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

System1-Arena study compares rapid decision-making by models with generative reasoning

The System1-Arena benchmark compares 30 models on 92 thousand questions and in 138 interactive environments. According to the authors, the largest System-1 models match the accuracy of fast closed LLMs on non-agentic tasks at a quarter of the latency; small models rarely succeed in interactive tasks.

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

CLUE improves recognition of social context in conversations

The study presents CLUE, a method for inferring social context from conversation. Compared with the baseline Qwen model, it reports relative improvements in average weighted F1 of 17.3 % in the original domain and 9.3 % outside it, with a reduction of more than 80% in the number of reasoning tokens.

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

Study introduces MeddeX for analyzing and comparing medical reports

The study describes MeddeX, which combines a LLaMA-3 (8B) model fine-tuned using QLoRA with RAG and Gemini Vision to analyze medical documents. The authors report improvements over LLaMA3-Med42-8B of 11.8 % in ROUGE-L and 71.4 % in BLEU; the available excerpt does not establish clinical effectiveness.

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

Study explores attention with linear time complexity and a fixed number of recurrent states

The authors approximate attention using Taylor series and store the history in a fixed number of recurrent states. They report a 20.8× speedup at a length of 8192 and a 76× speedup during decoding at a length of 65 536. However, increasing the order of the approximation increases the size of the states combinatorially.

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

Study introduces robotic skin with more than 10 000 sensing units

Researchers introduced robotic skin with more than 10 000 sensing units and hierarchical signal processing. According to the study, it transmits 100 frames per second and enables robots to distinguish surface details as small as 300 µm and respond to harmful stimuli.

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

IRIS model recognizes traces of cell signaling in gene activity

According to a study in the journal Nature Methods, researchers at MIT have developed the IRIS model, which infers the signals received by cells from gene activity. It also recognizes characteristic patterns in cell types it did not encounter during training.

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

GPT-6-Astra solved 49 of 50 problems in the TheoPhysEval test, according to a study

A team of 35 researchers introduced TheoPhysEval, a benchmark with 50 unpublished theoretical physics problems. According to the authors, the GPT-6-Astra model running in Codex solved 49 problems at a total cost of just under 116 dollars. The test assesses the application of known methods, not the resolution of fundamental open questions.

The Decoder (daily AI news) Original source ↗
Research only one source so far

Study examines how to teach voice agents to follow new regulated procedures

The authors fine-tuned a model with 9 billion parameters on 2 506 conversations. In a test of a Thai-language voice agent, the number of successful scenarios rose from 15/43 to 40/43. However, an audit of the shared specification uncovered 11 flaws, three of which their own metric failed to detect.

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

The Progress-Aware Distillation method aims to reduce repetition of unsuccessful approaches by small AI agents

The study introduces Progress-Aware Distillation, a method that penalizes agents for stagnation during training. The authors report higher success rates and greater efficiency across eight models with 0.6 to 8 billion parameters; the available abstract does not specify the size of the improvement.

arXiv cs.LG (Machine Learning) Original source ↗