6arXiv cs.CL (Computation and Language)·
Mental World Modeling (MWM) framework proposes coupling physical and mental state in world models
Researchers introduce Mental World Modeling (MWM), a theoretical framework that extends world models beyond physical scene simulation to explicitly track hidden mental states — beliefs, desires, intentions, emotions, and social norms — as core components rather than post-hoc rationales. They instantiate the framework in MENTIS, a training-free baseline that decomposes mental state tracking into structured sub-processes. Experiments across 8 LLM-based world models on a multimodal benchmark of situated decision scenarios show that explicit mental state modeling is essential for predicting human decisions. The work identifies current bottlenecks in mental world modeling and positions MWM as a next stage beyond physical world simulation.
Related guides (3)
Related events (8)
MMBench2 paper: hallucination in world models is predictable and preventable via coverage signals
Researchers introduce MMBench2, a 427-hour, 210-task dataset for visual world modeling, and train a 350M-parameter world model to study hallucination in generative world models. The paper identifies three distinct hallucination modes (perceptual, action-marginalized, scene-diverging) and develops lightweight signals that predict where models will fail. A coverage-aware sampling technique and curiosity-reward-based data collection enable efficient finetuning to unseen environments with as few as 50 real trajectories. The central finding is that world model hallucination is fundamentally a data coverage problem, with the same signals serving both detection and mitigation.
WorldEvolver: Self-Evolving World Models for LLM Agent Planning via Test-Time Memory Revision
Researchers introduce WorldEvolver, a framework that equips LLM agents with self-improving world models that revise their context at deployment time without updating model parameters. The system combines episodic memory (retrieval-based simulation), semantic memory (heuristic rule extraction from prediction errors), and selective foresight (confidence-based filtering). Evaluated on ALFWorld and ScienceWorld benchmarks, WorldEvolver achieves state-of-the-art world model prediction accuracy and improved downstream agent success rates across three backbone models. The work addresses a key challenge in long-horizon agent planning: unreliable foresight that can degrade rather than improve decision-making.
Looped World Models introduce iterative latent depth as a new scaling axis for world simulation
A new arXiv preprint introduces Looped World Models (LoopWM), a parameter-shared transformer architecture that iteratively refines latent environment states to achieve up to 100x parameter efficiency over conventional world models. The approach uses adaptive computation to scale depth dynamically per prediction step, addressing the tension between long-horizon simulation fidelity and deployment cost. The authors position iterative latent depth as a new scaling axis orthogonal to model size and training data.
BadWAM: Adversarial attacks that decouple action from imagination in world-action models
Researchers introduce BadWAM, a framework for adversarial attacks on World-Action Models (WAMs) — embodied control architectures that jointly predict future world states and actions. The paper demonstrates a new attack class called World-Action Drift Attacks, which use small visual perturbations to break alignment between a model's imagined future and its executed actions, reducing task success from 96.5% to 43.1% in closed-loop evaluation. A stealth variant (imagination-preserving attack) further shows that a WAM can appear to predict a plausible future while executing a desynchronized, harmful action — undermining the safety assumption that world prediction provides a check on action generation.
Theoretical framework proposes native meta-architecture to replace application-layer LLM cognitive simulation
A preprint from arXiv proposes a theoretical framework for embedding cognitive architecture natively into LLMs rather than simulating it via prompt engineering and context management. The framework introduces three mechanisms: Structural Tension (an endogenous loss function from information-manifold conflict), an Offline Recurrent Loop (sandboxed self-processing without external input), and Inference-time Plasticity (topology reconfiguration without weight modification). The authors argue these mechanisms could produce heterogeneous model instances with distinct topological structures through path-dependent evolution, while remaining within governance rails. The paper is primarily theoretical, offering operational definitions, reconfiguration operators, and falsification criteria rather than empirical results.
MeetingToM benchmark evaluates multimodal LLMs on Theory-of-Mind reasoning in multi-party meetings
Researchers introduce MeetingToM, a benchmark for evaluating multimodal LLMs on Theory-of-Mind (ToM) reasoning in naturalistic multi-party meeting settings. The benchmark is hierarchically structured across subject-level mental state prediction, dyadic addressee understanding, and group-level consensus reasoning, with particular attention to phenomena like pseudo-consensus where apparent agreement masks private dissent. Systematic evaluation of representative MLLMs reveals persistent limitations in integrating non-verbal cues, inferring hidden attitudes, and distinguishing genuine from surface-level consensus. The work addresses gaps in existing multimodal ToM benchmarks that focus primarily on overt, externally verifiable signals.
Paper argues LLMs are a degenerate special case of world models, maps continuous spectrum from NTP to JEPA
A new arXiv preprint reframes the LLM-vs-world-model debate by arguing that LLMs are a degenerate special case of world models rather than a fundamentally different paradigm, with the state space being token sequences and the only action being token appending. The paper maps a continuous spectrum from next-token prediction through multi-token prediction, future-summary prediction, and next-latent prediction up to JEPA-style architectures. It identifies two open research challenges in moving along this spectrum: the data cliff from self-supervised text to action-labeled environments, and whether transformers generalize to continuous-state prediction or require a new architectural primitive. The work directly engages with Yann LeCun's 2022 argument that general intelligence requires abandoning autoregressive prediction.
Qwen releases AgentWorld-35B-A3B: a world-model and environment-simulation MoE for agents
Qwen has released Qwen-AgentWorld-35B-A3B on Hugging Face, a 35B-parameter MoE model (3B active) built on the Qwen3.5 MoE architecture. The model is tagged for world-model and environment-simulation use cases, suggesting it is designed to simulate environments for agent training or evaluation. It is paired with a dataset called AgentWorldBench, indicating an associated evaluation suite. Early engagement is minimal (0 downloads, 4 likes) but the model represents a notable direction in agent-environment modeling from a major open-weights lab.