This article discusses how security teams can translate technical cyber risk into financial terms that boards understand, and introduces how Databricks Genie enables real-time, data-driven cyber risk quantification for better governance.

The FTC fined Cox Media Group, MindSift, and 1010 Digital Works nearly $1 million for falsely claiming their “Active Listening” AI marketing service could listen to consumers' conversations via smart devices. The service actually resold email lists, and the FTC noted that hidden opt-ins in terms of service do not constitute valid consent.

A controlled study finds that Claude Haiku compresses 10 English intensity words into 5 distinct median outputs, with the model’s interpretation heavily dependent on system state, and displays three behavioral modes near feasibility limits.

This paper introduces OGCaReBench, a retrieval-focused benchmark for evaluating LLMs on clinical questions beyond standard guidelines. Experiments show that even GPT-5.2 achieves only 56% accuracy, but retrieval augmentation boosts performance to 82%, highlighting the need for evidence-grounded reasoning in medicine.

Reflective Prompt Tuning (RPT) is a framework that uses LLM function calling to simulate the iterative workflow of human prompt engineers. RPT employs a diagnostic function to evaluate the target model over an optimization set, summarizes recurring failure modes, and uses accumulated memory to revise prompts iteratively. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points, remains competitive with state of the art, and improves confidence calibration.

This paper proposes PromptNCE, which estimates pointwise mutual information (PMI) zero-shot using only LLMs and prompts, avoiding the need for task-specific critics. It achieves Spearman correlation up to 0.82 on three datasets and demonstrates applicability in low-data educational settings.

RankJudge is a novel benchmark generator for evaluating LLMs as judges in multi-turn conversations. It creates paired conversations where one has a single flaw injected into one turn, enabling unambiguous labeling and precise isolation of failure categories. Implemented across machine learning, biomedicine, and finance, it evaluated 21 frontier LLM judges and ranked them via the Bradley-Terry model. The approach also uses difficulty ratings to dynamically curate evaluations, reducing label noise.

This paper proposes a model-agnostic probabilistic token attribution measure that uses Bayes' rule to invert next-token log-probabilities, capturing the model's internal representation of token sequence distributions and revealing LLM behavior through entropy and attribution scores.

A new method, Sem-Detect, distinguishes human-written peer reviews from AI-generated ones by combining textual features with claim-level semantic analysis. It leverages the observation that different AI models converge on similar points while human reviewers are more diverse. On over 20,000 reviews from ICLR and NeurIPS, it improves [email protected]% FPR by 25.5% in binary setting. In three-class scenario, fewer than 3.5% of LLM-refined human reviews are misclassified as AI-generated.

A new framework called CR4T proposes rewriting unsafe or refusal-oriented LLM outputs into age-appropriate guidance for adolescents, offering a more human-centered alternative to traditional safety guardrails.

DualOptim+ is a novel optimization framework that introduces base and delta states to bridge shared and decoupled optimizer states adaptively, improving machine unlearning in LLMs. Experiments show superior trade-offs across forgetting, safety alignment, and multi-task learning tasks.

This study introduces yvsoucom-iterkit, a deterministic log-driven automated machine learning framework that encodes pipeline optimization as traceable log entities, enabling analysis of component attribution, interactions, similarity, and cross-seed robustness. Experiments on the Pima Indians Diabetes and Stroke datasets with over 18,000 pipeline configurations reveal a structured and partially redundant search space, where performance is governed by a small subset of interacting components. Random Forest importance analysis identifies augmentation, model choice, and imbalance handling as key drivers on Pima, while imbalance handling dominates Stroke. Component similarity analysis shows strong redundancy. The framework achieves strong performance using ensemble models, with a lower Macro…

This paper studies harness engineering for LLM agents, an inference-time technique that uses task decomposition and guided execution. It reveals that more elaborate harnesses are not always better, identifying failure modes like over-decomposition, over-pruning, and hallucinated execution. Experiments show that partial harnesses, which specify only initial steps, can achieve higher pass rates than fully structured workflows.

The research paper introduces a coin-flip model to predict program performance from few examples, finding that symbolic programs have all-or-nothing priors while prompt programs have diffuse priors, explaining why passing a few tests certifies symbolic but not prompt programs. RAP method retrieves similar tasks to construct proxy priors.

HealthCraft is a novel RL environment for evaluating LLM safety in emergency medicine, built on FHIR R4 with 24 MCP tools and a dual-layer rubric. Tests reveal that frontier models nearly fail on multi-step workflows, and infrastructure bugs can alter model rankings. The environment and tasks are open-sourced.

A simulation study explores double descent in overparametrized linear regression with contaminated data, finding that least-squares interpolation outperforms robust estimators.

GOEN achieves 0.9483 OOD AUROC on CIFAR-10, outperforming deep ensembles and other baselines. CenterLoss is found to degrade OOD detection despite improving classification accuracy. Training takes under 20 minutes on a single GPU.

A new study proves that no feature ranking can be simultaneously faithful, stable, and complete when features are collinear. The impossibility is quantified for four model classes, and an ensemble averaging method called DASH is proposed as a solution. A survey of 77 public datasets shows 68% exhibit attribution instability. The framework includes practical diagnostics and implications for fairness auditing.

This study explores whether language models can predict the success of research ideas before experiments. Using a dataset of 11,488 idea pairs, fine-tuned 8B models achieve 77.1% accuracy, outperforming GPT-5. RLVR further improves interpretability, demonstrating the potential of small models for autonomous scientific discovery.

Researchers propose the Temporal Contrastive Transformer (TCT), a self-supervised representation learning framework for capturing temporal dynamics in financial transaction sequences. Experiments show that TCT embeddings alone achieve AUC 0.8644, but when combined with domain-engineered features, no improvement is observed (AUC 0.9205 vs 0.9245), indicating overlap with existing features. The work suggests that self-supervised learning can reduce reliance on feature engineering, though further research is needed.

Mahjax is a fully vectorized Riichi Mahjong environment implemented in JAX, enabling large-scale rollout parallelization on GPUs. It achieves throughputs of up to 2 million and 1 million steps per second on eight NVIDIA A100 GPUs under no-red and red rules, respectively. Designed for tabula rasa reinforcement learning, it also includes a visualization tool. Experiments show agents can effectively improve their rank against baseline policies.