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

Study uses the SteerScope suite to assess the effectiveness and side effects of steering LLM behavior

The study introduces SteerScope with 15 metrics and compares 23 methods. According to the authors, none of the evaluated activation steering methods outperformed Prompt Steering in effectiveness without greater overall side effects. For inputs outside the original distribution, side effects often intensified.

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

Study introduces HARP for evaluating LLMs on a fixed budget

HARP continually reallocates unused budget during LLM evaluation. The authors prove that the method stays within a fixed resource limit and establish statistical guarantees for the estimates. Experiments cover task completion by agents, jailbreaking, toxic content and hallucinations when using RAG.

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

Study introduces HARP for coordinating AI agents with hidden preferences

The study introduces HARP, which estimates the preferences of individual agents using explicit probabilities updated by Bayes' rule. The HARP+ variant supports steps that refine these estimates and, according to the authors, achieved the best results among methods without knowledge of the true preferences.

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

Study introduces WASD method for LLM distillation using token meaning

The WASD method accounts for semantic similarity between tokens during LLM distillation using their embeddings and Sinkhorn divergence. According to the authors, it improves performance on instruction following, mathematical reasoning and code generation. The implementation is publicly available.

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

Study examines model compression for translating biomedical texts

The authors investigated a combination of knowledge distillation and quantization for biomedical translation from French into English. According to their experiments, model size decreased by 69 %, inference speed increased by 98.21 % and CO2 emissions fell by 98.46 % without any loss in translation quality.

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

Study used AI to estimate Great Lakes water levels and explain their changes

The study authors trained eight algorithms on monthly data for four of the Great Lakes from 1982–2022. According to their results, the four best-performing algorithms had an error of about 12 cm compared with 14–21 cm for the simplest models. The SHAP method explained the contributions of the input variables.

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

Study introduces JEPA-Anything model for predicting states across domains

A team led by PhAI Labs introduced JEPA-Anything, which combines several component predictions into a single prediction. According to the authors, in a test using a simplified Pong environment, it reduced error by 35 % compared with standard JEPA; for unseen combinations of interventions, it reduced error by 13 %.

The Decoder (daily AI news) Original source ↗
Research only one source so far

Study: malicious instructions in comments led Code LLM models to generate vulnerable code

A study of ten Code LLM models with 3–13 billion parameters found vulnerabilities of at least medium severity in 77.4–92.3 % of outputs generated with planted instructions in comments, compared with 1.7–5.1 % without them. The most effective follow-up check missed approximately a third of vulnerable outputs.

arXiv cs.LG (Machine Learning) Original source ↗