Tag
Kyutai Labs introduces REKI, a 2.4B vision-language-action model pretrained exclusively on public data that sets a new Pareto frontier between performance and model size, using a simple design with register-token memory in the action expert. Full code, data curation pipelines, and all pretrained/finetuned weights are released.
该论文系统研究了VLM驱动的GUI智能体中的记忆模态选择问题，发现最优记忆表示是随交互步骤变化的，并提出SwitchMem路由框架在每一步动态选择文本或视觉记忆，在三个GUI基准上超越了使用静态记忆表示的基线方法。
DP-MIND 提出一种记忆引导的框架，将差分隐私推理扩展到百万 token 长上下文场景，通过在扰动输入上用强化学习优化记忆构建策略并结合本地去噪模块，在长上下文基准上优于现有 DP 与长上下文推理基线。
SkillForge proposes an agentic reinforcement learning method that co-evolves an LLM agent and its skill library through a fitness-driven lifecycle of trial, active, stable, and retired states. It reports up to 7.8% relative improvement over strong baselines on interactive agent benchmarks and releases SkillFurnace, a 5k+ record dataset for studying skill quality and lifecycle management.
RecurTrace introduces loop-time memory and adaptive halting to improve latent reasoning in language models, achieving higher accuracy on MathQA with optimized compute compared to fixed-loop methods.
This paper introduces CreaEval, an automated creativity evaluator for complex multi-step tasks that decouples analysis and judging to reduce biases and improve evaluation reliability, demonstrating a 22.74% average performance improvement over baselines.
PriorEye introduces geospatial visual priors for end-to-end autonomous driving, enhancing anticipatory behavior and robustness through a dual-memory architecture, as presented at ECCV 2026.
This paper introduces memory-augmented speculative execution for LLM agents, using three online memory systems to improve prediction accuracy by 19-39% on action prediction and up to 2.5x on observation prediction, all while being lossless with zero added wall-clock cost.
TF-Engram introduces a train-free system that stores phrase-level semantic memory on a GPU-DRAM-SSD hierarchy for LLMs, using predictive prefetching to hide latency, and demonstrates improved downstream performance on Qwen3-0.6B.
This paper introduces MA-DLE, a memory-based feature augmentation method for speech-based automatic depression level estimation, achieving state-of-the-art performance on the DAIC-WOZ and E-DAIC datasets.
Introduces MPT benchmark and PRefine method for cross-session personalized tool calling that captures user choice reasoning with minimal token overhead.