PreprintNot peer reviewedFrom RopediaResearchers report a cross-embodiment world model trained on human and robot videoOriginal title: Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human ExperienceA preprint describes a world model that predicts loco-manipulation outcomes across humans and robots using shared kinematic conditioning.Added 9 Nov 2026
PreprintNot peer reviewedFrom Google / Google DeepMindResearchers report multi-agent system MaxKernel automates TPU kernel writingOriginal title: MaxKernel: Agentic Kernel Generation for TPUsA preprint describes MaxKernel, a multi-agent LLM system that writes and tunes TPU kernels, reporting speedups over compiler baselines.Added 9 Nov 2026
PreprintNot peer reviewedResearchers report a multiscale method speeding up zeroth-order LLM fine-tuningOriginal title: ENGINE: Endogenous Variational MultiScale Optimization for Zeroth-Order LLM Fine-TuningA preprint introduces Engine, a zeroth-order fine-tuning method that mixes low- and full-dimensional updates for faster, more accurate LLM training.Zhuoli Ouyang et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint reports curriculum training can make safety alignment more robustOriginal title: Curriculum Learning for Safety AlignmentResearchers report a curriculum-based method for DPO safety training cuts harmful outputs and jailbreak success rates.Sandeep Kumar et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint reports faster private LLM fine-tuning by fixing noisy gradient structureOriginal title: Re-examining low-rank adaptation for private LLM fine-tuningA preprint proposes denoising DP-SGD gradients to restore their low-rank structure, reportedly speeding up private LLM fine-tuning.Ali Dadsetan et al. · Added 9 Nov 2026
PreprintNot peer reviewedFrom NVIDIAResearchers scale egocentric human video to 50,000 hours for robot foundation modelsOriginal title: Open-HumanScale: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation ModelsA preprint reports a log-linear scaling law linking egocentric human video pretraining hours to robot task success, plus a hybrid human-robot model.Added 9 Nov 2026
PreprintNot peer reviewedResearchers propose UniSD, a unified self-distillation framework for LLMsOriginal title: UniSD: Towards a Unified Self-Distillation Framework for Large Language ModelsA preprint presents UniSD, a framework combining self-distillation techniques that the authors report improves over baselines without external teacher models.Yiqiao Jin et al. · Added 9 Nov 2026
Conference paperReview status unknownFrom MetaA survey maps techniques for fitting LLMs onto edge devicesOriginal title: Large Language Model Deployment on Resource-Constrained Edge Devices: A Practitioners SurveyA conference paper reviews software and hardware approaches for running large language models on memory-limited edge devices.Added 9 Nov 2026 · Released Jan 2026
PreprintNot peer reviewedFrom China Telecom (TeleAI)A preprint reports a training method to shrink edge multimodal AI models with little accuracy lossOriginal title: ProQuant: Progressive Quantization-aware Training for Edge MLLMsResearchers describe ProQuant, a quantization-aware training method that keeps small multimodal models accurate at low bit-widths for edge devices.Added 9 Nov 2026
PreprintNot peer reviewedA preprint reports scale-dependent recipes for training code verifiers cut RLVR costsOriginal title: Aletheia: What Makes RLVR For Code Verifiers Tick?The authors report that which parts of the RLVR pipeline matter most for code verifiers depends on model size, offering cheaper training recipes.Vatsal Venkatkrishna et al. · Added 9 Nov 2026
PreprintNot peer reviewedResearchers propose a cross-domain benchmark for LLM agents doing black-box optimizationOriginal title: A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box OptimizationA preprint introduces AgenticBBO-Bench, testing LLM agents against direct LLM methods and numerical optimizers across five task domains.Ming Chen et al. · Added 9 Nov 2026
PreprintNot peer reviewedResearchers report clinical LLM safety depends on deployment setup, not accuracy gainsOriginal title: Safety and accuracy diverge across deployment conditions in clinical large language modelsA preprint finds that curated evidence, not model scale, drives safer radiology question-answering in LLMs.Sebastian Wind et al. · Added 9 Nov 2026
PreprintNot peer reviewedPreprint proposes federated GRPO method using reward signals to cut communicationOriginal title: Fed-GRPO: Reward-Signal-Driven Federated Group Relative Policy OptimizationA preprint describes Fed-GRPO, which uses GRPO's own reward statistics to guide aggregation, local training and compression in federated reasoning training.Pengxin Guo et al. · Added 9 Nov 2026
PreprintNot peer reviewedPreprint finds humans far outpace AI models at learning grid-world dynamicsOriginal title: Benchmarking World-Model Learning with Environment-Level QueriesA preprint reports humans substantially outperform five frontier AI models on a new benchmark testing learned world models.Archana Warrier et al. · Added 9 Nov 2026 · Released May 2026
PreprintNot peer reviewedResearchers report a scalable way to build concept vectors for AI alignmentOriginal title: From Stories to Features: Scaling Concept Vectors for AI Alignment & InterpretabilityA preprint describes extracting over a thousand interpretable concept directions from language models using synthetic story pairs.Dmitry Scherbakov et al. · Added 9 Nov 2026
PreprintNot peer reviewedFrom SalesforceResearchers propose unified benchmark suite for enterprise LLM agent capabilitiesOriginal title: EGI-BENCH: An Enterprise General Intelligence Benchmark SuiteA preprint introduces EGI-Bench, combining eight enterprise agent benchmarks into one scored framework across six capability dimensions.Added 9 Nov 2026
PreprintNot peer reviewedResearchers report a benchmark for testing LLM agents on geospatial API tasksOriginal title: GeoNatureAgent (GNA): A Framework and Benchmark for Pre-Production Evaluation of Tool-Using Agents on Geospatial and Environmental TasksA preprint introduces a geospatial agent benchmark, finding that even the best large language model struggles with strict end-to-end tool use.Gabriel Diaz-Ireland et al. · Added 9 Nov 2026
PreprintNot peer reviewedPreprint proposes graph-based memory architecture for LLM long-term dialogueOriginal title: From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term DialogueA preprint introduces CogMem, a schema-based memory system that separates opinions from facts to improve long-term dialogue reasoning.Zirui Liao et al. · Added 9 Nov 2026
PreprintNot peer reviewedFrom SalesforceResearchers propose turning retrieved text into model memory instead of live lookupsOriginal title: From Lookups to Lessons: Turning Retrieved Knowledge into Parametric MemoryA preprint describes ResMem, which trains a parametric memory from retrieved text to cut storage while preserving correct answers.Added 9 Nov 2026
PreprintNot peer reviewedFrom GoogleResearchers trace how state-controlled media spreads through AI training dataOriginal title: The World as Viewed by 136 Countries: An Open, Curated, Annotated, and Multilingual Dataset of 118 Million State Media ArticlesA new dataset tracks state media's rising share of web data used to train and ground AI language models.Added 9 Nov 2026
