RTM-2 (ReThinking Mind 2)is an experimental reasoning project built on top of HRM (Hierarchical Reasoning Model), with a local research branch focused on making the model more adaptive, more context-aware, and a little closer to what we informally kept calling a "sixth-sense" style of reasoning.
The goal is not just to make a model solve structured tasks, but to push it toward learning how it should approach different tasks:
- when to lean on high-level planning,
- when to rely on fast low-level computation,
- and how to infer the right reasoning style from the task context itself.
In practice, this repo currently explores that idea through a variant called HRM-Free-Meta, which adds:
- latent task-context modeling,
- dynamic weighting between hierarchical modules,
- meta-learning inspired regularization,
- and a cleaned local training / evaluation / resume workflow.
This repository is an experimental research workspace, not a polished library.
What is already in decent shape:
- a coherent local training entrypoint,
- checkpoint save / resume support,
- checkpoint evaluation,
- basic smoke tests,
- cleaner documentation than before.
What is still true:
- some files remain exploratory or debug-oriented,
- full runtime verification depends on a local PyTorch + CUDA environment,
- this is still a research repo, not a finished product.
The active code lives under HRM.
The most important entrypoints are:
- HRM/unified_training.py — main training and resume pipeline
- HRM/evaluate.py — evaluate saved checkpoints
- HRM/test.py — quick smoke-test runner
- HRM/README.md — detailed project usage
The bigger research idea behind RTM-2 is simple:
Not just solving tasks — learning how to approach them.
That means this repo is less about scaling parameter count blindly, and more about experimenting with:
- hierarchical reasoning,
- adaptive computation,
- task-aware control,
- context-sensitive routing,
- and meta-level signals that influence reasoning behavior.
See the main project guide here:
- HRM/README.md
If you mainly want resume examples:
- HRM/resume_quickstart.md
Experimental HRM fork exploring more intuitive task-aware reasoning with meta-learning, latent context modeling, dynamic module weighting, and a cleaned training/evaluation workflow.
pytorchdeep-learningmachine-learningmeta-learningreasoninghierarchical-reasoningadaptive-reasoningcontext-awareresearchexperimentalarc-agi