The study examines whether Mixture-of-Experts (MoE) language models develop linguistically structured expert routing during bilingual language acquisition. Using a decoder-only English‑German MoE Transformer trained with sequential language exposure, the authors probe token‑level routing distributions and measure category‑dependent specialization via mutual information, routing entropy, and Jensen‑Shannon distance. Results show that a curriculum‑trained model peaks at a mutual information of 0.1148 at layer 5, while a no‑curriculum baseline trained on mixed data achieves a higher peak of 0.2599 at the same layer, indicating stronger aggregate specialization. Replication with a second training seed reveals that the no‑curriculum condition’s specialization focuses on a single language in a seed‑dependent way, whereas the curriculum consistently yields a stable, language‑balanced routing profile, suggesting that staged bilingual exposure reduces single‑language dominance.
By Amrit Gopinath, Raghul, Durairaj Thenmozhi
arXiv:2608. 28564v1 Announce Type: cross Abstract: We study kernel ridge regression under anisotropic Gaussian data, where the input covariance decays as a power law with exponent $\alpha\geq 0$ for polynomial inner-product kernels.
By Lorenzo Rizzi, Arie Wortsman Zurich, Bruno Loureiro
The paper introduces a framework for synthetic‑augmented inference that balances the number of synthetic observations with their assigned weight. It defines a size‑weight frontier, estimating for each weight the maximum synthetic sample size that still guarantees target task‑marginal coverage for all smaller sizes. The authors provide finite‑sample coverage guarantees for configurations on or below this frontier and demonstrate that, when applied to augment opinion survey data with large language model responses, the method achieves the desired coverage while significantly tightening confidence intervals.
By Chengpiao Huang, Kaizheng Wang
The paper introduces Marginal Coverage Credit for Parallel State Entropy maximization (MCC‑PGPSE), a method that assigns policy‑specific credit by combining leave‑one‑policy‑out coverage with state‑owner specialization. MCC‑PGPSE retains the original pooled objective of PGPSE while redistributing non‑negative intrinsic rewards to discourage redundant exploration and encourage complementary coverage. Experiments on controlled tasks, seven public discrete‑state benchmarks, and Room/Maze environments show that MCC‑PGPSE consistently improves normalized team state entropy and state support compared to the Entropy baseline, with most gains attributed to the leave‑one‑policy‑out coverage component.
By Junhao Cao, Hongyi Xia, Jianian Wu, Xiaopeng Yi, Lixia Huang, Ping Guo
This paper investigates whether prosodic features—pitch, energy, and timing—are preserved when speech is translated between languages. Using multilingual dubbing data for English‑German, English‑Spanish, and English‑French pairs, the authors conduct a fine‑grained cross‑lingual analysis to quantify similarities and differences in prosody. The study identifies inherent cross‑lingual correlations in prosodic structure and explores how linguistic and alignment factors influence these patterns.
By Haopeng Xie, Ismail Rasim Ulgen, Sofia Son, Berrak Sisman, Philipp Koehn
The paper introduces Speculative Probing, a method that repurposes the speculative‑decoding module of large language models for real‑time classification tasks. By appending a trained soft prompt to the target sequence, the approach leverages the already‑cached KV store during inference, adding negligible overhead while achieving higher accuracy than traditional hidden‑state probes. Experiments on four classification tasks across multiple models show that these lightweight probes outperform zero‑shot GPT‑5.4‑mini and rival or surpass specialized 8B safety classifiers without running a full LLM.
By Collin Zhang, Tingwei Zhang, Vitaly Shmatikov
The paper investigates cultural biases in large language models (LLMs) by introducing the Culture-Related Open Questions (CROQ) dataset, which contains 24‑language questions about generic culture. Experiments reveal that LLMs disproportionately favor Japan in their responses, especially when prompted in high‑resource languages, while low‑resource languages tend to highlight countries where the language is official. The study also finds that these biases emerge after supervised fine‑tuning rather than during pre‑training.
By Joseba Fernandez de Landa, Carla Perez-Almendros, Jose Camacho-Collados
The paper investigates whether large language models (LLMs) replicate socio‑cognitive effects of power asymmetry observed in human communication. By assigning high or low status personas to LLMs in simulated multi‑turn dialogues across diverse professions, the study measures language coordination, pronoun usage, persuasion success, and compliance with unsafe requests. Results indicate that LLMs exhibit key power‑related socio‑cognitive behaviors, though with nuances and variability, linking these simulated interactions to both desirable and unsafe outcomes.
