes fundamental transitions but struggles to decompose complex examples into atomic transitions. Layer‑specific causal interventions reveal plasticity windows where freezing layers delays or prevents stage completion, offering a detailed view of capability evolution during training.
By Rohan Saha, Farzane Aminmansour, Alona Fyshe
TeleAntiFraud 2.0 is a monthly‑frozen, audio‑based benchmark for telecom fraud detection that incorporates newly observed scam patterns while preserving earlier test sets. It uses a Mixed‑Tree Anti‑Fraud Generation Pipeline to create profile‑grounded scenarios, expands them into mixed‑tree dialogues, and renders validated speech for 900 Chinese calls (600 fraud, 300 near‑domain non‑fraud) each month. Experiments show that classifiers perform well against unrelated negatives but drop significantly against near‑domain negatives, highlighting the need for near‑domain construction and collapse‑aware reporting in realistic evaluation settings.
By Huiyuan Liu, Zhiming Ma, Yanxing Liu, Shun Zhang, Qifan Wang, Di Liu, Yifan Wang, Yuyang Deng, Haoyang Meng, Yijin Zhou, Yuxi Zhao, Chengxian Hu, Peidong Wang, Peng Chen