Tag
A personal project describes a village simulation where a 144M-param julia-1 'gut' model scores villagers' actions on CPU, embeddingGemma retrieves relevant memories, and Gemma 4 e4b (6GB laptop GPU) is only invoked when plans fail or surprises occur. Results show the gut is rarely uncertain, and nightly rule-learning from mistakes cut injuries from 2.73 to 0.53 per valley.
The paper proposes QUAL, a label-free framework that aligns quantiles of time-series foundation models using the model pool as a probability reference, then applies Huber aggregation for robust zero-shot forecasting, reducing average member MASE by 7.4% and CRPS by 9.8% on GIFT-Eval.
The paper provides a formal mathematical framework defining 'any-task inference' (zero-shot capable) foundation models, decomposing them via a universal task reduction into encoder, model, and decoder components and stratifying error into design and realized-model terms, illustrated with five real-world examples.
PokeBot introduces RoboAPI, powered by its Poke-Aurora world-action model, allowing owners of 5 supported robot arms to send natural-language commands via a simple API at 30Hz — no task-specific data collection or training required.
MedJev 是一个可在笔记本本地运行的开源 System-One 医疗模型，在十项医学基准上进行零样本评测，性能接近 GPT-4，并提出两参数再校准方法显著改善患者级概率校准。
该论文提出 RE-ZS-CoT 框架，通过整合领域专家角色来增强大语言模型的零样本链式推理，并在 20 个 LLM 和 7 个基准上评估，最高提升达 44 个百分点。
The paper argues that multimodal reranking can be framed as a single-pass System One choice question for a general vision-language model, showing zero-shot Qwen3.8-27B matching or beating dedicated supervised rerankers on several retrieval benchmarks.
Mistral 发布 Timestral 时间序列基础模型家族，包含 5M 到 209M 参数四个版本，其 209M 模型在 GIFT-Eval、TIME 和 fev-bench 三大基准上均排名第一，支持零样本概率多变量预测且可在 CPU 上推理。
The paper describes team alkhder's submission to Subtask 1b of ImageEval 2026, using a zero-shot 7B vision-language model with constrained selection and Latin-square rotation-averaged logit scoring for multimodal hallucination detection in English and Modern Standard Arabic, reducing CI scores by over 80% against 3B baselines.
LEMON is a graph neural network monitor that uses equivariant learning over attention maps to detect LLM misbehavior zero-shot across unseen models, outperforming baselines by up to 14 ROC-AUC points on hallucination and jailbreak monitoring across 16 LLMs and six datasets.
The paper introduces Modality Self-Play (MSP), a training framework that enables multimodal foundation models to compose across modalities without real paired cross-modal data, using proxy textual descriptions as semantic bridges and achieving state-of-the-art or competitive results on multiple multimodal benchmarks.
微软的研究论文提出在用户轮次边界进行上下文感知的模型路由，发现将会话历史提供给路由器可在相同卸载率下提升多轮电子表格智能体的任务通过率，并分析了提示缓存切换的不对称成本与廉价模型的成本优势衰减。
该论文展示预训练语言模型无需任务特定训练即可零样本完成记录链接与姓名族群推断两项社会科学常规任务；作者提出 EnsembleLink 算法，在记录链接上匹敌或超越既有方法，并证明 LLM 在美国选民文件上的族群推断准确率高达 83.8%，优于 name-supplemented BISG，并可推广到宗教、种姓等其他群体类别。
HiPPT是一个生理学基础的可穿戴PPG表示学习框架，通过多范围文本对齐（记录级形态对齐与跨记录上下文对齐）在超过400万条腕部PPG记录上预训练，在血压、血管僵硬度等六项下游任务的线性探测中取得最佳结果，并支持年龄组、性别和BMI的零样本画像。
Edge0/Audio8-TTS-Realistic-Air is an Apache-2.0 text-to-speech model finetuned from Edge0/Audio8-TTS-Preview-0.6b, supporting zero-shot voice cloning and multilingual synthesis across 25+ languages (zh, en, ja, ko, de, fr, es, and more).
Edge0 released Audio8-TTS-Realistic-0.6B, a compact 0.6B DualAR text-to-speech model with improved speech realism, zero-shot voice cloning, and support for 23 languages (12 more than the Preview version), bundled with a 44.1 kHz neural audio codec and an SGLang serving adapter.
Natura-1 被宣称为面向物理世界的首个通用基础模型，采用500亿参数混合专家（MoE）架构，在一百万小时第一人称手部操作视频上预训练，能够零样本遵循自然语言指令并迁移到从未训练过的任务。
NVIDIA researchers propose NV-Tesseract-AD, a curriculum-enhanced conditional diffusion framework with a Transformer backbone for multivariate time-series anomaly detection, achieving ~3.7× F1 improvement over non-curriculum baselines and strong zero-shot transfer to EasyTSAD benchmarks.
A follow-up project showing a tiny 31K-parameter encoder-only transformer trained only on synthetic T1DM simulator data that zero-shot predicts blood glucose from real CGM traces, with LoRA fine-tuning and ExecuTorch deployment on an Android app.
MotorMind paper介绍了一种方法：冻结的通用VLM可通过中层动作(mid-level actions)和异步反馈像人类遥操作员一样操控机器人，在零样本、无需任务特定训练的情况下实现强机械臂操作能力。