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

Study introduces AMIFound model for enhancing medical images

The study evaluates the AMIFound model on 3.76 million image pairs from seven imaging modalities. According to the authors, it outperformed 38 comparison methods and improved results on 44 disease detection tasks, as well as predictions of tumor stage and survival risk.

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

The MASS method improves a language model but may reinforce its own errors

Researchers at Sakana AI and UC Berkeley introduced the MASS method. According to the study, after two rounds of self-improvement, the model achieved a success rate of nearly 70% against the baseline model. However, its own evaluator sometimes accepts incorrect results, which are then used in further training.

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

Humanity’s Sixth Sense test highlights weaknesses in models’ intuitive visual reasoning

Elorian and Scale AI released Humanity’s Sixth Sense, a test with 522 questions about images and videos. According to the reported results, humans achieved a pass@1 score of 93.1, while the highest-scoring model evaluated, GPT 6 Astra, scored 53.6.

Part of an overview of multiple AI topics; only this event has been covered.

Import AI (Jack Clark, Anthropic co-founder) Original source ↗
Research only one source so far

Study describes adaptive control of athletic movements in humanoid robots

The paper describes separating motion planning from motion tracking and using learned speed adaptation to transfer from simulation to hardware. According to the excerpt, the approach was validated on two humanoid platforms, including a serve performed without a motion capture system.

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

Study introduces POrchestra for ongoing adjustments to AI agent orchestration

The study introduces POrchestra, in which subagents request changes to their environment as they work. Across three test sets, the authors report an average relative increase in success rate of 9.1 % compared with the strongest baseline method; the OPOD method improved the success rate of a small controller agent by 10.0 %.

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

Study introduces DCOCO for faster generation using speculative decoding

The authors introduce DCOCO, which accounts for the context between tokens when proposing them in parallel. Across four verifier models and ten test sets, they report up to a 6.6-fold speedup over standard generation and a 6–23 % advantage over the strongest method in the comparison when handling concurrent requests.

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

Study introduces CB-Orchestrator for adaptive selection of LLM agent workflows

The CB-Orchestrator framework separates workflow generation from workflow selection using Contextual Bandit. On the MATH dataset, the authors report better results and a reduction in total API token consumption during training, including workflow generation, of at least 29.73 % compared with the methods evaluated.

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

RTI uses evaluation criteria to improve the training of AI agents

The study introduces Rollout-Time Intervention (RTI), a method that corrects flawed approaches during training based on evaluation criteria and filters out solution trajectories that differ too much. According to the authors, it outperformed baseline reinforcement learning methods in five benchmarks.

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

Study explores cheaper multimodal reward modeling with Qwen3-VL

The authors train a scoring layer on top of a frozen Qwen3-VL model using modified captions. Annotated data are used only to select the layer. According to the study, the 32B variant outperformed GPT-4o on VL-RewardBench by 6.6 points at roughly one-thousandth of the computational cost of fine-tuning.

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

Study introduces AIQ quantization for LLM inference

The authors propose AIQ quantization based on algebraic integers. According to the study, it achieves accuracy comparable to the MX format at the same number of bits per parameter. It converts computations to integer arithmetic.

arXiv cs.LG (Machine Learning) Original source ↗