Ψ-0.6: Adaptive Thought Flow in a Unified Vision-Language World Model

Chain-of-Thought has elevated the reasoning capabilities of large models to a new level, but it has a fundamental limitation: it can only think in language. When facing the physical world, linguistic descriptions often fall short—you can clearly say "the cup is at the edge of the table," but it's hard to precisely express its depth, orientation, and the direction in which it's about to fall using words alone.

A research team from NeuroAI Lab has released Ψ-0.6, attempting to answer a more fundamental question: what should the "chain of thought" for the physical world look like? Their answer is—let the model think simultaneously in multiple representations, and learn on its own when to use which one.

One Model, Five "Thinking Languages"

The core design of Ψ-0.6 extends reasoning from a single language modality to a combination of multiple representational forms. During the reasoning process, the model can invoke text, depth maps, motion information, camera poses, and imagined/generated scene images. This is fundamentally different from traditional VLM "see and tell": the latter treats vision as input and language as output, whereas Ψ-0.6 treats vision-related representations themselves as intermediate steps in reasoning.

In other words, when asked a question involving complex spatial relationships, the model may no longer simply "read" the image content and then organize language to answer; instead, it internally constructs a scene representation with depth and viewpoint information, and uses that representation to derive the answer. This ability to "manipulate objects in imagination" is something the language modality can hardly replace.

Vision-Language Models (VLMs) are key background for understanding the design intent of this paper. A VLM typically consists of a vision encoder (e.g., ViT) and a language model. The vision encoder compresses the image into a set of feature vectors, which are then mapped through a projection layer into the input space of the language model; the model subsequently generates text autoregressively. The limitation of this architecture is that visual information is used only once at the input stage, and the reasoning process occurs entirely in language space. Ψ-0.6 takes a different approach: it repeatedly invokes structured representations such as depth maps and camera poses as intermediate products in the reasoning chain, effectively letting visual representations participate in "thinking" rather than serving merely as "the question." This shift requires the model to have the ability to generate and parse outputs in multiple modalities, and it also imposes new requirements on how training data is organized.

Camera pose refers to the position and orientation of a camera in 3D space, usually represented by a translation vector and a rotation matrix (or quaternion), totaling six degrees of freedom. It is the key quantity connecting 2D images and the 3D world: the same object looks vastly different under different poses, and only by knowing the pose at capture time can the model back-project pixels into a unified 3D coordinate system, thereby judging the true distances and relative orientations between objects. In robotic manipulation scenarios, pose estimation directly determines whether the grasp point is accurate. Incorporating camera pose into the reasoning chain means the model can internally reconstruct the correspondence of "from which angle, what is seen," thereby handling spatial ambiguity caused by viewpoint changes.

Adaptive Selection: RL Teaches the Model When to Use Which Kind of Thinking

The real difficulty of multimodal reasoning is not "whether it can be used" but "whether it should be used." For some problems, language reasoning is fast and accurate; for others, spatial representations are necessary. Ψ-0.6 acquires this scheduling policy through reinforcement learning—during training, the model learns to judge which reasoning path is suitable for the current task.

This is one of the most valuable parts of the work. Most multimodal models adopt a fixed fusion architecture with hard-coded reasoning paths; Ψ-0.6 turns "choosing a representation" itself into a learnable decision, giving the model the ability to trade off between efficiency and accuracy.

Training large models with reinforcement learning (RL) differs from supervised fine-tuning in a way worth explaining. Supervised fine-tuning makes the model imitate given answers, whereas reinforcement learning provides only a reward signal, and the model explores on its own which outputs can obtain higher returns. In the Ψ-0.6 setting, the reward comes from the correctness of the final answer, so the model needs to discover strategic knowledge such as "depth representations are more effective for this type of problem." Commonly used algorithms include PPO, GRPO, etc., which update the policy by sampling multiple reasoning paths and comparing their returns. The advantage of this paradigm is that it can acquire scheduling strategies that are hard to express directly with labeled data; the cost is training instability, sensitivity to reward design, and a tendency to overfit specific benchmarks.

Empirical Performance: Spatial Reasoning and World Prediction

According to the evaluation results released by the authors, Ψ-0.6 surpasses leading open-weight models in a zero-shot manner on MindCube, a complex spatial reasoning benchmark. On FlowBench, it outperforms other world models in predicting object motion. These two tests correspond respectively to "understanding static spatial relationships" and "inferring dynamic physical processes," precisely covering the two key dimensions of physical-world reasoning.

More noteworthy is transferability. The model is trained on simulated room data yet can transfer to real robot tabletop manipulation scenarios. This suggests that what the model learns is not specific textures or statistical patterns of the simulation environment, but some more general spatial and physical representation.

World models are a core concept in the embodied intelligence field in recent years, referring to models that can predict future states of the environment—that is, given the current observation and action, infer the observation or state change at the next moment. Their difference from traditional video prediction lies in emphasizing modeling of physical laws rather than pixel-level fitting. The value of world models is that they allow agents to trial and error in "imagination," thereby reducing the cost of interaction in real environments; this is also the foundation of model-based RL. The "predicting object motion" mentioned in this article is precisely the basic capability of a world model, and whether it can generalize to unseen scenarios is the key criterion for judging whether it has learned real physical laws rather than surface-level statistics of the data.

One Model, Five Roles

Ψ-0.6's positioning is quite aggressive: the same model simultaneously serves as a perceiver, a vision-language model, a world model, a reasoner, and a policy network. This "full-stack" design has clear engineering advantages—it avoids interface loss and information loss when multiple modules are connected in series, and it also makes end-to-end training possible.

Of course, this also raises questions: if a single model takes on so many responsibilities, will it fail to be sufficiently specialized in each role? The evidence given by the authors is that it is competitive on multiple benchmarks, but the mutual constraints among different capabilities still require finer-grained ablation experiments for verification.

A New Paradigm for Physical-World Reasoning?

The concept of "adaptive thought flow" proposed by Ψ-0.6 points to a possible direction: future embodied intelligence models will no longer treat language as the sole medium of reasoning, but will dynamically switch among multiple representations according to the nature of the task. This is especially important for scenarios requiring precise spatial understanding, such as robotic manipulation, autonomous driving, and AR interaction.

However, this should also be viewed calmly. Currently available public information consists mainly of evaluations self-reported by the authors, and third-party reproduction and broader benchmark comparisons remain to be supplemented. The interpretability of multimodal reasoning paths is also an open question—when the model chooses to reason with depth representations rather than language, it becomes harder for humans to trace the basis of its decisions.

The paper, code, and project homepage have all been made public. Whether this approach can be validated and promoted by the community is worth continued observation.

One-line Summary

Ψ-0.6 enables models to reason with multiple representations such as text, depth, and pose, and to adaptively select among them using reinforcement learning.

Paragraph Summary

Ψ-0.6 from NeuroAI Lab attempts to break through the limitation that chain-of-thought can only think in language, extending the reasoning medium to five representations: text, depth maps, motion information, camera poses, and imagined/generated scene images. The model acquires a scheduling policy through reinforcement learning and can judge, based on the nature of the task, whether to use language reasoning or spatial-representation reasoning, rather than relying on a hard-coded fusion architecture. Evaluations show that it surpasses leading open-source models on MindCube zero-shot spatial reasoning, outperforms other world models on FlowBench object motion prediction, and can transfer from simulated rooms to real robot tabletop manipulation. The model simultaneously takes on five roles—perceiver, VLM, world model, reasoner, and policy network—and the authors call it "adaptive thought flow," offering a new reasoning paradigm beyond language for embodied intelligence.