New benchmarks and studies are mapping LLM progress toward action, interaction, scientific use, and specialized multimodal tasks. Researchers are also questioning priority claims and highlighting open precedents for non-autoregressive architectures.
inclusionAI has published two checkpoints on Hugging Face: Realtime-Venus-Omni, a 9B audio-visual interaction model with proactive and full-duplex behavior, and Realtime-Venus-Audio for audio-driven conversations. Both share a streaming backbone and ...
MDT turns divergence among face, voice, and text into the central signal for clinical analysis, reaching 0.7408 Macro F1 on the BAH test and training in less than twenty minutes on a single GPU. For AI-Radar the key insight is operational: a model en...
MDT turns contradictions between face, voice, and language into an explicit signal. With nine tokens, FiLM modulation, and LoRA fine-tuning, it reaches 0.7408 Macro F1 on the BAH test split and 0.7368 on the private leaderboard, outperforming the str...
A dual-stage framework combines fine-tuning with a diversity reward and query-level credit assignment to balance individual suggestion quality and intent coverage. Offline and online A/B tests on production data show improvements in CTR, quality, and...
A systematic mapping of 14,767 arXiv preprints published between January 2022 and August 2026 shows how LLM benchmarks are shifting toward action, interaction, and professional applications. Model participation is asymmetric: LLM-based scoring grows,...
A researcher claims they published a non-autoregressive architecture with model, dataset, and paper on arXiv in March 2025, later echoed by a frontier lab without open weights or dataset. The case highlights a structural problem: independent open sou...
A University of Washington Medicine study uses an LLM to extract features from unstructured clinical notes and feeds them into a logistic regression pipeline alongside structured data. The result is improved extubation failure prediction. The authors...
A new study on grokking shows that generalization in transformers is not a module switch: it emerges as a distributed gain with a prospective block-0 attention bias and spectral recoding. Selected degree-two modes account for 67–92% of the addition c...
A two-tier protocol certifies that an updated model is no worse than the previous one without labeling all traffic. The risk difference concentrates on inputs where the models disagree: only disagreement is checked. Across 14,000+ streams and 785 pai...
An independent researcher shared on arXiv and Hugging Face a non-autoregressive architecture for probability prediction with a JSON schema, a year before a frontier lab presented it as a breakthrough. His vertical model uses PPO over sequence embeddi...
Researchers propose CNSF, a causal neural set filter for multi-target tracking that encodes only current measurements and carries past evidence in a recursive state. On a simulated three-regime test set, CNSF reduces GOSPA and T-GOSPA relative to Tra...
InternLM introduces Intern-S2-397B, a 397-billion-parameter multimodal model for scientific intelligence and long-horizon agents. It learns from raw scientific literature pages without intermediate parsing, combining symbolic and visual reasoning, an...
