AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection is a multimodal system that evaluates both email content and embedded URLs to detect spam and phishing. It uses a two-layer approach: first, a URL classifier estimates prediction uncertainty, and only messages with high uncertainty are passed to a fine-tuned transformer encoder for deeper semantic analysis. Evaluated on eight diverse training corpora and two real-world datasets covering a decade of attacks, AURA achieves a macro F1-score of 0.9858 in-distribution and maintains scores above 0.94 on the NazPhish-Eval and GuenterTrap-Eval datasets, demonstrating strong generalization to new attack scenarios.
By Omran Berjawi, Walid fahs, Rida Khatoun
The paper presents a modality‑agnostic, hierarchical Transformer framework for assessing cognitive workload using heterogeneous biosignals. In a pilot study, the authors evaluated all 31 combinations of five modalities (ECG, EDA, RESP, SpO₂, EEG) across three tasks (IQ, MATH, GAME) and found that EEG alone performed best, while adding more modalities did not consistently improve results. The full five‑modality model achieved the highest average accuracy (73.02% on IQ, 68.08% overall) and reduced model size by about 50% compared to late‑fusion approaches.
By Stefanos Gkikas, Christian Arzate Cruz, Calvin Joseph, Giorgos Giannakakis, Raul Fernandez Rojas
The study evaluates zero‑shot essay scoring using sub‑3B open language models that run locally on a single 8 GB consumer GPU. Four instruction‑tuned models (Qwen2.5‑0.5B, 1.5B, 3B and SmolLM2‑1.7B) were tested on all eight ASAP‑AES prompts, comparing rubric‑decomposed versus holistic prompting, different aggregation methods, and trait‑mapping strategies. Results show rubric‑decomposed prompting consistently outperforms holistic prompting, trait‑mapping is sensitive to calibration, and longer essays reduce error, yet the best local configuration (macro QWK 0.388) still falls short of human agreement and a length‑only baseline, suggesting these models are best suited for formative, human‑supervised feedback.
By Nguyen Dung Son, Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, Nguyen Thai Anh
The paper compares the parallelism capabilities of three diffusion large language model paradigms—masked, uniform, and Gaussian diffusion. It proves that uniform and Gaussian diffusion can sample with a number of forward passes scaling with the dual total correlation of the distribution, potentially much less than the context length, whereas masked diffusion may require more passes. The study establishes a provable separation in parallelism, showing that masked diffusion’s critical windows are asymptotically narrower than those of the other two approaches.
By Sitan Chen, Liye Wang
The paper introduces a method called score centering to address the training‑inference mismatch (TIM) that destabilizes reinforcement learning for large language models. By adding an additive correction term that cancels drift between training and inference engines, score centering stabilizes RL and can match or surpass importance‑sampling techniques, especially as model size and mismatch severity increase. The approach also composes with importance sampling, yielding further performance gains in staleness experiments.
By Martin Marek, Max Ryabinin
The paper introduces ALIBI, a semantic cover story attack that injects a small, non-executed read‑only section into compiled binaries to mislead large language model (LLM) malware analyzers. By embedding a coherent but false security narrative, ALIBI can cause LLMs such as Gemini 2.5 Pro, GPT‑5.5 Pro, and Claude Opus 4.7 to downgrade or flip the verdicts of malicious samples. The attack also transfers to ELF binaries, and even a verification‑guided defense prompt only partially mitigates the effect, leaving a significant portion of malicious samples classified as benign.
By Hyeongjun Choi, Wonyoung Jung, Haehoon Seo, Sungyup Nam
Relational BabyLM is a decoder‑only Transformer that replaces standard self‑attention with a Dual Attention Transformer (DAT) to separate object‑level lexical features from structural/relational information. The model incorporates a Next‑Latent Prediction objective to compress history into a dense belief state and introduces a RoPE‑based symbol‑retrieval mechanism. On the BabyLM 2026 challenge, the best model ranks 6th overall and 3rd on the NLP‑task subset, outperforming GPT‑2 on most benchmarks and achieving the highest EWoK score among strict‑track entries.
