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 introduces SPEM for selecting and combining LLM agent skills

The SPEM system stores individual skills and successful combinations of them for reuse. The authors report a success rate of 87.1 % in the ALFWorld environment and a reward of 38.38 % on the SkillsBench suite, with lower token usage and shorter runtime than the methods used for comparison.

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

Study introduces MAMath-Bench to evaluate mathematical reasoning by multiple AI agents

MAMath-Bench contains 1 003 problems across 11 mathematical categories, incorporating text, images and audio. According to the study, collaboration among multiple agents does not consistently improve answer accuracy compared with strong standalone LLMs, and the systems remain vulnerable to injected errors in the reasoning process.

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

Study introduces Navalia framework for training agents on verifiable financial tasks

The Navalia framework turns corporate financial reports into tasks with traceable supporting evidence and programmatically verified answers for subsequent agent training. The authors report improvements over the baseline model Qwen3.6-35B-A3B across five test suites; the abstract does not state the size of the improvements.

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

The ALTK-Evolve method adds rules derived from previous attempts to customer service agents

According to the authors, an agent using the GPT-4.1 model and rules created with the ALTK-Evolve method achieved a score of 67.5 % Pass^4 on τ²-bench retail: it succeeded in all four attempts on this proportion of tasks. The rules are generated by analyzing previous attempts without weight updates or human annotation.

Hugging Face Blog Original source ↗
Research only one source so far

None of the LLM agents tested in the OR-License evaluation matched the overall quality of experts

The OR-License study tests 11 LLMs on 31 tasks across 10 industrial scenarios. None matched the overall quality of experts. The best-performing primary configuration of the Claude Code tool met strict feasibility requirements in 88.39 % of cases and achieved a DAQ score of 0.87 against a reference of 1.

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

FlashMap accelerates Qwen3-8B inference by up to a factor of 6.1, according to a study

The study introduces FlashMap, a method that proposes an entire block of tokens in a single pass and captures the dependencies between them. For Qwen3-8B, the authors report up to a 6.1-fold speedup with no loss of quality and an average speedup 7 % higher than that of the comparison method DFlash.

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

Study presents LoRA training on a phone using an NPU

The authors present a system that offloads computations involving frozen weights to the NPU during LoRA training, while keeping adapters and optimization on the CPU. They evaluated it on two Qualcomm Snapdragon platforms with four models from the Qwen and Llama families. The abstract does not report numerical results.

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

Study improves drug name pronunciation in a conversational TTS model

According to the study, adding phonemes increased the accuracy of drug name pronunciation, as assessed by the Gemini model, from 54.6 % to 90.7 %. Human evaluators rated 84 % of recordings as correct, compared with 82 % for the best comparison method; the confidence intervals overlapped.

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

Study introduces FusionRare for diagnosing rare diseases

The authors introduce FusionRare, which combines four AI approaches. According to the study, it achieved Recall @1 of 51 % using phenotype data and 79 % when combining phenotype and genotype data across 2 528 clinical cases. In an evaluation involving physicians, it reduced the time to diagnosis by 21 %, according to the authors.

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

CME-SpectrumBench tests spectral data analysis using LLMs

The study introduces CME-SpectrumBench, with 190 questions based on simulated spectral data. When prompted directly, the models underperformed hand-written code; with access to a Python interpreter, they achieved comparable results, according to the authors.

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

Study uses synthetic stories to interpret and influence language models

The authors created vectors for 1 036 concepts using 2 million synthetic stories, which they are making publicly available. According to the abstract, the vectors identify the corresponding concepts across six models and influence model behavior when used to intervene in activations.

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

Persistent memory only slightly improved the success rate of an agent for drawing Chinese characters

The study compared five variants of an agent on 668 tasks. Memory with free-form notes increased the first-attempt success rate by 4.6 percentage points, while memory supplemented with reusable code increased it by 2.8 points. The difference between the two designs was not statistically significant overall.

arXiv cs.AI (Artificial Intelligence) Original source ↗