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

Academic study introduces MoE²-LoRA, a method for efficient fine-tuning of Mixture-of-Experts models

Researchers described MoE²-LoRA on arXiv, a method for fine-tuning Mixture-of-Experts language models. The method uses a shared pool of LoRA experts and routing from the original model; according to the authors, it achieves the best accuracy results while preserving the general capabilities of the model.

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

TRACE-Router: Consistent and adaptive online routing for agentic AI

Research on a framework for routing between LLM models in long-running agentic workflows. Unlike traditional routers that make decisions separately for each call, TRACE-Router assigns one model to a task and all LLM calls use it. The policy is updated according to the final outcome…

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

Style attacks on multimodal models: optimizing vulnerabilities through GRPO

Research identifies an asymmetry in multimodal LLM safety: although they understand content regardless of visual style, their defenses can be bypassed with style-optimized modifications. The authors propose ASO, a module that uses a GRPO agent to amplify these attacks. The code is publicly available.

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

Persistent computational state: A session-centric runtime for generative world models

The research article shows that video world model failures in simulation tasks are caused by runtime infrastructure rather than the models. Request-centric serving loses non-recomputable state that is critical for backtracking. The authors propose a session-centric runtime with Persistent Computational State. Checkpoint/restore…

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

A monotonic framework for evaluating operational risk in wildfire firefighting operations

The study proposes a new framework for evaluating wildfire risk prediction systems that, instead of traditional metrics, measures whether increasing risk scores consistently correspond to greater operational demands. A comparison of the expert-based DFE index, GRU models and FARS (a hybrid system with an LLM) in French…

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

FlowEvo: Agents with self-evolving workflows and executable skills

FlowEvo, a research framework, enables LLM agents to improve continuously without training by compiling successful solutions into reusable skills. It links resource extraction, retrieval for future tasks, and filtering of unhelpful skills. On ALFWorld benchmarks…

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

Agent-based root-cause diagnosis of industrial anomalies with grounded reasoning

A research team created AgentRCA, a framework for automated fault diagnosis in industrial equipment. The system combines a model of normal dynamics with the capabilities of a language model and achieves competitive results without training on specific faults. It also provides readable explanations of the diagnosis.

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

Conversational agents tailored to language proficiency to reduce foreign language anxiety

The research paper describes a multi-loop system of LLM agents for language learning that generates dialogue tailored to students’ CEFR level. In a pilot study with Japanese students, 87.4 % of sentences were within the expected level (compared with 54.1 % without adaptation). A statistically significant reduction in anxiety was not demonstrated…

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

A discrete action space as a prerequisite for GRPO convergence in continuous control

The research tests Group Relative Policy Optimization (GRPO) on quadcopter control using the Qwen-0.5B model. Without modifications, the model collapses to a zero action (0% success), but replacing the continuous action space with a 5-way choice of PID presets achieves 98.6–100% success. The classical PID method achieves…

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

A hybrid system for predicting roadblocks in Bolivia

A research paper combines time series modeling (Prophet) with NLP on a 6-year corpus of Bolivian newspapers. The system achieves an AUC-ROC of 0.677 for one-day forecasting and reduces Brier Score by 10.9% compared with statistical models, capturing signals of social tension in media discourse to predict road…

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

What red-team evaluations of AI models can and cannot prove

Researchers define a mathematical "evidential ceiling" framework for measuring what red-team benchmarks can establish about AI safety under a limited budget. They found that current tests are adequate for common harms, but millions of times insufficient for rare catastrophic incidents.

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

Physiological signals for detecting talking-face deepfakes

A research team proposed a detection system for talking-face deepfakes that synthesize video from a photograph and audio. The method extracts physiological signals (rPPG) through RhythmFormer and trains lightweight classifiers. On Celeb-DF++, a 1D ResNet achieved an AUC of 0.806; performance varies considerably…

arXiv cs.LG (Machine Learning) Original source ↗