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

Ψ-0.6. Adaptive Streams of Thought in a Unified Vision-Language World Model. By Khai Loong Aw and seven others, from Stanford University. Published on the twenty-third of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Humans reason about the physical world by flexibly combining different forms of inference as a problem unfolds. Reaching that kind of flexibility is a major open challenge in AI. Capabilities often stay separated across specialized models, one for each kind of reasoning, and that separation limits how they can be composed and jointly adapted to the task in front of them. A question about where things are, what they look like, and what would happen if one of them moved draws on several of these at once. When each lives in a different model, nothing decides how to combine them for the problem at hand.

Ψ-0.6 is a unified vision-language world model. It adaptively composes perception, language, geometry and simulation inside one shared reasoning process, which the authors call streams of thought. A stream interleaves images, text, and explicit representations of geometry and motion. The model can generate and interpret all of these through a shared token interface, so every form in the stream is something it can both produce and read back.

Two capabilities follow from that. The model can reason through geometric intermediates on the way to an answer. And it supports counterfactual simulation: generating and interpreting imagined scenes to decide, for example, whether moving an object would obstruct a pathway. On top of this, reinforcement learning from task feedback jointly optimizes the selection, the generation, and the use of intermediate representations. Training shapes not only what the model produces, but which representation it reaches for.

Zero-shot, Ψ-0.6 surpasses leading open-weight models on multi-view spatial reasoning. Reasoning through geometric intermediates brings further gains on top of that. The reinforcement learning stage improves visual question-answering accuracy, and those gains transfer to unseen benchmarks. The abstract states these results without figures. Taken together, they cover both halves of the design: the shared stream that lets the model think in geometry, and the training that teaches it when to.

Most systems decide in advance what medium a problem gets solved in. Ψ-0.6 makes that choice part of the reasoning itself, and then trains it. The authors frame this as a path toward adaptive physical reasoning: a model that does not only predict the physical world, but also chooses the representations through which to reason about it.

Ψ-0.6. Adaptive Streams of Thought in a Unified Vision-Language World Model. By Khai Loong Aw and seven others, from Stanford University. Published on the twenty-third of October. This is the long version: the whole paper rather than the abstract.

A vision-language model can describe a kitchen. A world model can predict what that kitchen looks like after the camera turns. A depth estimator can say how far away the stove is. Now ask a question that needs all three: if the box slides to the right, will it hide the cup? Those abilities usually live in separate models. And when a language model calls a frozen tool, the tool's predictions cannot be shaped by whether the final answer was right. The authors, from Stanford, ask how a system can learn which representations are useful for a physical problem. Their answer is one transformer of about eight billion parameters, called Psi zero point six.

Two ideas carry everything that follows. The first is one vocabulary. Images, depth maps, optical flow, relative camera pose, point tracks and text are all serialized as tokens in one sequence. Each visual patch is a pointer token, saying which modality and where and when, followed by content codes saying what. So a prompt can fill some patches and leave others open.

The second idea is that every task is just an ordering of tokens. Give an image and a camera motion, ask for an image, and that is novel view synthesis. Give two images, ask for the camera motion, and that is pose estimation. The same camera pose can be an estimate, a command, or an intermediate step. A stream of thought is a chain of these predictions, each output becoming context for the next.

Training runs in stages, about one point two million supervised steps in all. It begins as a video world model, predicting RGB frames only, then adds depth, flow and camera conditioning. Language comes later: first web text, then captions, then instruction data with question answering, with the earlier predictive tasks kept in the mixture. A standard vision-language model starts from text; this one starts from predicting how scenes change.

For questions about moving an object, two decoding tricks matter. First, a single forward pass estimates, for every patch, the probability that it moves, and flow is generated from the likeliest movers first. A typical question moves one small object, so this spends the sequential budget where the motion is. Second, a form of classifier-free guidance: at each token, the model compares its predictions with and without the text instruction and amplifies the difference. It then generates the future frame and reasons to an answer.

The adaptive part is learned. Given a video and a question, the model acts as a policy and emits a short notation naming which streams to compute: depth at this frame, camera motion and flow between those two. The reward is how much those streams lower the loss of the correct answer compared with answering from RGB alone, judged by a scorer built from the same base model. Training starts with filtered behaviour cloning on sampled notations that beat RGB only, then moves to a variant of GRPO. Sixteen notations are sampled per question, each rewarded against the group mean, so above-average choices are reinforced and the rest suppressed. The policy picks its streams from the video and question alone, without seeing the answer options.

On MindCube, a benchmark of multi-view spatial reasoning, the frozen model scores between forty-nine point seven and fifty-one point three percent, above every open generalist and spatial specialist it is compared with. Composing a stream, predicting relative camera poses and then reading the answer off the matching real view, lifts it to fifty-seven point seven. On predicting the flow of a counterfactual move, motion ordering plus guidance reaches fifty-six point four percent accuracy. A modular pipeline of expert models reaches fifty point eight, despite receiving camera intrinsics as an oracle input, and image editors given the true object mask top out at sixteen point four.

The comparison that tests the thesis most directly sits inside the reinforcement learning figure, which plots a random choice of streams next to the learned policy. Random streams already beat RGB only almost everywhere. On EgoDex, the learned policy reaches fifty-three point five percent against fifty point two for random. On OpenVid it sits slightly below random. Averaged, the mean relative gain over video-only answering is five point six percent on held-out data and nine percent across four unseen benchmarks. Taking the best of sixteen sampled notations reaches seventy-three point three on EgoDex, which the authors read as headroom.

The paper has a short limitations section with two items. The first is scale: the model was built with far less data, compute and engineering than frontier systems, and the authors expect scaling to help, though they cite only preliminary experiments. The second is transfer: they show gains from composing capabilities at inference, but do not isolate positive and negative transfer among those capabilities during training.

The evidence adds more. Before fine-tuning, the model scores twenty-four point five percent on the simulated counterfactual questions, below the thirty-five point eight chance level. On the real robot split, never seen in training, the tuned model reaches fifty-one point four, behind a Qwen model tuned on the same data at fifty-nine point six. After MindCube fine-tuning, the Rotation category stays near chance, which the authors attribute to low scene diversity in that training data.

The idea to carry away is that the medium of reasoning can itself be a choice. In text chain of thought, the medium is fixed. Here depth, flow, camera pose and imagined frames are tokens the same model writes and reads, so which one to think through becomes an action that task feedback can train. The paper shows this as a proof of concept on video questions, with modest learned gains on top of a larger effect from simply having geometry available.