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

Convergence of stochastic low-rank adaptation

The research paper improves the theoretical analysis of convergence for LoRA (Low-Rank Adaptation). It shows that finding a stationary point requires only O(ε⁻⁴) gradient evaluations instead of an exponential number. It proposes the LoRA-NSGDM and LoRA-STORM algorithms with improved oracle complexity for the stochastic setting.

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

Progress reward modeling in robotic learning: a survey

A survey of progress reward modeling in robotics. Terminal signals are insufficient; robots need feedback while performing tasks. The survey organizes the field from three perspectives: model interface, signal construction methods, and available data and benchmarks.

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

Role Drift in compound LLM systems

Research identifies Role Drift — a problem where modules in compound LLM systems deviate from their assigned roles. The authors propose Role Anchor, a regularizer for detecting and controlling behavior. Experiments showed that 86 % of optimization gains disappeared when the decomposer was kept in its role.

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

Defining AI-native Systems: Autonomy as Authority over Revision

The scientific article proposes a precise technical definition of AI-native: a system in which AI autonomously rewrites its own implementations with an escalation detector, a verification procedure and a guaranteed fallback. It focuses on revision authority (who is allowed to change decisions), not on the capabilities of the underlying…

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

RIS-Kernel: An architecture for long LLM context using sparse attention

RIS-Kernel reduces self-attention complexity from O(N²) to O(N log N) without modifying weights. On Qwen2-1.5B-Instruct, 1% density achieved 75% accuracy at 32k tokens, outperforming the native baseline (71.88%). At 65k tokens, it runs on CPU without GPU, with a 14% gain over zero-context.

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

Hidden representations of Colombian identity in the model Qwen2.5-7B

Pilot research explores whether the model Qwen2.5-7B-Instruct internally represents Colombian identity, socioeconomic status and stereotypes. Using natural language autoencoders, researchers analyzed layer 20 activations on 30 examples (15 Spanish-English pairs). Qualitative evidence examined without…

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

Depth scalability of Logic Gate Networks

Research identifies obstacles to training deeper Logic Gate Networks and introduces Input-Anchored LGN, which consistently improve accuracy up to 100+ layers. The solution connects every layer to the original input, stabilizing training and preserving computational paths. Tested on MNIST…

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

New sampling technique for balanced learning on long-tailed data

Researchers introduced Sharpness-Guided Equilibrium Sampling (SGS), a method for addressing the problem of imbalanced data in deep learning. The method dynamically adjusts class sampling during training based on class frequency and loss sharpness. On the CIFAR-100 dataset, it achieved an improvement of 10.85 points in accuracy…

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

MA-DAR: Continual Learning in Spatiotemporal Knowledge Graphs

The new MA-DAR method addresses the integration of new facts with historical knowledge in continual learning. It eliminates representation conflicts (normalization dominance and semantic blurring) through alignment on a shared manifold and adaptive routing. It improves performance on four benchmark suites.

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

Evaluating adapters for robot locomotion policies using counterfactual analysis

A scientific paper on arXiv proposes a methodology for deciding whether to add a learned adapter to a frozen control policy. Tests on the Go2 robot with 200 iterations show minimal actual improvement (5.2 % of potential, 0.55 % detectable gain); direct queries return a decision of “no”, the VGCC and MPC strategies…

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

A multilayer taxonomy of large language model capabilities

A research paper presents a taxonomy of 14 domain capabilities and 91 LLM subskills, organized into three layers based on cognitive science. An analysis of 15 934 papers from 2023–2025 found that research focuses primarily on linguistic-semantic competence (22.3 %) and reasoning (21.3 %), while six…

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

Cross-domain policy evaluation and learning in contextual bandits

A research approach to evaluating and learning new policies in bandit problems. It combines historical data from the target domain and other domains, addressing limited-data scenarios, deterministic policy logging, and new actions that existing methods could not handle.

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

Detecting household moves in heterogeneous data using entity resolution with an LLM

The research paper proposes a framework for detecting when multiple people move together based on heterogeneous data. It combines an LLM for extracting names and addresses, semantic embeddings, and graph reasoning. On synthetic data, recall increased by 8–15 % and the F1-score by 6–8 % compared with the pairwise baseline.

arXiv cs.AI (Artificial Intelligence) Original source ↗