By Anvesh Rao Vijjini, Sagar Manjunath, Snigdha Chaturvedi
The paper exposes a new vulnerability in Federated Retrieval-Augmented Generation (FedRAG) called Routing Hijacking, where a malicious client forges its semantic profile to attract target queries despite lacking relevant data. Experiments across three FedRAG routing architectures show that this attack consistently misroutes queries, causing downstream failures such as missing evidence, poisoning, incorrect answers, hallucinations, and sycophantic behavior. The authors propose a trust‑aware post‑routing framework that reweights clients based on evidence feedback, which effectively suppresses persistent hijacking and transfers to a learned neural router.
By Junjie Mu, Qiongxiu Li
The paper introduces HiFTS, a unified autoregressive framework that generates hierarchical chain-of-thought (CoT) feedback before predicting trait-level and holistic scores for multi-trait automated essay scoring. HiFTS distills rubric-grounded CoT feedback from a teacher large language model and trains student models to jointly produce feedback and scores, employing Group Relative Policy Optimization to balance score agreement, calibration, feedback quality, and structural validity. The authors also present CFMS-34, a new Chinese multi-trait AES dataset, and demonstrate that HiFTS achieves strong scoring performance while producing coherent, rubric-aligned feedback on CFMS-34 and ASAP++.
By Shihang Yang, Sanwoo Lee, Ningning Zhao, Yunfang Wu
The paper introduces Cognitive Chain-of-Thought (CoCoT), a structured reasoning framework for vision‑language models that divides multimodal social reasoning into three cognitively inspired stages: Perception, Situation, and Norm. CoCoT improves performance across diverse tasks—multimodal intent disambiguation, theory of mind, social commonsense reasoning, and safety instruction following—by 5.9% to 4.6% on average. Fine‑tuning on CoCoT‑structured traces further boosts accuracy by 5–6% without explicit prompting, indicating that models internalize the structured reasoning pattern and that the approach enhances interpretability and social alignment in multimodal systems.
By Eunkyu Park, Wesley Hanwen Deng, Gunhee Kim, Motahhare Eslami, Maarten Sap
The paper investigates how model initialization affects federated short‑term load forecasting (STLF) when client data are heterogeneous. It introduces two strategies: a global pretrained initialization using auxiliary public load data to reduce client drift, and a local sequential initialization (SLIAvg) that lets clients start from progressively adapted models each round. Experiments on real smart‑meter data show these methods improve convergence and lower forecasting errors while remaining compatible with existing federated learning frameworks.
By Jianing Chen, Vajiheh Farhadi, Yan Li, Thomas La Porta
The paper introduces Safety-aware Contrastive Decoding (SafeCoDe), a lightweight, model‑agnostic framework designed to improve context‑aware safety in Multimodal Large Language Models (MLLMs). SafeCoDe operates in two stages: a contrastive decoding step that highlights tokens sensitive to visual context by contrasting real and Gaussian‑noised images, and a global‑aware token modulation strategy that adjusts refusals based on scene‑level reasoning and predicted safety verdicts. Experiments across various MLLM architectures and safety benchmarks demonstrate that SafeCoDe consistently enhances context‑sensitive refusal behaviors while maintaining model helpfulness.
By Zheyuan Liu, Zhangchen Xu, Guangyao Dou, Xiangchi Yuan, Zhaoxuan Tan, Radha Poovendran, Meng Jiang
The paper introduces egRUE, an explainable uncertainty estimation method that merges uncertainty quantification with feature‑level explanations for medical AI predictions. egRUE incorporates prediction explanations into its uncertainty calculation and decomposes uncertainty into contributions from individual features. Experiments and a user study with medical experts show that egRUE improves reliability, interpretability, and calibrated trust compared to existing methods.
By Li Rong Wang, Jamie Duell, Xinran Xu, Thomas C. Henderson, Yu Yue Hew, Pik Wan Erica Chiang, Xiao Wei Alstar Ang, Bingwen Eugene Fan, Xiuyi Fan
The paper introduces Replicant, a deep reinforcement learning framework that learns to evade malware detectors under a strict label‑only black‑box threat model. Replicant generates reusable policies for modifying malware samples and deciding when to query the target, and it transfers across different samples, detectors, and feature spaces. In experiments on seven Android malware detectors and three feature spaces, Replicant achieves a mean attack success rate of 78.8%, outperforming state‑of‑the‑art methods by 20.9%–39.2% and providing a stronger signal for adversarial training to harden detectors.
