Skip to content

Research separate from news

What could become important next

A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.

1381 published research events

Latest work

A significant claim from a single source is published only after further confirmation.

Research only one source so far

Engineering trustworthy agentic AI systems for critical applications

A scientific survey of the trustworthiness of agentic AI systems in engineering. It defines five dimensions of trustworthiness (safety, robustness, transparency, auditability, privacy) and their application in four domains: energy, autonomous vehicles, HPC and communication networks. The aim is to create…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

From agent failure paths to quantified residual risk: A compositional framework for resilient agentic AI

A research paper introduces CPSAINT, a framework for analyzing failures of autonomous agents across seven layers (physical state, sensors, data, compute, actuators, environment, time), and FRIESA-K, a function for quantifying residual risks. The methodology maps failures to measurable risk using an absorbing…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Validating adaptive federated aggregation on heterogeneous cardiovascular datasets under differential privacy

A research paper validates FedCVR (federated learning with server-side adaptive optimization and differential privacy) on five real cardiovascular datasets. The model achieved an F1-score of 79.2 % and an AUC of 0.96 under an operational privacy budget (epsilon ≈ 4.2), statistically outperforming standard…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Type-aware repair allocation for prompt optimization in image generation

A research paper introduces TARA, a method for repairing incorrect prompts in text-to-image generators. The method diagnoses and resolves errors in object counting, attribute confusion and illegible text. It improves accuracy by 5.6 points over VisualPrompter on the DSG benchmark and runs 20 % faster.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Air Quality Arena: A dataset and benchmark for air quality forecasting

A new research dataset (AQA-Data) and benchmark (AQA-Bench) for forecasting air pollution. It covers 6 pollutants over 3 years across 7 countries, 4 continents and 14 000+ stations. It tested 11 time-series foundation models, which outperformed classic baselines; the top model uses a cross-modal architecture with…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Estimating total variation distance in autoregressive models

A research paper studies methods for estimating the total variation distance between autoregressive distributions (e.g. between different LLM engines). It proposes three approaches based on the type of data access: sampling, logits and noisy logits. It empirically validates them on SGLang vs vLLM. Code is available online.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

PhoenixRepair: Rethinking the Repair Search Strategy in Software Agents

The research paper presents a multi-agent framework for automatic code repair. PhoenixRepair explores multiple potential edit locations and performs iterative reflection to generate better patches. On SWE-bench-Verified, it achieved a 7.8% improvement over SWE-agent with DeepSeek-V3.1 and 76.0% Pass@1 with…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Semantic primes as explanations for emotions in language models

A study examined whether semantic primes from Natural Semantic Metalanguage explain emotional representations in LLMs better than appraisal-based approaches. Across four models (Llama-1B, Gemma-2B, Gemma-9B, OLMo-7B), it found that NSM primes enable control of emotions that is three times as strong and twice as precise.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Reward-aware population scaling for evolution strategies in LLM fine-tuning

The preprint shows that when fine-tuning LLMs using evolution strategies, population size behaves more as a function of reward design than as a fixed constraint. With z-score normalization disabled, N=2 achieves better results on GSM8K and TREC, where the normalized variant fails. The article mathematically models…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

SynPre-FL: Federated Learning with Pretraining on Synthetic Data

A research paper introduces SynPre-FL, a framework combining synthetic healthcare data generation (a latent autoencoder with a diffusion model) with federated learning. The method addresses client heterogeneity and class imbalance, protects privacy against attacks (membership inference, reconstruction), and improves…

arXiv cs.LG (Machine Learning) Original source ↗