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

Lumen-bench study: a verbal refusal may be accompanied by a harmful tool call

The Lumen-bench study evaluates 30 models across five languages in 194 400 trials. According to the authors, refusals were accompanied by harmful tool calls in 8.57 % of cases involving an actual call. Replacing ethical system instructions with neutral ones increased the violation rate from 2.76 % to 13.64 %.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study introduces the CoDA framework to mitigate prompt injection attacks

The study introduces the CoDA framework, which enforces safety rules during generation in diffusion language models. The authors report a substantial reduction in the success rate of both static and adaptive prompt injection attacks while preserving utility; the available abstract does not include specific figures.

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

Study compares 17 models performing customer support tasks

The authors evaluated 17 models on customer support workflows using MCP-compatible tools. According to the study, GPT-5.4 leads overall; QWEN3.5-9B leads among open-weight models with fewer than 20 billion parameters. Average inter-rater agreement reached a kappa of 0.524.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Fin_Agent study evaluates rules and language models for multilingual financial NLP

The Fin_Agent study combines rules and the Qwen2.5-32B model for Greek named entity recognition (NER), rule-based scoring for Japanese, and BGE-M3 with the Qwen3-32B model for question answering. The authors report an F1 score of 0.8884 on a non-public NER dataset and test accuracy of 98 % for Japanese; they highlight the dependence on data.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study: Proximal Policy Optimization agent learns the rules of the Crafter environment

The authors used interventions that changed a single property of the state to examine the value function of a Proximal Policy Optimization agent. According to the study, it captures the rules for collecting diamonds and the effects of water and food on health, even though training focused solely on reward.

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

Study finds only partial agreement between LLM-generated behavior and epidemiological data

A study of three LLMs using hepatitis C as a case study found the correct direction for 46–68 % of 28 behavioral relationships. When asked directly about the same relationships, the models achieved 82–93 % agreement. The authors recommend further validation before using generated outputs in epidemiological models.

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

Study uses Whisper-small to detect language switching in speech

The study authors detect language switching directly from speech using activations from a pretrained speech encoder. With the Whisper-small model, they achieved an average F1 of 0.798 across 14 language pairs in the CS-FLEURS dataset; comparison methods achieved 0.776 and 0.751.

Apple Machine Learning Research Original source ↗