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

RACE teaches AI agents to decide when they need further reasoning

The study presents RACE, a method that teaches AI agents to use previous reasoning and skip redundant steps. According to the authors, it reduced reasoning costs across four test suites while maintaining or improving task success rates. The excerpt does not specify the size of the savings.

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

Study introduces STI-OPD for distilling multi-turn AI agents

STI-OPD probabilistically replaces the student model’s actions with the teacher model’s actions based on the difference between their policies. The authors report better results than the strongest previous OPD method across all evaluated tests and student model sizes.

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

Study: image models generate stereotypical features associated with sexual orientation and criminal history

According to the study, GPT Image 1 Mini and Gemini 2.5 Flash Image accepted requests for stereotypical facial edits based on sexual orientation in more than 70 % and 99 % of cases, respectively. Both models accepted edits based on alleged criminal history in more than 97 % of cases.

The Conversation — Artificial Intelligence Original source ↗
Research only one source so far

The Caddie method trains advisors for LLM agents based on task outcomes

A study introduces Caddie, a method that trains an advisor using reinforcement learning without modifying the base model. According to the authors, an advisor based on Qwen3-4B increased the success rate of Qwen3-4B on the MuSiQue test by more than 25 percentage points and also helped other models.

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

Study introduces the GraphOPD method for training LLM agents

The GraphOPD method combines a graph of dependencies between agent steps with differences between the teacher and student. According to the authors, across three model sizes on ALFWorld, WebShop and SearchQA, it outperforms the strongest comparison method by up to 5.8 percentage points.

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

Study introduces ToolRACER for training agents on challenging conversations

The study introduces ToolRACER for generating synthetic conversations and ToolRACERBench with 5.6 thousand runs across six domains. Approximately 66 % include failure-prone scenarios. The authors report improved agent accuracy after training; the excerpt does not specify the size of the improvement.

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

Study introduces LatentGRM for faster evaluation of AI responses

The authors introduce LatentGRM, which evaluates pairs of responses without generating written assessments. According to the study, the LatentGRM-8B variant reduced evaluation time across four test domains by a factor of 6.1–7.0 at vote@5 while maintaining comparable accuracy in assessing preferences.

arXiv cs.CL (Computation and Language / NLP) Original source ↗