A new study shows that while medical LLM hallucination can be reliably detected (AUROC 0.77-0.86) using a simple probe, the underlying neural signal is distributed and redundant. Crucially, detection does not translate to control: steering hallucination-associated neurons fails to correct outputs, revealing a fundamental gap between readability and controllability.

Researchers introduce a comprehensive hate speech dataset covering six topics in Turkish and Arabic, and develop state-of-the-art BERT-based models for hate category classification, intensity prediction, target identification, and span detection.

A study evaluates frontier LLMs on Arabic cultural and sociolinguistic knowledge using human expert grading. The cross-evaluation framework tests models on Egyptian and Iraqi Arabic, finding that GPT-5.4 is the most reliable judge and that implicit cultural reasoning remains a major challenge.

A research paper presented at ACL 2026 BigPicture Workshop introduces techniques to harness language model latent spaces for better control and trust, including steering vectors and model calibrators.

Large language models face a binary failure in creative writing, oscillating between safe surface editing and uncontrolled plot expansion. Researchers propose Loom, a framework based on the narratological distinction between story and discourse, using a three-layer pipeline with intent-centered semiotic chain-of-thought to achieve precise control over narrative intent and rendering density. Evaluation shows Loom resolves this tension, achieving highest quality scores with gains in factual integrity and descriptive intensity.

A new paper challenges Beckmann & Butlin's ontological framework for LLM individuation, presenting four empirical wedges from persona-topology experiments and proposing regime-indexed individuation.

Federated learning is bandwidth-bound on two orthogonal axes: model size, which limits how often parameter-averaging methods can afford to merge, and class count, which makes per-probe soft-label distillation prohibitive at large vocabularies. TallyTrain collapses the class-count axis to ⌈log2C⌉ bits per probe by transmitting only each peer's argmax class index. Under non-IID training, hard-label majority voting filters noise where soft-label averaging amplifies it. TallyTrain matches or beats soft-label distillation at up to three orders of magnitude less communication. A bandwidth-bridge variant combining hard-label consensus with sparse parameter merges Pareto-dominates FedAvg, FedProx, and FedDF.

Thermodynamic computing devices based on the Ising model show great promise for low-power AI inference and edge computing, but scalable training methods remain limited. This paper presents a scalable backpropagation-based algorithm for training deep convolutional networks on Ising machines, achieving 94.9% on CIFAR-10 and 76.0% on CIFAR-100. It also develops a theory for the tradeoff between inference cost and accuracy, and discusses implications for hardware development.

This paper proposes a reinforcement learning method with a novel verifiable, label-free reward to train calibrated probabilistic forecasters. Applied to NFL win probability prediction, a 7B model trained solely with this reward matches betting market calibration without human labels or supervised fine-tuning, outperforming zero-shot frontier models.

Parameter-efficient fine-tuning (PEFT) reparameterizes weight updates in a fixed basis: low-rank adapters operate in the spatial domain, while spectral methods operate in a fixed Fourier domain. This paper introduces Fractional-Fourier Mixture of Experts, where each expert has a learnable fractional-Fourier order that interpolates between spatial and Fourier domains. Routing tokens through experts of different orders allows low-rank updates to be placed in their most compact domain, and the experts are naturally decorrelated, reducing interference and improving multi-task composition. The method adds negligible cost and outperforms strong baselines on LLaMA-3.1-8B and Qwen2.5-7B across various benchmarks.

Evolutionary neural architecture design for multivariate time-series forecasting remains underexplored. This paper introduces EVOTS, an evolutionary search framework that uses a modular genome representation and repair mechanism to discover task-adaptive Transformer-like models, achieving competitive results on ETT benchmarks.

A new paper proves that three popular methods for training language models to reason—GRPO, Dr. GRPO, and DAPO—are all adjusting the same number: the standard deviation of answer correctness across multiple samples. The group-standard-deviation identity shows that disagreement directly determines the size of the training update, validated on the Big-Math dataset and in controlled experiments.

FoGS is a novel synthetic data method for survival analysis that selects samples from a pool of multiple generators instead of generating directly, addressing data scarcity and privacy restrictions in clinical settings. On 16 public datasets, it significantly improves model performance over single-generator approaches while maintaining privacy.