By Adrian Brasoveanu, Ece Takmaz, Jakub Dotla\v{c}il
The paper introduces dQwen3.5, a family of diffusion language models derived from the hybrid-attention architecture of Qwen3.5 at 0.8B, 2B, 4B, and 9B parameters. It demonstrates that adapting a hybrid backbone—combining attention and RNN layers—can be more efficient than full-attention models, reaching a target training loss in roughly half the tokens. Across scales, dQwen3.5 exhibits full-attention-like behavior in any-order decoding and strong performance with parallel decoding.
By Anton Xue, Litu Rout, Aditya Akella, Adam Klivans, Sujay Sanghavi, Sanjay Shakkottai
The paper introduces QUATRO, a reinforcement‑learning approach for fine‑tuning large language models that enforces trust‑region constraints directly rather than relying on heuristic clipping. By deriving a principled objective, QUATRO provides explicit control over policy updates and stabilizes entropy during training. Experiments on mathematical reasoning benchmarks demonstrate that QUATRO maintains stable training even with higher learning rates and increased policy staleness.
By Doyeon Lee, Eunyi Lyou, Hyunsoo Cho, Sookyung Kim, Joonseok Lee, Jaemoo Choi
The paper introduces Post-Boundary Bridge (PBB), a hybrid transformer architecture that keeps causal attention within blocks while adding direct connections across block boundaries. PBB focuses on within-block modeling and nearby exchange, delegating long-range communication to full-attention layers. Experiments on dense and mixture-of-experts models ranging from 205 million to 2.07 billion parameters show that PBB hybrids maintain near-full perplexity, competitive downstream performance, and improved source retrieval, while Flash-PBB achieves 1.82× faster decoding with half the local key‑value cache compared to Flash‑SWA.
By Zhibo Yang
The study analyzes 17,012 app‑store reviews for six major generative‑AI apps, using BERTopic and RoBERTa to uncover topics and sentiment. Negative sentiment is most common around advertising, authentication, server reliability, and subscription pricing, with significant differences across apps—Claude shows the highest negative sentiment yet a highly enthusiastic user base. The authors also note geopolitical and privacy concerns for DeepSeek and propose a Trust Friction Score to quantify trust and usability barriers.
By Md Jafrin Hossain, Umme Nusrat Jahan, Shouvaggo Sharif Shammo
The study investigates how category‑specific harmfulness signals, isolated by removing the shared general harmfulness component, influence large language model (LLM) safety. Using activation steering across 11 risk categories in three instruction‑tuned LLMs, the authors find that the presence of harmfulness in these category residuals varies by category and that the pattern of inducing refusal is even more model‑dependent. Additionally, category residuals were shown to enhance the models’ downstream alignment with the shared general harmfulness representation, indicating that fine‑grained signals play a role beyond the general component.
By Soyeon Park (KAIST), Seogyeong Jeong (KAIST), Sunwoo Kim (KAIST), Alice Oh (KAIST)
The paper introduces SwitchSD, an adaptive framework that controls speculative decoding by distinguishing genuine copy intent from accidental repetitions using lightweight probes on a model’s internal representations. SwitchSD dynamically switches between neural drafting and context-based copying, achieving up to 15% throughput gains over existing baselines such as EAGLE3. The approach turns copying from a noisy heuristic into a principled, model-aware decoding regime.
By Roy Eisenstadt, Ido Cohen, Edo Cohen-Karlik, Lior Wolf, Itamar Zimerman
The paper introduces a Retrieval-Augmented Generation (RAG) framework for scientific image quality assessment, targeting both the understanding (SIQA-U) and scoring (SIQA-S) tracks of the SIQA challenge. It builds a multimodal index that merges textual semantics with fine‑grained visual features and employs a multi‑route retrieval and fusion mechanism to supply large language models with relevant reference cases, improving their evaluation of complex scientific images. The approach aligns well with human expert judgment and secured first place in the SIQA-U track at the ICME 2026 Grand Challenges.
