xxang/AStar-Thought-V2-Qwen3.6-27B

AStar-Thought-Latent-Qwen3.5-0.8B-deepseek-v3.2-speciale-openr1-math-3k-filtered-v10-angle90-Qwen3.6-27B-v2.12-strc0.1

This model is a fine-tuned version of Qwen/Qwen3.6-27B for the A\-Thought-V2 framework, as described in the paper [A\-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM](https://huggingface.co/papers/2609.07821).

A\*-Thought-V2 is an explicit-implicit interleaved efficient reasoning architecture guided by LLM dynamics. By interweaving implicit latent-space reasoning with explicit text, it performs lossless compression of the chain-of-thought: aligned reasoning steps remain explicit text, while deviating steps are compressed into continuous latent tokens.

Code for data compression, training, and evaluation is available at the AStar-Thought GitHub repository.

Training and evaluation data

Trained using the A\*-Thought-V2 compression pipeline on the OpenR1-Math-3k dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 1e-05
• —trainbatchsize: 1
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 8
• —gradientaccumulationsteps: 8
• —totaltrainbatch_size: 64
• —totalevalbatch_size: 64
• —optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• —lrschedulertype: cosine
• —lrschedulerwarmup_steps: 0.1
• —num_epochs: 3.0

Training results

• —Final training loss: 0.3388

Framework versions

• —Transformers 5.2.0
• —Pytorch 2.10.0+cu128
• —Datasets 4.0.0
• —Tokenizers 0.22.2