arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.06130cs.LGcs.AIcs.CL
收藏

具有潜在动作的自我改进世界建模

Self-Improving World Modelling with Latent Actions

University of Edinburgh(爱丁堡大学);Nvidia Research(Nvidia研究);University of Groningen(格罗宁根大学);University of Cambridge(剑桥大学)

Yifu Qiu, Zheng Zhao, Waylon Li, Yftah Ziser, Anna Korhonen, Shay B. Cohen, Edoardo M. Ponti

AI总结:

SWIRL通过将动作视为潜在变量,结合前向世界建模和逆动态建模,实现了对LLM和VLM的自我改进世界建模,在多个基准测试中取得了显著提升。

摘要:

世界内部建模——在动作Z下预测先前状态X和后续状态Y之间的转换——对于LLM和VLM的推理和规划至关重要。学习此类模型通常需要成本高昂的动作标记轨迹。我们提出了SWIRL,一种自我改进框架,通过将动作视为潜在变量并交替使用前向世界建模(FWM)P_θ(Y|X,Z)和逆动态建模(IDM)Q_ϕ(Z|X,Y)来学习仅基于状态的序列。SWIRL迭代两个阶段:(1)变分信息最大化,通过更新FWM生成最大化条件互信息与潜在动作的后续状态,促进可识别的一致性;(2)ELBO最大化,通过更新IDM来解释观察到的转换,有效地执行坐标上升。两种模型均使用强化学习(特别是GRPO)进行训练,以相反冻结模型的对数概率作为奖励信号。我们为两种更新提供了理论可学习性保证,并在多个环境中评估了SWIRL:单轮和多轮开放世界视觉动态以及物理、网络和工具调用的合成文本环境。SWIRL在AURORABench上实现了16%的提升,在ByteMorph上实现了28%的提升,在WorldPredictionBench上实现了16%的提升,在StableToolBench上实现了14%的提升。

英文摘要:

Internal modelling of the world -- predicting transitions between previous states $X$ and next states $Y$ under actions $Z$ -- is essential to reasoning and planning for LLMs and VLMs. Learning such models typically requires costly action-labelled trajectories. We propose SWIRL, a self-improvement framework that learns from state-only sequences by treating actions as a latent variable and alternating between Forward World Modelling (FWM) $P_θ(Y|X,Z)$ and an Inverse Dynamics Modelling (IDM) $Q_ϕ(Z|X,Y)$. SWIRL iterates two phases: (1) Variational Information Maximisation, which updates the FWM to generate next states that maximise conditional mutual information with latent actions given prior states, encouraging identifiable consistency; and (2) ELBO Maximisation, which updates the IDM to explain observed transitions, effectively performing coordinate ascent. Both models are trained with reinforcement learning (specifically, GRPO) with the opposite frozen model's log-probability as a reward signal. We provide theoretical learnability guarantees for both updates, and evaluate SWIRL on LLMs and VLMs across multiple environments: single-turn and multi-turn open-world visual dynamics and synthetic textual environments for physics, web, and tool calling. SWIRL achieves gains of 16% on AURORABench, 28% on ByteMorph, 16% on WorldPredictionBench, and 14% on StableToolBench.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