The paper introduces PRISM, a method for recognizing procedural mistakes in videos by grounding video segments in language and mapping them to procedural states using a lightweight hidden Markov model trained on correct executions. It aligns reference annotations with four edit-based error types—insertions, deletions, substitutions, and transpositions—to evaluate across heterogeneous benchmarks. Experiments on three datasets show that using language grounding, semantic role labeling, and a trained local decoder consistently improves mistake detection performance compared to nearest-candidate matching.
By Olga Loginova, Oleksandr Bezrukov, Frank Keller
VEILED‑SH is a new benchmark for evaluating AI responses to indirect expressions of suicide, self‑harm, and distress. It contains 10,003 prompts drawn from counselor chats, forum posts, platform searches, and triage, all grounded in Singaporean contexts and public datasets. The dataset uses an expert‑validated multi‑label taxonomy to classify intent, communicative function, and response behavior, and it includes six automated profiles calibrated to peer‑support practitioners and clinicians for proxy‑alignment scoring.
By Leanne Tan, Keith Chua, Shaun Khoo, Onno P. Kampman, Lim Ruling Charmaine, Hazirah Hoosainsah, Kellie Yu Hui Sim, Kai He, Kenny Tsu Wei Choo, Roy Ka-Wei Lee
The paper proposes a multi‑branch conditional GAN that fuses visible, near‑infrared, and structural features for multispectral document binarization. By applying spatial‑ and channel‑wise softmax gating at encoder depth 3 and guiding the fusion with edge maps from selected VIS‑NIR band pairs, the authors evaluate nine configurations and find that the F3 + NIR edge‑guided model achieves the best performance, improving F1, p‑FM, and IoU over a fixed‑average baseline with statistically significant gains. The study demonstrates that targeted cross‑spectral structural guidance can effectively regulate intermediate fusion and enhance foreground‑overlap accuracy while preserving all four spectral bands.
By Yasas Savinda, Maheshi Buddhinee Dissanayake
The paper introduces ReMAS, a reinforcement learning-based framework for black-box red-teaming of multi-agent systems (MAS) that fine-tunes attacker LLMs to craft adversarial prompts. ReMAS operates in two stages: a rewrite stage that refines base attack prompts to boost extraction success, and a template generation stage that builds attack templates to better invoke specific sub-agents and expose their system prompts. Experiments demonstrate that ReMAS significantly outperforms existing black-box attacks, showing high success rates and transferability across different backend LLMs and MAS structures.
By Yuqi Jia, Hangsheng Zhang, Chenxi Liu, Junjie Liang, Xun Chen
The paper introduces OP-Print, a unified framework for membership inference attacks on on‑policy self‑distillation models. It shows that the supervision signal in such distillation is sparse and structured, focusing on uncertainty‑sensitive correction sites, and exploits this property to construct attack signals for both gray‑box and black‑box threat models. Experiments demonstrate that OP‑Print outperforms baseline attacks, which remain near random guessing, across various models and distillation schemes.
By Qi Li, Xinchao Wang
The paper investigates how depth growth and looping in language models influence in-context learning (ICL). It shows that models with these architectural features rely more on contextual information, acquire ICL earlier, and retain it better than fully dense counterparts. Attention analysis reveals that depth growth propagates early induction heads across layers, and ablations confirm that these models depend more on induction-ranked heads for few‑shot performance and token prediction.
By Ferdinand Kapl, Kaitlin Maile, Stefan Bauer, Johannes von Oswald
The paper introduces the Fragility Index (FI), a new performance metric that measures the tail risk of confident misclassifications in classification models. FI is framed within a robust satisficing framework to assess classifier sensitivity under distributional uncertainty, and the authors develop a surrogate-based training method with provable bounds linking training fragility to ranking-error FI. The approach is extended to common loss functions and includes an FI-based regularizer for deep neural networks, with empirical studies on medical diagnosis tasks showing that FI reveals risk not captured by accuracy or AUC and leads to models with lower operational costs in a human–AI diagnostic workflow.
