← Back to all news
arXiv AI

Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs

The paper investigates how few-shot demonstrations help large language models (LLMs) build and use in-context world representations. By providing demonstrations from different worlds, the authors show that LLMs can better encode graph structures in low-dimensional hidden states, leading to improved predictions across six models from four families. The study also demonstrates that these representations shift to nearly orthogonal subspaces and that targeted interventions on them more significantly impair performance, indicating their critical role in downstream tasks such as ARC-AGI, web agent tasks, and Othello.

Related stories

arXiv AI
Sep 17

CSWAM: Better Causal Semantic Representations for Out-of-Distribution Generalization in World Action Models

The paper introduces CSWAM, a Causal Semantic World Action Model that enhances FastWAM by integrating a causal semantic expert based on V-JEPA 2.1. This expert provides temporally grounded, appearance‑agnostic representations of semantic state changes and motion, leveraging sparse observation history and causal attention to improve action‑only inference. Experiments on simulation and real‑robot tasks show that CSWAM significantly boosts out‑of‑distribution generalization, raising success rates from 10.16% to 45.18% on RoboTwin 2.0 and from 27.5% to 70.0% across real‑robot tasks.

By Tianbin Liu, Jian Zhu, Taiyi Su, Jianjun Zhang, Chong Ma, Zitai Huang, Yi Xu
robotics
More like this →
arXiv AI
Oct 16

Learning Intrinsic Dynamics in Latent World Models

The paper investigates how encouraging action predictability in latent world models can preserve intrinsic dynamics, which are crucial for planning. By deriving a theoretical lower bound and conducting empirical tests, the authors show that action predictability prevents the encoder from merging distinct action outcomes, even when visual appearance varies. Using a method called Concat IDM, they achieve a substantial increase in planning success on the Push‑T benchmark, from 9.1% to 77.2%, demonstrating the practical benefit of this inductive bias.

By Jiamin Shi, Yuanhong Zhang, Zhiyang Liu, Jiawei Sun, Jingwen Fu
safety
More like this →
arXiv AI
3d ago

Vision-Language Agents Can Learn in Their Own Latent Worlds

The paper introduces Dwell, a model‑based reinforcement learning framework that treats a pretrained vision‑language model (VLM) as a recurrent latent transition operator. Dwell enables the VLM to predict, adjust, and act within its own latent space, allowing the agent to learn from real interactions and latent imagination without reconstructing full futures. Experiments across five interactive visual environments show that Dwell outperforms other trainable methods while significantly reducing computational demands for imagination.

By Haoxi Li, Sikai Bai, Jianfei Ma, Qinglin Hou, Xiaocheng Lu, Tao Han, Jie Zhang, Song Guo
llmsagentsdiffusionreinforcement-learningmultimodal
More like this →
arXiv AI
Sep 10

WorldAgen: Unified State-Action Prediction with Test-Time World Model Training

WorldAgen is a unified framework that jointly learns world modeling and action prediction using a shared Transformer backbone with two specialized heads. It introduces a Mixed Unidirectional Attention Mask to separate the world model and agent model, and enables Test-Time Training (TTT) by sampling exploratory actions and updating the world model with real state transitions. Experiments on CALVIN and LIBERO show that WorldAgen matches or surpasses state‑of‑the‑art methods, especially when TTT is applied to a few samples.

By Chi Wan, Kangrui Wang, Yuan Si, Pingyue Zhang, Manling Li
llmsagentsmultimodalbenchmarks
More like this →
arXiv AI
Sep 21

FOCAL-VLA: Subtask-Guided Geometry Distillation and Implicit World Modeling for Vision-Language-Action Models

FOCAL‑VLA is a framework that improves vision‑language‑action models by combining subtask‑guided geometry distillation with implicit world modeling. It transfers geometric knowledge from VGGT to focus on subtask‑relevant image regions and uses Track4World features to capture future 3D evolution, guiding action generation without running these models at inference time. Experiments demonstrate that FOCAL‑VLA outperforms baselines on both simulation benchmarks and real‑world manipulation tasks.

By Zhiyuan Gao, Di Wen, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Sch\"afer, Kunyu Peng, Michael Beetz
llmsroboticsefficiencymultimodalbenchmarks
More like this →