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

Uncertainty-aware trust estimation in LLM systems

A study proposes weighting multiple language models by the quality of their uncertainty rather than assuming equal trustworthiness. A Cooke-style method penalizes confidently wrong predictions. Tests on MMLU and MMLU-Pro showed robustness to unreliable models in heterogeneous panels.

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

PlanE: Data, tuning and inference planning for extractive LLMs

The PlanE research framework optimizes datasets for tailoring LLMs to specific tasks with the aim of reducing annotation costs. It includes a DTI planning algorithm for selecting the optimal model and parameter combinations. The code is publicly available.

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

PersonaTrail: A Benchmark for Personalized Web Agents Through Browsing Histories

A research paper on arXiv introduces the PersonaTrail benchmark for evaluating web agents capable of working with underspecified instructions and inferring context from browsing history. It includes the Preference-Aware Contextual Memory (PACMem) framework, which distinguishes between factual and preference memories for…

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

Exact realization of ReLU in affine iterations of refinement operators using residual memories and offset frames

A mathematical study of how neural networks with the ReLU activation function can be constructed to exactly implement certain iterative mathematical operations. The main contribution is a new residual memory controller algorithm, which enables efficient implementation with a depth of O(n).

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

Adaptive reinforcement learning with multiple time horizons

Researchers propose a reinforcement learning method that adaptively selects and combines multiple time horizons instead of a fixed discount factor. It enables better adaptation to changes in the reward structure without manual tuning. The method was tested in MiniGrid environments, including continual learning with…

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

LLM watermarks compromise the quality of medical text

A study tested 5 watermarking schemes (tracing marks) on 11 LLMs and 7 VLMs in medical tasks. It found that watermarks cause performance degradation — lexical errors, hallucinated specialist terminology and errors in interpreting medical images. Aggregate metrics usually hide these problems.

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

CEDAR: Detecting causal edges in autoregressive processes

A scientific article on CEDAR, a method for detecting causal edges in sparse autoregressive time series. It combines screening of candidate lags using U-centered distance correlation with conditional independence tests, requires O(d²) tests and is most effective in data-scarce situations.

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

SevDiff: A diffusion model for generating vehicle trajectories with controlled collision severity

A research paper on arXiv introduces SevDiff, a diffusion model for generating realistic vehicle trajectories with precisely controlled collision severity (Time-to-Collision). The model achieves 100% accuracy for TTC 0.5–1.5 s and 97–99% for 2.0–2.5 s, trained on 468 interaction sequences from express…

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

When Repetition Becomes an Algorithm: Convergence in Looped Transformers

The research examines how transformers with weighted layers applied T times implement algorithms. The authors identify four key findings: (1) budget law — a linear computational bound related to training conditions (v ~ n_train/T_train), (2) architecture determines the type of algorithm (parallel vs.…

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

Neural predicates for investor views in the Black-Litterman model

An academic preprint proposes a formal approach to the Black-Litterman portfolio construction model that uses neural predicates to automatically generate investor views and their uncertainty. The model is interpretable and fully differentiable, enabling end-to-end learning from financial data.

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

Attention degradation in transformers: A mechanism or a side effect

A study tested whether attention degradation in transformers (GPT-2, LLaMA, OPT) limits contextual retrieval. Adding commas at syntactic boundaries reduces degradation over 40–80 tokens, but an attention intervention (Relay-Aware Attention) did not improve performance. Conclusion: function tokens contribute through…

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

Using biokinetic prior knowledge to model bioprocesses with limited data

A research study examines how to incorporate biokinetic knowledge from ODE models into neural networks for bioreactor prediction. It compares a data-based approach (pretraining on simulations) with an architectural approach (embedding the ODE in the network). Both outperform the baseline without prior knowledge on 11…

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

Geometric deep learning in polypharmacological drug design

An academic review of geometric deep learning (GDL) applications in designing drugs that target multiple proteins simultaneously. It analyzes architectures from graph neural networks to SE(3)-equivariant diffusion models. The emphasis is on generative models with reinforcement learning for resolving conflicts between multiple binding sites.

arXiv cs.LG (Machine Learning) Original source ↗