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

MA-SafeDiffuser: safe multi-agent planning with diffusion models

A research paper presented MA-SafeDiffuser for safe multi-agent planning. The method extends SafeDiffuser with diffusion probabilistic models and guarantees safe trajectories using barrier functions. It reduces safety constraint violations while maintaining planning quality. An implementation is av…

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

Analysis of rhetorical structures in legislative hearings using an NLP pipeline

The research paper presents an NLP pipeline for automatic analysis of rhetorical structures in legislative hearings. Logistic regression with grammar-based masking achieved an F1 score of 52.8% in classifying discourse sections. The proposed grammar successfully parsed 55.21% of Californian legislative…

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

Fractal geometry of basins and the limits of predictability in deep learning

Research in Nature Machine Intelligence shows that very small changes in parameter initialization (even a single bit) lead to different training solutions. The authors propose the fractal geometry of "riddled basins" as a mechanism explaining the inherent unpredictability and non-reproducibility of deep neural networks.

Nature Machine Intelligence Original source ↗
Research only one source so far

MIT experiment: AI model fixed errors in a foreign neural network, increasing accuracy from 8 to 48 percent

MIT research verified the ability of the AI model Astra to diagnose and fix errors in foreign neural networks. After intervention in the Qwen model, success rate in tests increased from 8 to 48 percent. The process requires analysis of internal representations; on larger models it is hindered by limited context length.

Research only one source so far

The open model inference market: exponential growth according to Interconnects

Analytical report from AI2: the open model inference market is growing exponentially. Fireworks serves 40+ TB of tokens daily to custom models, a 2x increase since summer. A comparison with the Google API shows a smaller segment than proprietary solutions, but with a growing share. The data comes from the Interconnects Dashboard.

Interconnects (Nathan Lambert, AI2) Original source ↗
Research only one source so far

Study identifies four critical barriers to adopting AI and blockchain in African horticulture

Academic researchers published a literature review (2005–2025) on AI and blockchain in horticulture in Africa. They found that blockchain cannot guarantee the correctness of the original data and that rules are missing for information ownership and the allocation of benefits to small farmers. They propose combining the predictive capabilities of AI with…

The Conversation — Artificial Intelligence Original source ↗
Research only one source so far

LLM capability degradation during fine-tuning: geometric theory and the AEGIS loss

Research explains the loss of general capabilities in large language models during fine-tuning (PCD) as a consequence of feature interference in the activation space. The authors propose the AEGIS loss, which preserves capabilities at a 1.3× computational cost. The improvement was verified on 3 model families and 14 tasks.

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

Testing the robustness of LLMs for causal inference

A study maps why LLMs perform poorly in causal reasoning. It tests 6 dimensions: model size, graph complexity, prompting strategy, metadata quality, data quality, and sensitivity to uncertainty. It finds "optimism bias" – models overestimate causal relationships and show low accuracy. Model scaling…

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

Speculative decoding with adversarial training improves the accuracy of long-horizon agents

Research from arXiv shows that speculative decoding (LLM acceleration) affects the decision-making quality of long-horizon agents. The authors demonstrated numerical noise in the verification process that alters the model's output distribution after 30 steps. A new approach trains the draft model on adversarial preferences in critical…

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

GlucoAgent: personalized glucose forecasting from multimodal data using an agentic system

The study presents GlucoAgent, an LLM-based agentic system that dynamically analyzes multimodal health data (CGM, smartwatch, meal logs) for glucose level forecasting. It achieves an average RMSE of 22.5 mg/dL, outperforming the Chronos-2 model (23.5 mg/dL). Multimodal data improves GlucoAgent's accuracy…

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

Benchmark of metalinguistic reasoning in language models across 195 languages

A new benchmark called Typological Feature Induction (TFI) tests the ability of language models to inductively infer typological features of unfamiliar languages. Tests on 195 languages from Grambank and multilingual corpora showed 59–80% accuracy in identification from 16 sentences, but the models rely on prior knowledge. Na…

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

Self-Privileged Fine-Tuning: research on tokenwise weighting for training language models

A research paper proposes Self-Privileged Fine-Tuning (SPFT), a technique for better training of language models. Instead of random token weighting, it uses a frozen copy of the model with privileged information (e.g., the correct answer). Experiments show improvement on mathematical reasoning and code generation tasks.

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

Latent Cluster Separation as a Regularizer in Lifelong Anomaly Detection

A research paper examines the separate effects of two mechanisms in VAE-based systems for continual anomaly detection: the regularizer (latent cluster separation) and scoring (distance-based). Findings: the regularizer improves detection and backward transfer independently, but the main instability stems from the scoring…

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

EvoGroup: adaptive memory strategies for group agents

The research paper presents EvoGroup, a system of memory strategies that self-evolve and adapt through interaction with the shared context of group agents. On the EverMemBench benchmark, it achieves an 89.58% success rate (an increase from 71.07%) and outperforms human-designed solutions. It generalizes effectively to…

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

AdSteer: Native Advertising in Large Language Model Generation

The paper describes the AdSteer framework for integrating native advertising into LLM outputs. The method adjusts token probabilities during decoding and achieves a 14.1% improvement in score across GEM-Bench with 22.5% lower tokenization requirements than rewriting methods.

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

Language Model in Decision Transformer for Offline Multi-Task Learning

The research paper proposes LaMA-DT, a method combining a language model with a Decision Transformer for learning from offline data. Natural language guides the augmentation of training trajectories. Experiments on Meta-World and MuJoCo show improvements, especially in low-data regimes.

arXiv cs.CL (Computation and Language / NLP) Original source ↗