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

Robust smart prediction with optimization

A research paper proposes a method for robustly integrating forecasting and optimization that accounts for data bias in feature space. It establishes theoretical guarantees of exponential decay in approximation error and experimentally shows improvements over standard methods.

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

The Hard Decision Layer: How transformers commit to predictions

Research identified a phenomenon in transformers where models abruptly commit to answers at a particular layer — the Hard Decision Layer (HDL). Tests on four models (Qwen, Llama, Granite, Mistral) and four datasets showed consistent HDL emergence without learned routing strategies. On…

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

TRACTA benchmark for temporal reasoning in complex systems

The article introduces the TRACTA benchmark with three tasks (early warning, pattern detection, run classification) for testing temporal reasoning in event-driven systems. Neuro-symbolic models working with semantically grounded trajectories achieved the best results compared with purely…

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

Detecting depression on social media using language model embeddings

Research compares large language models and traditional classifiers for depression detection. LLMs dominated binary classification but lagged in severity prediction. Supervised models trained on LLM summary embeddings were more accurate, particularly for multiclass classification.

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

Quasi-Monte Carlo initialization for meta-reinforcement learning

A research paper examines quasi-Monte Carlo weight initialization for meta-reinforcement learning. It shows improved training convergence on similar unseen tasks compared with orthogonal defaults (SB3), but the orthogonal method performed better on dissimilar tasks.

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

Evaluation-as-a-Service architecture: a microservices approach to AI evaluation with coverage guarantees

A scientific paper describes an Evaluation-as-a-Service architecture with six Kubernetes microservices for AI evaluation. It combines conformal prediction, fairness monitoring, and drift detection. Validation: empirical coverage at the target level, 100% drift detection power, and latency below 2ms for a batch of 100.

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

Variance-reduced distributed Q-learning in changing networks

A research paper introduces VRDQ, an algorithm for multi-agent reinforcement learning over distributed networks. Agents share information about the Bellman operator and achieve linear speedup through collaboration with minimal communication overhead.

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

Towards standardized testing of personal LLM agents over time

A research paper proposes a new framework for evaluating personal LLM agents. The authors identify four conditions for effective testing: explicit temporal interventions, persistent user state, cross-component effects, and configuration variation. After auditing published benchmarks, they found no protocol…

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

Portable latency prediction for optimizing LLM deployment on heterogeneous edge devices

The research team presents a framework for more accurate latency prediction when deploying LLMs on mobile and edge devices. On Pixel 8, it improved decoding accuracy from R² 0.957 to 0.973, and on Pixel 8 Pro, prefill latency from -1.383 to 0.966. The method combines static parameters with dynamic hardware telemetry.

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

Improving compositional generation in diffusion models through intrinsic reward

The research paper introduces TILT, a method for improving image generation from text descriptions. It addresses diffusion model failures on complex prompts involving multiple concepts. The method uses the model's intrinsic reward without external supervision. Results on the T2ICompBench benchmark show improvement without reducing quality…

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

CEL: A library and benchmark for AI explanation methods

Introduces CEL, a library with 18 datasets and 14 counterfactual explanation methods. It includes a unified evaluation protocol with metrics for validity, coverage, sparsity, and plausibility. The first comprehensive benchmark for systematically comparing methods within a unified framework.

arXiv cs.LG (Machine Learning) Original source ↗