Physical Reasoning is an open source research collective studying world models: systems that learn the dynamics of the physical world and predict what happens next. We work across robotics, self-driving, prediction markets, and the sciences.
Read our blog →Most world-model research is conducted within a small number of industrial labs and evaluated on proprietary data. Evaluation protocols are not standardized, and whether a model has learned the underlying dynamics or only the surface statistics of its training distribution is rarely measured directly.
The physics of the world should be open. We conduct our research in public, and every result is published, including negative ones, together with the code to reproduce it.
Anyone can contribute. Start here for what we work on and what to read, or pick one of the open problems.
We test whether a model adds information to prediction-market prices by scoring its forecasts out of sample against the market's own price at the same moment, with proper scoring rules and pre-registered sealed tests. On Kalshi daily temperature markets, free weather forecasts beat the market by a wide margin in 2023 and by very little since 2025 (isotherm). The model is live as an API and an MCP server that agents can call (reference).
We test what a self-supervised encoder has learned by training probes on its frozen representations, against raw-feature and untrained-encoder controls. On Kalshi strike ladders, JEPA representations never beat the raw inputs, and standard collapse diagnostics pointed the wrong way (causal-jepa, write-up).
We freeze and hash every model before a round opens and score it only on outcomes that resolve afterwards, against copying the last value and against the market on the same outcome. In dry runs, ECMWF's learned weather model beats its physics model once both are calibrated, and no released time-series foundation model beats copying the last price (results, forward). The first sealed round freezes on 31 October 2026.
We argue that a world model should be judged on whether its imagined outcome of an intervention matches what actually happens, not on prediction error alone. This is a proposed evaluation. The first benchmark replays order books we record to score what a limit order we never placed would have done (forward task 3); it is a pipeline check so far. A test on physical environments is an open problem.