By Yinuo Zhang, Bingshuo Liu, Zhiying Tu, Dianhui Chu, Qingbin Liu, Xi Chen, Jiang Bian, Xiaoyan Yu, Dianbo Sui
The paper investigates how alignment training, specifically reinforcement learning from human feedback (RLHF), affects the internal partisan structure of a large language model. Using a mechanistic case study on Llama 3.1 8B, the authors find that alignment training does not erase the model’s partisan geometry but compresses its variance, producing consistently balanced, non‑partisan outputs. Sparse autoencoder analysis and feature‑level steering experiments reveal that policy‑encoding features become inactive in the aligned model, indicating a causal disconnect rather than structural removal of partisan knowledge.
By Wendy K. Tam
The paper introduces a training‑free method for uncovering prompt‑conditional stylistic axes in large language models (LLMs). By repeatedly sampling completions of a single prompt at high temperature and applying Principal Component Analysis (PCA) to the pooled hidden activations, the authors automatically label the resulting axes using the extreme (pole) generations. Validation against 245 human‑elicited stylistic annotations shows that, for the Qwen‑3.5‑4B‑Instruct model, the top two axes align with human dimensions with 72.8% precision and 43.6% macro‑recall, and 75.6% of validity ratings confirm the axes’ polar generations, while other models exhibit varying degrees of discoverability.
By Ajit Mallavarapu, Ziwei Gu
ViTOED is a new dataset for target‑oriented emotion detection in Vietnamese social media, containing 10,985 user comments and 21,244 manually annotated opinion quadruples (source, target, expression, polarity). The dataset uncovers Vietnamese‑specific linguistic phenomena such as implicit sources and targets and vocabulary ambiguities, and it serves as a benchmark for evaluating Vietnamese pre‑trained language models. A baseline using structured sentiment graphs shows that span detection and relation extraction remain challenging, indicating significant room for improvement in Vietnamese target‑oriented emotion detection tasks.
By Chanh Vo, Son T. Luu, Ngan Luu-Thuy Nguyen
The paper examines how deliberative (System 2) reasoning affects a Retrieval-Augmented Generation (RAG) model’s vulnerability to knowledge‑poisoning attacks. Using two metrics—Cordon Rate and Leakage Rate—it evaluates six model configurations on 200 SciFact questions. Results show that enabling reasoning lowers both Cordon and Leakage Rates for DeepSeek‑V4‑Flash, indicating reduced behavioral impact from poisoned evidence, though overall attack success increases.
By Mehrdad Ghassabi, Audrina Ebrahimi, Sadra Hakim, Hamidreza Baradaran Kashani
The paper introduces WZPlanner, a new dataset and model for safe autonomous driving in work zones. The dataset, WorkZonePlan, contains over 149,000 synthetic and 5,000 real-world samples with 3D annotations for lane and work zone boundaries, plus 228 evaluation routes in CARLA. The proposed BoundaryFormer (BF) and its enhanced BF++ variants jointly predict lane/work‑zone boundaries and driving trajectories, achieving higher Driving Scores while being significantly smaller than competing models.
By Nishad Sahu (Raj), Changzhong Qian (Raj), Guangzhou Cai (Raj), Shounak Sural (Raj), Ragunathan (Raj), Rajkumar
The paper investigates how attention sparsity behaves in autoregressive image generation, finding a distinct diagonal sparsity pattern due to spatial locality of visual tokens. It introduces a diagonal‑aware sparse attention mechanism that skips KV entries along the diagonal within a recent window, achieving up to 3.1× higher throughput and 1.19× lower latency with less than 2% quality loss compared to dense inference.
By Daeun Kim, Junwha Hong, Changhun Oh, Yoonsung Kim, Yoonhyeong Lee, Jongse Park