This study investigates whether internal signals for figurative language generation in multilingual LLMs are language-specific or cross-lingually reusable. Using activation steering, the authors estimate a direction from figurative-literal activation differences in one language and apply it to generation. Across five figurative categories, six languages, and four LLMs, these directions steer reliably within the same language, especially for metaphor and simile. More importantly, they transfer across languages: a direction learned in one language increases the target behavior in another, with German being most receptive. Combinations of directions from multiple languages can match or exceed native directions, while removing shared components weakens steering. These findings provide direct…

A new study uses historical cosmology as a controlled experiment to investigate how domain adaptation reshapes explanatory behavior in language models. Phase 1 trains a small model from scratch on a pre-Copernican corpus; Phase 2 fine-tunes a larger model with QLoRA. Results show domain adaptation primarily shifts explanatory framing (premodern vs modern) rather than directly altering cosmological stance, suggesting stance changes are secondary to framework shifts.

Large Language Models (LLMs) struggle to adapt to specialized domains without high-quality domain-specific data. This paper introduces Autonomous Agentic Data Engineering, where LLMs act as autonomous data engineers to plan, generate, and iteratively optimize training data. GPT-5.2 achieved a 57.29% improvement in a student model through agent-driven data adaptation.

We present a protocol to evaluate ChatGPT's ability to generate disease-centric biomedical associations. It outlines how we generate the associations, validate the biological entities using biomedical ontologies, and verify associations using literature. The protocol includes a self-consistency strategy to assess generative reliability across ChatGPT models. To address ontology exact-match limitations, we provide a use case performing semantic verification through a workflow enabled by Retrieval-Augmented Generation (RAG) powered by open-source large language models (LLMs). This enables LLMs to establish truth over content generated by other LLMs and expose hallucination.

This paper formalizes calibration for probabilistic label ranking, developing a hierarchy of notions covering full, sub-, and top-k rankings. It proves full-rank calibration implies others but not conversely. Empirical results show popular label ranking models are poorly calibrated, and calibration correlates with RLHF reward model accuracy beyond top-1.

Existing benchmarks mostly evaluate isolated or short interactive tasks, failing to test agents' ability to track evolving analytical context over long horizons. Researchers introduce LongDS, a benchmark with 68 tasks from real Kaggle notebooks spanning 2,225 turns across six domains. Evaluation shows best model achieves only 48.45% average accuracy, performance drops 47 points from early to late turns, and long-horizon errors account for 52%-69% of failures. Additional steps don't necessarily improve performance; key bottleneck is maintaining correct analytical state.

NumLeak is a measurement framework to detect memorization of public numeric benchmarks in foundation models. Top LLMs recall exact values with high Pearson r (0.97-0.99) on financial and economic data, but performance collapses on recent holdouts. White-box logprob ranking detects memorization better than open-ended generation, and a simple system prompt defense blocks 99.8% of suffix attacks.

This paper introduces the Maritime Anomaly Detection Quality Index (MADQI), a novel composite metric for evaluating unsupervised anomaly detection models on AIS data without labeled data. It integrates four metrics: ARC, PPS, SDS, and ECE, normalized via multi-chunk evaluation and adaptive scaling. Experiments achieved an MADQI score of 80.37%, with ECE and ARC scores of 0.907 and 1.000, demonstrating effectiveness in detecting extreme anomalies and maintaining consistency.

A new fMRI generative framework called Dual-Spectral Flow Matching (DSFM) is proposed, cascading discrete wavelet transform and discrete cosine transform to capture multi-scale transients and low-frequency energy compaction of BOLD signals. Class-conditioned cosine-frequency representations are generated via spectral flow matching, then reconstructed into physiologically plausible time series via inverse transforms, improving downstream brain network classification. Accepted at ICLR 2026; code open-sourced.

This paper introduces a novel LLM architecture based on RBF networks that eliminates deep neural networks, finds the global optimum of the loss function in closed form in one iteration, and offers improved explainability and accuracy.

