MYTH-Lab/GoT-R1-14B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

GoT-R1-14B is a 14 billion parameter causal language model developed by MYTH-Lab (Wuhan University & Shanghai Jiao Tong University) with a 32768 token context length. It utilizes a novel Graph-of-Thought (GoT) reasoning framework, moving beyond linear Chain-of-Thought to internalize structured reasoning. This model is optimized for high-density logical problem-solving with minimal verbosity and high token efficiency, making it suitable for complex reasoning tasks.

Loading preview...

GoT-R1-14B: High-Density Reasoning with Graph-of-Thought

GoT-R1-14B, developed by MYTH-Lab (Wuhan University & Shanghai Jiao Tong University), is a 14 billion parameter causal language model that redefines complex problem-solving by internalizing a Graph-of-Thought (GoT) reasoning framework. Unlike traditional Chain-of-Thought (CoT) models that can generate verbose and redundant outputs, GoT-R1 constructs a high-density, structured reasoning graph internally, ensuring each reasoning node is an atomic logical primitive.

Key Capabilities & Features

  • Internalized Graph-of-Thought: Shifts from linear CoT to a structured, internal GoT for enhanced logical consistency.
  • High-Density Reasoning: Decouples pure logic from conversational filler, outputting graph topological steps.
  • Extreme Token Efficiency: Achieves state-of-the-art accuracy with significantly reduced token usage (e.g., 1.8% of the token budget compared to external search methods like Tree-of-Thought).
  • Reduced Verbosity: Strictly penalizes redundant narration during RL training, avoiding unnecessary conversational filler.
  • Automated Structural Synthesis: Trained on high-fidelity logical skeletons derived from teacher-model CoT traces without manual graph labeling.
  • Improved Accuracy: Demonstrates superior performance across reasoning benchmarks like GSM8K, IFEval, TruthfulQA, and Winogrande, with an 18% improvement on TruthfulQA at the 8B scale.

Training Methodology

The model undergoes a two-stage training regimen: Supervised Fine-Tuning (SFT) to master GoT syntax and structural formatting, followed by Group Relative Policy Optimization (GRPO) to reinforce topological integrity and penalize verbosity.

Ideal Use Cases

GoT-R1-14B is particularly well-suited for applications requiring precise, efficient, and logically consistent reasoning, especially in scenarios where minimizing token usage and avoiding verbose outputs are critical. This includes complex problem-solving, logical inference, and tasks demanding high accuracy in structured reasoning.