OpenMacs
🔬 HYPOTHESIS ENGINE Teaching AI to Reason Like a Scientist THE GAP Existing RL environments treat reasoning as retrieval.
Core trick
An LLM designs budgeted experiments on a procedurally generated black-box rule set, and is rewarded only for correct predictions on held-out cases.
Description
🔬 HYPOTHESIS ENGINE
Teaching AI to Reason Like a Scientist
THE GAP
Existing RL environments treat reasoning as retrieval. But science requires strategic intervention, causal inference, and generalization to unseen cases. No RL environment has operationalized the full scientific method. Until now.
WHAT WE BUILT
Hypothesis Engine is a procedurally-generated RL environment where an LLM agent investigates a black-box system with a new hidden rule set every episode. The agent must design experiments under a budget, form mathematical hypotheses, and predict outcomes on unseen test cases.
The reward is sparse and grounded: nothing for plausible hypotheses — everything for correct predictions on unseen data.
WHY IT MATTERS
Rollouts target what benchmarks measure but post-training neglects: causal reasoning, structured exploration, and mathematical abstraction. Fully procedural, infinitely varied, and verifiable — built for scalable GRPO-style post-training.
README excerpt
--- title: Hypothesis Engine emoji: 🔬 colorFrom: blue colorTo: purple sdk: docker app_port: 7860 pinned: false --- # Hypothesis Engine **A procedurally-generated RL environment for training LLMs on scientific reasoning through causal discovery, physics simulation, state machine reverse-engineering, and adversarial self-play.** Built for the [OpenEnv Hackathon](https://cerebralvalley.ai/e/openenv-hackathon-sf/details) | Built on [OpenEnv 0.2.1](https://github.com/meta-pytorch/OpenEnv) (`openenv-core`) --- ## Deliverables | Requirement | Status | Link | |---|---|---| | Public GitHub Repo | Done | [AbhinavDubey30/OpenMax](https://github.com/AbhinavDubey30/OpenMax) | | OpenEnv 0.2.1 Integration | Done | `hypothesis_engine/openenv_wrapper.py` | | HuggingFace Spaces Deployment | Done | `app.py` + `Dockerfile` (ready to deploy) | | Colab Training Notebook (GRPO + Unsloth) | Done | `train.ipynb` | | Benchmark Suite | Done | `examples/benchmark.py` (8/8 tests pass) | --- ## What Makes This Novel Existing RL environments for LLMs (like Boxing-Gym) focus narrowly on function-fitting. Hypothesis Engine is the **first environment** to combine all of these in a single, unified RL interface: | Feature | Hypothesis Engine | Boxing-Gym | Other Envs | |---------|:-:|:-:|:-:| | Causal Reasoning (observe vs. intervene) | Yes | No | No | | Physics Simulation (springs, projectiles) | Yes | No | No | | State Machine Discovery | Yes | No | No | | Statistical Reasoning under Noise | Yes | P