Specialty
A preprint reports favorable shifts on several gender-bias measures across three language models, while general-capability averages remained close to the corresponding base models.
Scientific publication5 min read
The model was strongest on rhythm abnormalities and weaker on conduction and morphological conditions, with results varying across diagnoses and datasets.
Health information science and systems4 min read
A methods study found higher reproducibility for claim-locked language-model reports in fMRI tests, while separate audits showed why stability is not the same as correctness.
Scientific publication4 min read
A modeling study found that one speech-act-based forecasting design led on three dialogue datasets and in low-data tests, while transfers between datasets remained uneven.
Scientific publication4 min read
A preprint benchmark reports that an episode-based AI memory system scored higher than tested full-context and retrieval methods while using much less retrieved text.
Scientific publication4 min read
A paper-derived training pipeline produced a 60 billion-token corpus and was linked to higher writing scores, while reasoning results stayed close to direct fine-tuning.
Scientific publication4 min read
A preprint study of graph-based question answering found that carefully controlled rewrites damaged performance mainly during subgraph construction, not entity linking or final decoding.
Scientific publication4 min read
DCGC reported the best average score in a six-benchmark comparison, while tests of shuffled drafts suggested that the original draft still mattered.
Scientific publication4 min read
A computational preprint reports lower judged misalignment when selected harmful training answers were replaced with corrections rather than deleted, while several comparisons remained unresolved.
Scientific publication4 min read
The proposal treats AI as a co-scientist for linguistic theory, but its examples are illustrative and do not measure whether the approach improves research.
Scientific publication3 min read
Jev scored higher than two reference models on most benchmark datasets, but its results varied widely across tasks and languages, while a calculation-heavy score raised questions about memorization.
Scientific publication4 min read
SignFLIP combines sign-language translation and generation in one system, but its low human ratings show why benchmark results do not settle whether generated signing is understandable.
Scientific publication3 min read
A proposed fine-tuning method matched sparse attention during training and inference, outperforming sequence-parallel training in most reported tests. The gains came with extra training time, and key results for LongBench V2 were still pending.
Scientific publication4 min read
A new benchmark suggests that large language models can handle parts of theoretical computer-science research in natural language, but struggle when ideas must become machine-checkable mathematics.
Scientific publication4 min read
A secondary analysis of four-person conversations reported that combining speech intensity, gaze and perceived interpersonal closeness distinguished gaps from overlaps better than gaze alone, including in noisy conditions.
Scientific publication4 min read
Twin Worlds compares answers before and after controlled entity substitutions, using inconsistent changes as a reason to withhold an answer.
Scientific publication4 min read
Models performed strongly on written signs of evasion but struggled more with vocal unconfidence, and speaker-level calibration helped only modestly.
Scientific publication4 min read
A new benchmark finds that language models often handle the meaning of Chinese internet neologisms better than the sounds, characters and source forms used to create them.
Scientific publication5 min read
H-Scale uses calibration activations to choose hardware-valid NVFP4 scales, with higher average scores reported across Qwen3 and LLaMA tests but a narrow evaluation scope.
Scientific publication4 min read
A new preprint benchmarks language models on questions aligned with CFA Levels I to III and FRM Parts I to II. Leading systems exceeded 97% on Easy items but fell sharply on Hard cases, while the gated result was 0.39 percentage points higher on held-out questions.
Scientific publication5 min read
A preprint reports better support-versus-attack retrieval in tests of three embedding models, alongside evidence of reduced sensitivity to the claim topic in some fine-tuned comparisons.
Scientific publication5 min read
A Bangla question-answering system reported strong results on a manually built test set, while a small student review found high satisfaction. The study did not test patient outcomes, clinical safety, or performance on an external benchmark.
Scientific publication4 min read
A conceptual framework proposes hybrid AI support for medication optimization under clinician oversight, but the paper reports no validation or real-world performance evidence.
International journal of clinical pharmacy3 min read
A benchmark evaluation of FIRE, a two-stage system that analyzes hate categories before generating replies, reported higher factuality and category accuracy, lower toxicity, greater diversity, and lower peak memory use.
Scientific publication5 min read
The preprint reports that CaRGo-T scored above prompting baselines on benchmark tests of multimodal humor understanding and detection, with gains varying by model and task.
Scientific publication4 min read
A benchmark of 31,316 Sanskrit verse, translation and glossary examples reported the strongest scores for instruction-fine-tuned models, while errors exposed persistent difficulty with phrase boundaries and morphology.
Scientific publication5 min read
An unsupervised system called GUIDE reported lower misspelling rates in a 10-day production comparison and stronger offline scores on two Chinese query datasets, though the study does not establish that the system caused the online gains.
Scientific publication4 min read
A four-language benchmark asked systems to mark hallucinated character spans in image-conditioned text; leading scores were moderate, rankings were uncertain, and empty predictions were common.
Scientific publication5 min read
A prompting method called CritICL uses failure patterns from weaker language models to guide stronger ones at inference time. Tests on Qwen and Llama systems reported modest gains over selected baselines and lower generation and token use, but the results varied across benchmarks.
Scientific publication4 min read
A preprint study found that stylometric features separated human-written from AI-generated text well, but AI-edited text was more often mistaken for human writing and was best separated with editing-aware signals.
Scientific publication5 min read
A preprint reports high detection scores on several adversarial text benchmarks, while results varied across model versions, target generators and attack settings.
Scientific publication3 min read
A benchmark of LLM-generated GPU code found a 2.11× speedup for GPT-5.5 on 22 analytical database queries, while results varied by model and query.
Scientific publication4 min read
A clinician-adjudicated mental-health slice of HealthBench put five leading AI models into a statistical tie. Cross-judge rankings were nearly identical, while only two models produced reproducible empty refusals.
Scientific publication3 min read
A new preprint introduces a dataset and evaluation framework for testing how language models question patients, gather evidence and arrive at a diagnosis across multi-turn conversations.
Chouayfati, Pia, et al. "MTDiag: A Multi-Turn Diagnostic Dataset Towards Clinically Meaningful LLM Evaluation." Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue. 20264 min read
A preprint study argues that sentence-level negative log-likelihood can make cross-language model comparisons less sensitive to tokenization and other surface differences, although important limits remain.
Scientific publication4 min read
A training-free framework called PACE front-loaded retrieved evidence and adapted its reranking budget, with higher reported recall and lower latency in the tested simulations.
Scientific publication4 min read
A modeling study of networked language-model agents found that rival persuaders were associated with more back-and-forth belief trajectories, while direct and peer-relayed exposure tracked with movement toward a persuader’s position. The result applies to a closed simulation, not to humans or real platforms.
Scientific publication5 min read
An analysis of four transformer language models found that token geometry carried a clear signal about grammatical class, while the layer pattern varied by architecture.
Scientific publication4 min read
A three-model comparison gave BanglaBERT the strongest Macro-F1 and external-test results, while BanglaMamba was faster and used less GPU memory.
Scientific publication6 min read
A preprint reports that GRIN, a mixed-policy reinforcement-learning method, performed better on harder questions that required combining supplied facts or drawing inferences while keeping direct recall comparable.
Scientific publication4 min read