已驗證來源
每個故事都連結到最有力的可用證據:可用的原始來源,否則明確歸因的報告。
簡單的英語
發生了什麼、為什麼重要以及值得關注的內容——無需行話。
無填料
當訊號很弱時,我們不會發布任何內容,而是填充提要。
更多故事
9 故事創新
ClueWeaver uses dual-agent reasoning to help compact LLMs analyze long narratives
A newly submitted paper presents ClueWeaver, a reward-trained framework that divides long-narrative question answering between evidence retrieval and interpretation in compact, locally deployable language models.arxiv.org創新
GRIN paper proposes reinforcement learning for continual knowledge injection in language models
A new arXiv paper proposes Golden-GRPO Injection, a three-stage self-learning framework intended to help large language models absorb newly injected knowledge and apply it across paraphrases, combined documents and reasoning tasks.arxiv.org創新
BanglaMamba trades some accuracy for faster, lower-memory Bangla fake-news detection
A new preprint reports that BanglaMamba, a state space model for Bangla fake-news detection, uses less GPU memory and processes data faster than tested BERT-based systems, while trailing BanglaBERT on accuracy and external-dataset performance.arxiv.org安全性
Preprint finds persuasion can spread through LLM agent networks
A new preprint reports that direct messages and peer relays can shift the stances of language-model agents, and argues that evaluations should track exposure histories and actions, not text alone.arxiv.org創新
Preprint finds common multilingual AI metrics can skew cross-language comparisons
A new preprint reports that several normalized metrics used to compare language models across languages can introduce biases linked to tokenization, encoding and orthographic differences. It argues that sentence-level negative log-likelihood over semantically equivalent sequences produces more consistent comparisons.arxiv.org創新
MTDiag tests whether medical LLMs can reason across multi-turn diagnostic conversations
A new dataset evaluates large language models as diagnostic agents in interactive clinical encounters rather than through isolated questions. MTDiag combines cases from emergency medicine, clinical records and published case reports, and proposes knowledge-grounded metrics beyond diagnostic accuracy.arxiv.org創新
HealthBench-Psych introduces an open benchmark for mental-health AI performance
Researchers created HealthBench-Psych, a 610-conversation mental-health subset of OpenAI’s HealthBench, and released its evaluation pipeline, model responses, grades and analysis code.arxiv.org創新
Survey maps how foundation models can learn from their own outputs after training
A survey accepted to Findings of EMNLP 2026 catalogs 80 methods for adapting foundation models without human labels, preference data, stronger teachers or executable verifiers. It warns that internal learning signals can either improve a model or recursively amplify its errors.arxiv.org創新
Study traces dialect bias through every stage of language-model development
A new preprint reports that language models encode dialectal English unequally during tokenization, pre-training, post-training and inference, even when paired texts have the same meaning.arxiv.org