AI Research

Researchers introduced SEATauBench, the first agent-focused evaluation framework for Southeast Asian sovereign AI, adapting TauBench to five languages. Testing across three models revealed that English agent capabilities…

Researchers introduced 5ting, a system for the SemEval-2026 Task 8 (MTRAGEval) that combines BGE-M3 dense retrieval, FAISS indexing, dual-query merged retrieval, and LLM-based reranking with role-separated generation con…

Researchers developed a structure-preserving Marathi-to-English document translation framework using a multi-stage LLM pipeline, integrating layout-aware OCR and HTML reconstruction to maintain formatting and domain term…

Researchers at arXiv found that using large language models (LLMs) like GPT-5.2 to label training data for entity matching can reduce manual labeling effort by hundreds of hours, with student models performing within two…

Researchers prove that mesh intelligence requires a liquid substrate—a continuous-time, multi-timescale network—because agents must estimate a changing latent from irregular, unscheduled observations without retraining. …

Reinforcement learning researchers need to clearly distinguish between using simulators as a proxy for real-world deployment and solving simulators as an end in itself, according to a new paper. The authors argue that co…

Researchers propose S-GAI, a spectral geometry-aware initialization framework for sigmoidal MLPs that encodes dataset geometry into network weights. Using SVD to extract class-wise spectral features, S-GAI initializes hi…

Researchers used polynomial algebra to study critical points in deep monomial networks, finding that for large activation degrees, criticality occurs at subnetworks where neurons are inactive or redundant. This provides …

Researchers propose an agentic AI pipeline for appliance-level energy anomaly detection in office buildings, combining deep time-series forecasting, variational anomaly detection, and LLM-based reasoning to generate prio…

Researchers proposed a scalable operator-learning framework, KL-DNN, for large-scale PDE problems, achieving 1.1 psi RMSE for pressure and 0.0146 for CO2 saturation in geological carbon storage simulations. The model tra…

Researchers propose Hierarchical Temporal Fusion (HTF), an extension of the Temporal Fusion Transformer that integrates hierarchical embeddings and a coherence-aware loss function to enforce forecast consistency across h…

Researchers used Lean 4's kernel-level tracking to show that the axiom of choice has a measurable geometric correlate in proof space, with classical proofs exhibiting distinct geometric signatures that decline with dista…

Researchers developed an incremental approximate message passing (IAMP) algorithm for empirical risk minimization under a multi-index model, achieving near-optimal training and test error in high-dimensional asymptotics.…

Researchers introduced SCSuff, an information-theoretic metric to evaluate whether free-text explanations from large language models (LLMs) are sufficient to explain their outputs. Testing on LLMs showed explanations are…

A new study reveals that test-time scaling in language models, where models sample multiple answers to a question, hits a ceiling beyond which additional samples do not improve selection accuracy and can even degrade per…

Researchers introduced RSEA, a Recursive Self-Evolving Agent that improves LLM agents by rewriting natural-language artifacts without weight updates, using a held-out selection gate to prevent regression. Across four ben…

Researchers introduced a 'capability slice' unit and a closed-loop system that links evaluation failures to targeted data interventions in LLM pre-training. In two case studies, the system correctly identified a masked l…

Researchers introduced GPTNT, a benchmark for real-time collaboration between multimodal agents based on the video game Keep Talking and Nobody Explodes. In tests, no current AI system could defuse a single bomb under ti…

Researchers introduced IMCBench, a benchmark for multimodal LLMs in image-grounded medical conversations, evaluating eight models across four families. Claude Opus 4.6 achieved the highest overall score (3.61), but no mo…

Researchers introduced DynaSteer, a dynamic representation editing framework that steers large language model reasoning trajectories toward truth by leveraging geometric insights into truth encoding and entropy-based int…