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

Autonomous laboratories with reward functions and experimental workflows

A paper from Nature Machine Intelligence argues that autonomous laboratories cannot be treated as a purely textual problem. The authors propose a framework with three principles: maximizing information efficiency per unit cost of an experiment, compiling hypotheses into experimental footprints with actions and rules, and…

Nature Machine Intelligence Original source ↗
Research only one source so far

MIT scientists deployed large language models to analyze code sharing in medical studies

A research team from MIT and the Dana-Farber Cancer Institute used LLMs to analyze nearly 4000 medical papers and found that code sharing in clinical studies of predictive models remains rare and often lacks the necessary documentation. The results, published in Nature Medicine, inform the development of guidelines…

MIT News – Artificial intelligence Original source ↗
Research only one source so far

Artificial intelligence in research work: criticism of opacity and overcrowding in scientific publishing

OpenAI solved over 700 math problems using AI, but mathematicians criticize the solutions for lacking conceptual understanding. AI is also accelerating the overload of scientific publishing: submissions to ICLR grew from 490 (2017) to 60 000 (2024), which threatens the quality of peer review.

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

Shepherd: a research framework for controlling and recovering AI agents

The Shepherd research project from Northeastern and Stanford University introduces a framework for overseeing AI agents. Meta-agents watch over worker agents, protect actions in a sandbox, and allow them to be rolled back or branched similarly to version control. The framework is written in Python and will be presented at…

IEEE Spectrum — Artificial Intelligence Original source ↗
Research only one source so far

MIT research: Data centers harm air quality and the health of residents far from their location

MIT research found that emissions from power plants responding to data center demand in the US caused about 650 premature deaths in 2023. Air pollution spreads as far as the Midwest, even though the centers are concentrated in Virginia and Texas. Data centers accounted for 4% of US energy consumption and 4.4% of emissions.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

Voluntary scaling policies of AI companies and the risk of regulatory capture

Researchers from the AI Now Institute analyze how AI companies, including Anthropic, are trying to regulate their models through voluntary safety policies. The text highlights the risk of regulatory capture – where companies maintain control over the regulatory process instead of independent regulation – as well as the contradiction between public…

AI Now Institute Original source ↗
Research only one source so far

Hybrid inference with timeout cascades reduces the costs of coding agents

Research evaluated 36 models from 2024 and found that small models achieve comparable accuracy on simple and medium tasks. A hybrid strategy that first tries a small model and escalates to a larger one only when needed reduces GPU costs by up to 33% without losing accuracy on SWE-bench Verified.

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

Tracking Representations in LLM Reasoning Traces for Iterated Games

A research study uses mechanistic interpretability to examine how LLM agents make decisions in iterative games with incomplete information. It finds that models build and iteratively revise internal representations, but once they prematurely commit to an incorrect belief, they rarely recover.

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

Jailbreak detection in language models using contrastive prototypes

A research team has proposed the JDPI method for detecting jailbreak attempts in LLMs. The method uses binary safety labels to create harmful and harmless prototypes without the need to train an additional classifier. Compared to the Llama Guard 3-8B tool, it achieves higher accuracy (ROC-AUC higher by…

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

The KnowledgeClaw system and the OmniKE-Bench benchmark for omnimodal knowledge engineering

A research team has published KnowledgeClaw, an agentic system for omnimodal knowledge engineering, together with the OmniKE-Bench benchmark (1829 items, five modalities). The system achieves a score of 55.9, 14.6 points higher than the closest baseline. The software and data are publicly available.

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

Do LLM models understand financial markets? The LUCID benchmark for testing causal reasoning

The paper introduces the LUCID benchmark for testing whether LLM models understand financial markets. Researchers tested 6 models on inferring causal structures from event descriptions and historical data. They found that graphs generated by the models usually reduce reconstruction quality compared to historical data alone.…

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

WikiEn-Ar190 research: comparing GPT-5-mini and AI agents in English-Arabic translation

The new WikiEn-Ar190 dataset (190 verified Wikipedia articles) is used to compare GPT-5-mini on its own and in combination with an AI agent in translation into Arabic. The standalone model achieved a COMET score of 0.8649; the combination with self-assessment improved lexical accuracy by 28.76% (BLEU 22.83).

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

NaviGATE: a robust generative model for multimodal reasoning over knowledge graphs

The research paper presents the NaviGATE method for multimodal knowledge graph reasoning, which combines structural representations with textual and visual evidence. It uses noise-aware dual-view instruction tuning and structure-constrained decoding. In experiments, it outperforms both unimodal and multimodal…

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

Hybrid RAG system for detecting fake reviews with RoBERTa-LoRA and LLM

Research demonstrates a hybrid RAG architecture for detecting spam reviews, combining RoBERTa with LoRA, FAISS retrieval, and an LLM. It achieves an F1 score of 0.90 in a cross-domain test, outperforming the OpenAI Detector (0.82) and competing with specialized SOTA models (fakeRoBERTa 0.97).

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

Audit of a deployed CV sorter: document length skews metrics, uncertainty sampling increases workload

A practical study of a deployed CV Priority Sorter system. It was found that document length was a misleading shortcut in the metrics; with length controlled for, an improvement of 95% relevant candidates after the first half remains. Uncertainty sampling increases reading time. A local cross-encoder achieves comparable performance to…

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