This work shows that $\beta$ entangles two distinct roles: it governs the effective inverse preference-noise scale and simultaneously rescales the optimization dynamics, coupling this scale with the effective step size, and proposes a centered-softplus reformulation that is argmin-equivalent to DPO for $\beta>0$, while making the inverse preference-noise-scale and learning-rate effects explicit and independently tunable.
Experimental results show that CritICL consistently outperforms standard in-context learning and achieves performance competitive with or superior to test-time scaling methods, while requiring significantly fewer generations and lower token cost.
Yu-Fan Wu, Yinghui He, Zhengyi Hu et al.· 1 citation
Large-scale, standardized datasets have driven many advances in AI-based scientific modeling, from protein structure prediction to natural language processing. Infectious disease epidemiology is increasingly adopting AI methods for forecasting, surveillance, and outbreak analytics, but the time-series data available to train them remains orders of magnitude smaller than the corpora behind the advances seen in other fields. Because the scope of real-world epidemiological data cannot practically reach the scale needed to train truly large-scale AI methods, simulated data provides a possible alternative. Here we introduce Reservoir, a large open simulator and dataset of realistic epidemic simulations in which every trajectory carries complete ground-truth labels, including quantities that cannot be measured directly in a real outbreak, such as true infection counts, time-varying reproduction numbers, and counterfactual intervention effects. Reservoir is generated by a stochastic simulator with realistic noise and reporting artifacts, together with interventions with configurable timing, compliance, and age-dependent efficacy. The current release contains 500,000 outbreak trajectories spanning one billion simulated days across diverse pathogen characteristics, population structures, and intervention regimes. Reservoir enables counterfactual experiments, surveillance-design studies, and training of epidemic models at a scale real-world datasets cannot provide.
This work introduces Visual Retrieval Heads (VRHs), a small subset of attention heads that are causally responsible for grounding text descriptions to image regions, and shows that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads.
Chanho Park, Daehyeon Choi, Jihyun Lee et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
This report presents an open pretraining recipe that trains a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs, and derives a Puro Cost Scaling Law that relates training cost to average model performance.
Kairong Luo, Jia-Rui Cui, Yao-Rui Yin et al.· 0 citations
This study provides a systematic comparison of several open-source LLMs on a structured intestinal symptom extraction task and concludes that Qwen3 models offer a favorable balance between accuracy and efficiency, making them suitable for resource-constrained scenarios.
Xinyue Zhang, Quanyu Wang, Beibei Liu et al.· Journal of Medical Internet...· 0 citations
CMPM, a Chinese Multi-Panel Meme benchmark with 1,214 annotated samples covering five structural types, ordering dependency, panel-order constraints, and optional comment context, is introduced and results indicate that canonical-display accuracy is not by itself evidence of order understanding.
Hai-Han Li, Hai-Hao Li, Zhengjie Xu et al.· 0 citations
FLARE is proposed, a novel framework that endows VLAs with robust error recovery capabilities through a ``Retry" and ``Reset" Paradigm, and significantly improves task success and robustness.
Ganlong Zhao, Zijia Tang, Xingping Chen et al.· 3 citations
The effect of introducing the manager-worker scaffold over a shared filesystem workspace, with no training and no per-benchmark tuning, measured against the same model answering in a single pass is investigated, finding several mechanisms behind the gains.
Victor Gao, Vida Khosrowshahi, Ali Khosrowshahi et al.· 0 citations
This paper presents Ankhdjet, an open-source compiler that lowers a HuggingFace ternary checkpoint to a via-mask program of a fixed compute-in-ROM macro on the open SKY130 PDK, verified end to end with open tools.
BekchiAI is a complementary web-based observability and control layer for deployed agents, providing full token and latency telemetry as well as remote run termination and the benchmark, evaluation tools, and platform are publicly released.
Results show that a sub-million-parameter recurrent generative policy can achieve strong performance on modern language-conditioned manipulation benchmarks while providing an explicit mechanism for prediction-error-driven online state correction.