By Zheng Cui, Daniel Zhuoyu Long, Jin Qi, Chen Yang, Ruohan Zhan
The paper investigates whether increased computational resources for an attacker can make adversarial attacks transferable across all defenders. It finds that for classification tasks, a quadratic resource gap guarantees either a transferable attack or a defense, whereas for generative tasks this guarantee fails. The authors also show that a defense that adds computation after seeing the input can be provably stronger than any fixed-model defense, suggesting that inference-time compute can enhance robustness.
By Greg Gluch, Shafi Goldwasser, Anastasiia Struss
GIFT (Gradient‑Informed Filtering for Distillation‑Augmented Reasoning RL) is a new method that resolves conflicts between reinforcement learning (RL) and distillation signals when training large language models for complex reasoning. It decomposes the distillation gradient into components parallel and orthogonal to the RL gradient, using the parallel part as a safety filter for conflict and the orthogonal part as a value measure for novel guidance. A progressive coverage‑annealing schedule then selectively applies token‑level guidance, gradually shifting from dense, filtered distillation to pure outcome‑grounded RL, yielding consistent performance gains—up to 9.5 % on challenging mathematical reasoning tasks—across multiple model scales.
By Shijie Li, Yilin Gao, Tieyuan Chen, Yun Luo, Weiyao Lin, Hang Yu
arXiv:2610. 30801v1 Announce Type: new Abstract: We study episodic linear Markov decision processes with unknown transitions, adversarial losses, and bandit feedback.
By Haolin Liu, Chen-Yu Wei, Hanbin Zhou
The paper investigates how learned multimodal representations of high‑dimensional health data can fail to preserve causal information, even when trained to reconstruct the data. It establishes conditions under which a representation is sufficient for causal identification, shows that interventional mediation requires stricter preservation than counterfactual comparison, and proves that reconstruction‑based objectives cannot certify sufficiency. The authors decompose the error of cross‑fitted doubly robust estimators into a first‑order representation bias and a second‑order nuisance‑estimation remainder, demonstrating that coverage can deteriorate with larger samples if the representation is insufficient. They propose a practical workflow that combines modality‑specific variational autoencoders, cross‑fitted doubly robust estimation, locked phenotypes, selective pathway analysis, portable representations, and diagnostics for representation insufficiency, exposure leakage, and sensitivity, and validate it in simulations where a rare confounder’s omission leads to severely under‑covered indirect‑effect intervals.
By Cong Cao
The paper proposes a method to detect large AI model training runs on a data‑centre power grid using Viterbi tracking of iteration cadence. On simulated data, the tracker identifies at least 96 % of training runs with a 1 % false‑alarm rate when the operator does not hide activity, and remains effective even against several adversarial hiding strategies. The approach requires only a modest passive power meter sampling at 2 Hz or faster, making hidden training both detectable and costly for the operator.
By Tom Kimpson, Mauricio Baker, Emlyn Graham
SIGNN++ is a learning-based star identification framework designed for celestial navigation in GNSS‑denied nighttime environments. It builds on the earlier SIGNN graph neural network by adding a realistic noise model, Fourier Transform of angular‑distance edge features, and a RANSAC‑based perspective star alignment step. Experiments show that SIGNN++ achieves higher star identification accuracy and solved rates than SIGNN and other publicly available baselines while maintaining comparable attitude accuracy and practical computation time for continuous operation.
By Jeongbin Sohn, Jina Lee, Max Yen, Dowan Gwon, Jeahn Han, Floyd Hepburn-Dickins, Brian Coltin, Ryan Soussan, Pyojin Kim
The paper evaluates five inference-time context augmentation methods across seven large language models and three datasets. It finds that accuracy improvements are often not genuine, as the methods can both correct and break predictions, especially when the added text only increases computational time. The study shows that methods adding no new content mainly affect predictions near the decision boundary.
