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

MissionBench: A benchmark for evaluating multimodal models in autonomous flight missions

MissionBench is introduced with 120 missions in 5 simulated 3D environments for evaluating 22 multimodal models. The best model achieved a success rate below 35 percent (humans: 84.4 percent), illustrating the complexity of multilevel embodied tasks. Results show that larger models have better capabilities…

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

SceneActBench: Evaluating agents' ability to act in 3D environments

Researchers introduced the SceneActBench benchmark for testing vision-language model agents capable of acting in 3D scenes. The benchmark contains five tasks with 520 cases. Testing eleven proprietary VLM configurations produced scores of 38.6–50.2 without consistent performance.

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

Context anxiety in large language models: Why they fail despite sufficient capabilities

The scientific study describes the phenomenon of ‘context anxiety’: reasoning models have the capabilities to solve problems but fail because of premature self-doubt and poor estimates of the tokens needed. The authors show that models can learn better strategies without this problem, without having to grow in size.

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

Benchmarking fine-tuning and retrieval strategies on the NRC Reactor Operator examination

A study evaluates Gemma 4 31B-IT on 14 NRC Reactor Operator examinations from 2015–2021. Comparing 8 fine-tuning and retrieval-augmented generation configurations showed that SFT with fixed-size chunking RAG passed 8 of 14 examinations (79.7 % accuracy). RAFT lagged behind SFT; the choice of chunking strategy…

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

A semiotic logical hexagon for logical reasoning in LLMs

A research paper proposes HexLogicAgent, a framework that improves logical reasoning in LLMs by organizing semantics. The key finding: failures arise more from weak semantic representations than from deductive logic. Logical hexagon theory models the complete structure of semantic oppositions, which slows…

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

J-CoT: Chain-of-Thought in J-Space

A research approach combining chain-of-thought with latent representation. J-CoT proposes an intermediate representation based on vocabulary-indexed coefficients instead of fully verbalizing each step. In tests, it outperformed existing latent-reasoning methods on mathematical, scientific, and coding tasks.

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

Evaluation design affects conclusions about MeSH terms in medical classification

A study examines how evaluation design affects comparisons of expert-assigned and automatic MeSH terms for classifying medical abstracts. On the Statins topic, the difference between approaches (WSS@95%) changes from +0.096 in 5-fold evaluation to +0.021 in 10-fold evaluation. In the 10-fold design, BiomedBERT and bag-of-words produce similar…

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

Causal dependence vs. attention: how to train sparse attention correctly in transformers

A study finds that attention patterns (where a model looks) differ from causal dependence (what its decisions depend on). Sparse attention trained on attention achieves 41 % accuracy, while training on evidence achieves 99 %. The problem also appears in pretrained models: Gemma-2-9B improved from 56 % to 99 % accuracy when it was…

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

Neural Atom Prevalence: A Bayesian method for neural network compression

A research paper introduces Neural Atom Prevalence (NAP), a Bayesian framework for structured neural network compression. The method reduces active nodes to 8 % of the original architecture on MNIST while retaining accuracy and providing calibrated uncertainty estimates (93.4 % coverage against a 95 % target).

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

Developing and validating a Spanish version of the psychological dependence on LLMs scale

A study validates the Spanish version of a psychological dependence on large language models scale (LLM-D12-SP). A sample of 386 participants confirmed a two-factor structure: instrumental dependence (using LLMs for tasks and decisions) and relational dependence (psychological attachment). Cronbach's…

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

A factorial study of synthetic data generation for machine translation of low-resource languages using grammar reference books

The research team used LLMs to extract grammatical rules and examples from grammar reference books and generated synthetic parallel corpora for fine-tuning translation models. Testing on three endangered languages (Kalamang, Tuatschin, Mandan) showed an improvement of +8.8 to +3.3 ChrF++ points. The study maps…

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

Dynamic commonsense coordination for generating empathetic responses

DCC, a research framework for generating empathetic responses, improves commonsense coordination through three modules (SCE-AttnRes, AGCF, ICAD). On Empathetic-Dialogues, it achieved higher emotion classification accuracy and better response diversity than the CEM baseline.

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

Stable quantum federated learning for intelligent services

An arXiv research contribution proposes DUQFL-Prox, a framework for quantum federated learning. It aims to address instability and unfair performance between clients in heterogeneous distributed environments without sharing local data. Experiments on fraud detection and genomic classification show improved stability and…

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

Multi-agent system-driven digital twins for predictive maintenance: Architectures, technologies and open research questions

A systematic review of 547 scientific articles analyzes the intersection of Multi-Agent Systems and Digital Twins focused on predictive maintenance in industry. The study identifies three main open research questions: deploying AI on resource-constrained microprocessors, distributed coordination through lightweight communication protocols and…

arXiv cs.AI (Artificial Intelligence) Original source ↗