By Shae McFadden, Ilias Tsingenopoulos, Mario D'Onghia, Alexander Herzog, Myles Foley, Chris Hicks, Lorenzo Cavallaro, Fabio Pierazzi
The paper presents a diagnostic framework for Alzheimer’s disease that uses the Large Brain Model (LaBraM), a foundation model pretrained on over 2,500 hours of EEG data, to generate high‑dimensional latent embeddings. These embeddings are fed into a non‑linear Random Forest classifier, achieving an ROC‑AUC of 89.36% ± 3.49%, PR AUC of 81.45% ± 4.43%, and Balanced Accuracy of 82.44% ± 4.34% in a subject‑independent 5‑fold cross‑validation setting, using only 8‑second EEG segments. Post‑hoc occlusion and neurophysiological alignment analyses confirm that the model captures clinically validated biomarkers such as occipital‑frontal Alpha and Theta rhythm degradation and correlates with cognitive performance and clinical severity.
By Maggie Lin, Chung-Lin Hou, Tzyy-Ping Jung
HARTS (Hybrid‑Attention RL over Tree Structures) is a new system that jointly plans microbatches, data‑parallel replica assignments, and microbatch‑slot schedules to efficiently train agentic reinforcement learning models with hybrid attention over arbitrary rollout trees. It uses prefix compression and a linear‑time algorithm for chunkwise linear attention to avoid recomputing shared prefixes, enabling activation recomputation and bounded state replay while preserving trajectory‑wise training benefits. In experiments on an Agentic RL workload derived from SWE‑bench tasks, HARTS delivers 4.81–4.87× speedups in forward, backward, and gradient computations across multiple parallel configurations, with numerical differences comparable to baseline self‑rerun variation and a similar reward trend over the first 120 training steps.
By Boyuan Meng (Ant Group, China), Peihua Bao (Ant Group, China), Hong Liu (Ant Group, China), Xiaowei Zhu (Ant Group, China), Chao Wang (Ant Group, China), Gen Li (Ant Group, China), Zhenxuan Pan (Ant Group, China)
The paper develops a mathematical theory of superposition in neural networks using frame theory and compressed sensing. It shows that a sparse binary vector of active features can be encoded by an overcomplete dictionary and recovered via a ReLU operation with a suitable bias. The authors prove recovery theorems for both random-support and worst-case support settings, providing high-probability guarantees for low-coherence dictionaries and a sharp criterion for sparsity levels, with explicit results for Gaussian random matrices and equiangular tight frames.
By Michael I. Ivanitskiy, John Jasper, Emily J. King, Dustin G. Mixon
The paper introduces a Mixture of Multicenter Experts (MoME) framework that leverages diverse clinical strategies to reduce bias in medical AI without sharing data across institutions. MoME integrates specialized expertise from multiple centers, improving generalizability and adaptability of a multimodal target volume delineation model for prostate cancer radiotherapy. The model, trained with few-shot imaging and clinical notes, outperformed baselines, especially in high inter‑center variability or limited data scenarios, and allows local customization without cross‑institutional data exchange.
By Yujin Oh, Sangjoon Park, Xiang Li, Pengfei Jin, Yi Wang, Jonathan Paly, Jason Efstathiou, Annie Chan, Jun Won Kim, Hwa Kyung Byun, Ik Jae Lee, Jaeho Cho, Chan Woo Wee, Peng Shu, Peilong Wang, Caiwen Jiang, Nathan Yu, Jason Holmes, Jong Chul Ye, Quanzheng Li, Wei Liu, Woong Sub Koom, Jin Sung Kim, Kyungsang Kim
The paper investigates how dense embedding models can be used for stance-aware argument retrieval, a task that requires both topic relevance and correct stance (support or attack) toward a claim. Experiments reveal that current models favor topical overlap and ignore stance, and that contrastive training to fix this bias leads to over-correction, where models focus too much on polarity keywords at the expense of topic relevance. To address this, the authors propose diagnostic word-ablation metrics and a data‑centric solution involving a balanced argument curriculum and LLM‑augmented stance‑inverted arguments, which helps powerful models learn deeper directional logic and improves stance‑aware retrieval performance.
By Angelo Sparacino, Francesca Toni, Adam Dejl