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

AscendGYM benchmark for optimizing LLM kernels on Ascend NPUs

AscendGYM is a new benchmark for optimizing LLM kernels on Ascend NPUs, featuring 48 kernels across 3 difficulty levels. The test cases come from real production tasks. It includes the Speedup-Token Score (STS) metric, which measures performance within a set token budget. The code is publicly available.

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

Training language models through internal disagreements without human data

Research introduces the Self-Disagreement method: a language model generates the strongest argument against its own answer, decides whether the argument holds up, and learns from that decision without human annotations or model preferences. This increases dialectical competence — the ability to know when to hold a position and when to change it…

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

The Reflect-Reasoning method for aligning language models with adaptive principles

A research paper from arXiv presents the Reflect-Reasoning method, which teaches language models to reason about adherence to principles instead of simply memorizing them. The method outperforms system prompting, generalizes to new domains (medicine, finance) without additional training, and reduces principle violations without…

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

CamSpokenNLI dataset for NLI transfer between Cameroon Pidgin English and Camfranglais

New pilot dataset CamSpokenNLI with 228 spoken NLI examples (114 each in Cameroon Pidgin English and Camfranglais). XLM-RoBERTa achieves 44.7-46.2% accuracy on text variants; cross-variety transfer 40.1-45.6%. Automatic speech recognition (Whisper) has a WER of 66.6-45.5% and poorly detects the language.

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

Multi-layer AI pipeline for generating BPMN diagrams from natural language

A research pipeline generates BPMN diagrams from natural language using fine-tuned RoBERTa transformers and LLMs. It achieves 90.7% accuracy in extracting control-flow relations. The method uses few-shot learning for better generalization with limited data and an ontological representation of processes. Tested on a real…

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

SCALAR: Parameter-Efficient Fine-Tuning

The SCALAR method achieves LoRA-level performance with 4% of the training parameters, with 20% lower memory usage and 6% higher performance on the Qwen2.5-7B model.

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

VEPO: A Method for Reducing Hallucinations in Vision-Language Models

Scientists have published the VEPO method, which reduces hallucinations in vision-language models without external tools. It transfers visual evidence from English to low-resource languages (Chinese, French, Hindi) using preference optimization. Tests have demonstrated a consistent decrease in hallucinations, with…

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

Improved regret analysis in policy optimization for adversarial MDPs

The research paper presents the VOPO algorithm with a novel regret analysis for adversarial MDPs. It achieves Õ(√(H³SAK)) regret, an improvement by a factor of √(H³) or √(S/H) over previous methods. The key innovation is the Stability-Matching Bonus for variance-aware analysis.

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

Spectral framework for universal proxy regularization of model merging

Scientific paper (arXiv:2611.17728) describes a spectral framework for solving parameter interference in merging models without data from other tasks. The authors propose three implementations of a proxy regularizer (Jacobian, Fisher, output-drift) and theoretically prove their properties, including equivalence under NTK…

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

Split-and-scale method for latent 3D representations

Apple Machine Learning Research presents a split-and-scale method for latent 3D representations. The method expands latent capacity at inference without increasing training costs by applying the tokenizer to smaller sub-regions of scenes. It achieves quality competitive with larger models without exceeding the memory of a single…

Apple Machine Learning Research Original source ↗
Research only one source so far

CGMap method for identifying cancer genes using graph learning

The CGMap model uses graph learning to identify oncogene candidates from biomolecular networks. Tested on 12 022 genes, 70 candidates were found. Results were validated by functional tests on cell lines; it outperforms existing methods in interpretability.

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

Cleaning the agent's history hurts when the error persists in the world

Researchers tested whether editing the error history in LLM agents improves recovery. Editing with an annotation increased the success rate to 66% (from 51%), but editing with an action decreased it to 44%. On AppWorld: a false success message decreased the success rate from 45% to 28% when the effect is not carried out. Cleaning the history without changing…

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

Persona-based agents for testing decisions

The research paper presents a framework and infrastructure for testing products and services with AI agents simulating users. The methodology conducts controlled what-if experiments with defined cohorts across multiple domains. A pilot study on 12 industrial use cases shows that persona attributes influence…

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