DolbyUUU/Logic-RL-Lite
Lightweight replication study of DeepSeek-R1-Zero. Interesting findings include "No Aha Moment", "Longer CoT ≠ Accuracy", and "Language Mixing in Instruct Models".
This project investigates how to post-train large language models (LLMs) to improve their reasoning capabilities using only reinforcement learning, without initial supervised fine-tuning. It takes base LLMs and a dataset of logic puzzles to analyze how different models learn to reason and identify key factors like model size, base model choice, and training algorithms. This is for AI researchers or machine learning engineers studying LLM reasoning and training methodologies.
No commits in the last 6 months.
Use this if you are researching how pure reinforcement learning impacts the reasoning abilities of language models, especially regarding logical puzzles.
Not ideal if you are looking for a ready-to-use LLM for general-purpose reasoning tasks or a tool for applying LLMs to business problems.
Stars
50
Forks
—
Language
Python
License
—
Category
Last pushed
Apr 01, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/llm-tools/DolbyUUU/Logic-RL-Lite"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
open-thought/reasoning-gym
[NeurIPS 2025 Spotlight] Reasoning Environments for Reinforcement Learning with Verifiable Rewards
Hmbown/Hegelion
Dialectical reasoning architecture for LLMs (Thesis → Antithesis → Synthesis)
LLM360/Reasoning360
A repo for open research on building large reasoning models
TsinghuaC3I/Awesome-RL-for-LRMs
A Survey of Reinforcement Learning for Large Reasoning Models
bowang-lab/BioReason
BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model | NeurIPS '25