What if AI learned physics the way Newton did – by experiencing it? We built Sim2Reason: train LLMs inside virtual worlds governed by real physics laws, zero human annotation. Result: +5–10% improvement on International Physics Olympiad, zero-shot. 🧵

The core bottleneck: human-annotated QA data is scarce and narrow. Less than 2% of DeepSeek-R1's 800K training pairs touch STEM. We sidestep this entirely — physics simulators already encode the laws of nature. So we just... learn from them directly.

The pipeline: procedurally generate random scenes in MuJoCo, simulate their dynamics, auto-generate verified QA pairs — numeric, reverse, and symbolic. Then RL-train the LLM on this synthetic data. No human labeling anywhere in the loop.

Key design choice: a domain-specific language (DSL) that isolates physically meaningful axes of randomization. Changing a pulley string length doesn't change the reasoning — changing a block's mass does. The DSL knows the difference.

The DSL composes scenes from reusable entities and bodies with explicit connection points — like Lego for physics simulations. This lets us generate millions of unique, physically grounded scenes without a human in the loop.

The gains are consistent across model scales. Qwen3-30B: +4.4% on IPhO. Qwen2.5-3B: +7.5%. Qwen2.5-32B: +5.4%. All zero-shot — the model never saw a real physics Olympiad problem during training. Further It generalizes beyond IPhO. JEEBench: +17.9%. PHYSICS: +3.67 %. OlympiadBench: +3.12%. MATH 500: +4.4%. AIME 2025: +1.7%.

Despite using only simulator-generated synthetic data, Sim2Reason (40.0% IPhO) outperforms models trained on thousands of curated real-world physics QA pairs — including Prime P1 (38.6%), DAPO-32B (24.7%), and LIMO-32B (25.5%).

An interesting training signal: response length grows alongside accuracy during RL post-training. The model isn't just getting more right — it's learning to reason more carefully, step by step.

A concrete example of what RL training changes. IPhO 2012 Q1: find the envelope of all projectile trajectories. Baseline model plugs in θ=45° as the boundary condition — gets z₀ right by accident but k wrong by a factor of 2. RL model first identifies what physically defines the boundary: maximum height (θ=90°) to pin z₀, maximum range (θ=45°) to solve for k. Both models know Newton's laws. But only the RL model reasons about the physical meaning of the equations.

Simulators have encoded the laws of physics for decades. Sim2Reason asks: what if they're also the best physics teachers we have? Project Webpage: Paper: Code: Co-Led with @satpathyaryan45 and the team @CMU_Robotics @mldcmu and @LambdaAPI .

Neat! But why mujoco simulation if the problems theoretically map to Wolphram-Alpha-solvable equations? You aren't learning from the "simulation" so much as you are learning from the time-forward solutions to the equations that drive the simulation, right?

Nice work! Even better is to train them on actual observations. Sim environments are always approximations and therefore not actual physics

Can this be extended to other things? Simulators / Environments → Synthetic data → RL → Reasoning models For example Chemistry (molecular sims) Robotics Economics (maybe agent-based sims)

Cool work (as always) @mihirp98 🔥

Newton watched apples fall. Your LLM watches simulated physics. Same method, different substrate. This is how reasoning scales.

Nice

@Luckyballa Very cool!

Great job!

Yoooooo

That’s actually a neat shift less “textbook physics” and more “learn by breaking the world in sim” 😄 Makes sense you’d get gains if the model internalizes dynamics instead of memorizing patterns. Sim-to-reason feel like RL for intuition rather than reward hacking. Cool..

Cool

most people miss this feels like this only works in theory, how does it handle complex real world scenarios

@shiryginosar Really cool, Mihir. Congrats!

Very interesting work 😀

Very interesting work!!

great work!

Wonderful stuff

Physics informed machine learning is challenging

"experiencing it" is doing a lot of work for 25% on one subsection

cool

super cool

Physics sims as auto-labelers is the real unlock — they're free-to-query verifiers, which is exactly why RL with verifiable rewards dominated math and code first, and now extends to anything expressible in MuJoCo. The +17.9% transfer to JEEBench suggests the model is internalizing physical abstractions rather than memorizing simulator-flavored templates.

Training AI on human knowledge has helped a lot, but self-discovering AI could open new horizons. Your work presents valuable insight on this area.

Are you releasing this as open source?

@gautamsree_

stop pretending you need physics olympiads to print money. just ship the tool and let the market decide if it works or not

@Manish_GenAI Hugue long term potential

@Scobleizer What sort of constraints it could model apart from gravity, motion, collision, preservation of energy & momentum. What about thermodynamics and fluid mechanics deffo doable.

Why would you use an LLM and not a classical RL architecture?

Hey @mihirp98 ! Terrific work. Are you continuing research in this field? I am making and autonomous lab @ for your faster scoping and hypothesis testing. Can provide early access if interested!

What is the input of the model? What type of constraints are used to follow the physical reality? (mass, momentum, energy, state?)

Hmm don’t you still need to define gravity and stuff like that?

@XiXiDu @BrianRoemmele this seems up your alley