By Haeun Yu, Geert Heyman, Pepa Atanasova, Isabelle Augenstein
The paper introduces a lightweight screening tool that evaluates individual patient vulnerability to membership inference attacks across thousands of medical AI models. By measuring attack scores for each patient, the authors found that certain records consistently rank high in risk regardless of training settings, and that stronger weight decay can reduce overall exposure but not the identity of the most at‑risk patients. The tool uses four small models to flag high‑risk patients before a production model is trained, aiming to integrate privacy assessment into clinical AI data preparation.
By Konstantinos Vrettos, Eleftherios Tzanis, Haridimos Kondylakis, Evangelia E. Vassalou, Florence X. Doo, Michail E. Klontzas
Reverse target fishing seeks protein targets for a query molecule, a key step in drug discovery. The authors reframe this as a generative retrieval problem, introducing GenFish, which uses residual vector quantization to map protein language model features to shared semantic identifiers and autoregressively predicts protein codes from a molecular fingerprint. They also present ProTargetBench, a proteome‑scale benchmark covering 20,431 human proteins that distinguishes experimentally inactive from unmeasured associations, and show that GenFish outperforms both discriminative neural and non‑neural baselines on full‑catalogue retrieval metrics.
By Zhichao Deng, Haichuan Tan, Xin Hong, Huaizhi Wang, Tao Huang, Bowen Gao, Yanyan Lan
The paper introduces LLM‑Auction, a learning‑based generative auction mechanism that integrates ad allocation with large language model (LLM) generation. It redefines the auction object from discrete ad slots to distributions over LLM outputs, aligning allocation with advertiser value and user experience while inherently modeling allocation externalities without extra inference cost. The authors prove allocation monotonicity and continuity, show a first‑price payment rule has favorable incentive properties, and demonstrate state‑of‑the‑art allocation efficiency in a simulated LLM‑as‑a‑judge environment.
By Chujie Zhao, Qun Hu, Shiping Song, Dagui Chen, Han Zhu, Jian Xu, Bo Zheng
The paper presents an empirical taxonomy of eleven hallucination sub‑types in large language models (LLMs) that generate UML class diagrams from natural language requirements, organized into Prompt Faithfulness, Code Faithfulness, and Reasoning Quality. The taxonomy was derived from 60 manually annotated DeepSeek‑R1 traces and applied to additional traces, revealing model‑dependent failure patterns and a negative association between requirement complexity and Prompt Faithfulness errors. A mitigation experiment with Qwen3‑based models showed that adding Prompt Faithfulness and Reasoning Quality constraints to the prompt improved the F1 score by 4.0 percentage points, while post‑hoc Code Faithfulness review offered no consistent benefit.
By Mojdeh Rahmanian, Foutse Khomh, Ashkan Sami, Yanchao Yu
The paper introduces DPIF, a plug‑and‑play dual‑path interaction framework that can be integrated with any encoder‑decoder segmentation network. DPIF aligns cross‑view distributions using Sinkhorn optimal transport, decouples features via conditional mutual information to transfer knowledge within shared semantic spaces, and applies persistent homology‑based topological regularization to encode anatomical invariants. Experiments on a private echocardiography dataset and two public datasets show that DPIF‑enhanced models outperform single‑view baselines and achieve comparable performance to full‑data single‑view models while using only 40–60% of the data.
By Shuhuan Wang, Yuanyuan Cui, Pengju Wang, Huimin Guo, Ling Yue, Yin Gu, Yanchao Zhang, Wei Feng, Yunhe Liu, Wei Qian, Mingjun Xu, Shuang Liu, He Ma
The paper introduces a new class of algorithms that embed existing automatic differentiation (AD) particle filter methods, enabling a bias‑variance tradeoff and reducing mean squared error in likelihood estimation for partially observed nonlinear stochastic dynamical systems. Unlike prior approaches, these algorithms only require a differentiable simulator of the latent dynamics, not explicit transition probabilities. Numerical experiments demonstrate that a hybrid method—using AD to refine a coarse solution from iterated filtering—outperforms current state‑of‑the‑art techniques on a challenging scientific benchmark.
By Kevin Tan, Aaron J. Abkemeier, Kunyang He, Giles Hooker, Edward L. Ionides