This study disentangles the effects of classifier tuning vs. joint optimization in semi-supervised learning (SSL) pipelines for security classification, finding that simply tuning the classifier with Bayesian Optimization recovers 86% of the performance gain, and a simplified recipe matches the full pipeline.

Generative models used as surrogates for physical simulation often fail to enforce physics constraints. Constrained sampling can enforce constraints at inference but is computationally expensive. This paper introduces SNAP-FM, which leverages sparse GPU nonlinear optimization to accelerate constraint projection. Using block-sparse Jacobian and KKT systems with ExaModels.jl and MadNLP.jl, the method achieves faster nonlinear constraint projection on PDE benchmarks while maintaining accuracy.

This paper introduces the Manifestation Unit Protocol, a structured representation scheme for mechanistic interpretability that organizes component-level analysis outputs into queryable, reusable fields, validated across multiple models.

Seed2.0 is a model series that takes a meaningful step toward solving complex, real-world tasks. It identifies user needs, builds a reliable evaluation system, and targets long-tail knowledge and complex instruction following. It also delivers world-leading reasoning, visual understanding, and search capabilities.

A study on Lewis signaling games with LLM agents shows that memory architecture, specifically a persistent private notebook, significantly improves coordination over stateless agents, and channel capacity alone cannot predict success.

This study introduces Epistemic AI Literacy (EAIL), reframing AI literacy as a process-oriented epistemic phenomenon. Using the AIR framework, it analyzes epistemic aims and processes in GenAI-supported co-programming. Analysis of a large human-AI dialogue dataset reveals prevalent lack of EAIL: 78.8% of interactions relied on non-mastery-oriented aims and less reliable strategies, while only 11.1% showed high epistemic engagement.

A new paper introduces a contextual-bandit team game with two-sided asymmetric information, studying runtime human oversight of AI agents. It provides exact one-shot characterizations of team-optimal and myopic oversight rules, revealing a gap of avoidable harm and the price of non-credible oversight communication.

RareDxR1 is an end-to-end reasoning-centric large language model designed for open-domain rare disease diagnosis directly from unstructured clinical notes. It bypasses traditional pipeline-based phenotype extraction or retrieval-augmented generation by synergizing knowledge internalization with autonomous evolutionary learning, and employs Reflection-Enhanced Reasoning Sampling and dual-level curriculum reinforcement learning to improve diagnostic accuracy. Experiments show state-of-the-art results across multiple benchmarks.

A new conflict-free path planning algorithm for en-route air traffic control is proposed, leveraging solution-space displays for interpretability and flexibility. It integrates three intent-based conflict detection methods and two search variants (SSPPV and SSPPE). Empirical results using MUAC data show SSPPV with zone-based detection achieves optimal performance with average computation time of 3.69 ms.

Proposes a framework that replaces free-form LLM-generated web scraper code with typed JSON configurations, combined with a six-type collector taxonomy, template and utility-function constraints, static Airflow DAG execution, rule-based quality checking, and structured feedback correction. Experiments show zero execution-stage LLM tokens and lowest average wall-clock time on verified tasks.

This paper introduces MMM, a data model designed for knowledge documentation and interoperability across disciplines. Combining normative constraints with free-text labels, it addresses limitations of document-centric systems and formal approaches. A reference implementation demonstrates its usability.

A new paper proposes Bounded Morality, a formal framework that extends bounded rationality to moral cognition, modeling ethical theories as locally efficient strategies within a tradeoff between moral breadth and depth.

A new paper on arXiv introduces Constructive Alignment, a paradigm that reframes AI alignment as controlling how AI shapes human preferences over time, rather than satisfying static preferences.

Just eight months after launching its commercial service, AI leaderboard provider Arena, which originated as a research project at UC Berkeley in 2023, has reached $100 million in annualized run-rate revenue.

As AI shifts from development to production inference, compute demand is growing and moving to continuously operating AI factories. NVIDIA introduces a new strategy to provide large-scale accelerated computing access to startups and enterprises through a revenue-sharing model, with initial deployments by Sharon AI and Firmus.

Meta has spent billions on AI and data centers. Now it plans to launch a cloud infrastructure business, selling AI compute and models, competing with AWS, Google Cloud, and Azure.

Anthropic reinstates Fable 5 with stricter safeguards after a jailbreak incident. The new classifier causes more false positives, especially in coding tasks, and the US government gains significant oversight with pre-release access and dedicated resources, signaling a shift toward state-controlled AI.