This study introduces a multi-model paradigm to study synthetic deception via LoRA fine-tuning of five transformer models. Linear probes detect deception with near-perfect AUC in early layers, and logistic regression probes outperform MLP probes, supporting the Linear Representation Hypothesis. Probes generalize across domains with minimal loss. Different models exhibit distinct representational regimes: collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. The results show that robust, domain-invariant deception representations can be rapidly entrenched through modest supervised fine-tuning, with implications for activation-based monitoring.

The Unicorn framework uses a latent prototype codebook to decouple correlation modeling from channel identities, enabling scalable pretraining across heterogeneous datasets and outperforming existing methods in few-shot transfer scenarios.

Researchers developed a deep learning framework to predict hip dynamics from gait kinematics, comparing LSTM, Transformer, and Mamba. Transformer performed best on 60 healthy subjects and retained moderate predictive ability on osteonecrosis patients in zero-shot testing. The study demonstrates feasibility but calls for broader validation.

Quantum computing in the NISQ era requires hardware-level capabilities such as mid-circuit measurement, classical feedback, precise timing, and pulse-level access, all exposed by OpenQASM-3. However, no dataset existed to train LLMs on these hardware-oriented features. QASM-Eval fills this gap with an expert-verified test set of 100 tasks and a training set of 4,000 tasks covering classical logic, timing scheduling, pulse control, and complex workflows. An extended verifier automatically checks syntax, quantum states, and timeline. Evaluation shows that state-of-the-art LLMs struggle significantly on OpenQASM-3 tasks, but fine-tuning on QASM-Eval yields substantial improvements, providing a crucial benchmark for developing reliable LLM assistants in quantum programming.

MAVEN (Modular Agentic Verification and Execution Network) is a lightweight symbolic reasoning scaffold designed to enhance generalization in tool-calling environments through structured decomposition, adaptive tool orchestration, and intermediate verification. On the MAVEN-Bench stress test, MAVEN improves the GPT-OSS-120b base model from 48% to 71% accuracy without additional training, using an open-weight backbone at roughly 1/10 the cost of proprietary baselines.

Existing healthcare AI benchmarks fix provider responses, failing to evaluate mechanisms by their resulting equilibrium. This research recasts hospital mechanism design as program synthesis for language models, using a multi-agent simulator (Medi-Sim) and LLM-guided evolutionary code search to synthesize an inspectable mixed-objective program that eliminates up-coding, halves rejection, and retains most baseline funds.

A new paper proposes three rerooter designs to enhance subgoal-based policy tree search by using implicit subtask decomposition via the sqrt(LTS) algorithm, achieving state-of-the-art online training efficiency without explicit subgoal generation.

EHRBench is an automated and reliable benchmark that leverages electronic health records (EHR) through an EHR-LLM-KB pipeline to generate nearly 1 million QA items covering diagnosis, treatment, and prognosis tasks. It evaluated over 30 LLMs, revealing consistent performance trends and highlighting gaps toward clinically reliable systems. The work has been accepted at KDD 2026.

LLM agents adapt by updating external harnesses (prompts, skills, memories, tools), but it's unclear if base task-solving ability predicts harness self-evolution capability. The study shows harness-updating is flat across models, while harness-benefit is non-monotonic: mid-tier models benefit most, with weak and strong tiers benefiting less. Recommendations include investing in the task-solving agent over the evolver and focusing on harness invocation and long-horizon instruction following.

A paper on arXiv explores using MAP-Elites, a quality diversity algorithm, to procedurally generate First-Person Shooter (FPS) maps. The authors introduce two novel map representations (Point-Line and Spatial-Layout) and define metrics for topological and emergent properties. Using MAP-Elites with Sliding Boundaries (MESB) to evolve maps, results show new representations yield higher diversity and quality than previous methods.

This paper investigates encoding factored tasks (FTS) in SAT, proposing several encoding strategies and analyzing the impact of task transformations and parallelism on SAT-based planner performance.

PhyDrawGen is a neuro-symbolic pipeline that generates physically accurate diagrams from text. It uses an LLM to extract a scene graph, a deterministic solver to encode physics constraints, and a fine-tuned Qwen-VL model to iteratively correct violations. Evaluated on 1,449 problems, it outperforms GPT-5-image and Gemini models.