CODI (Continuous Chain-of-Thought via Self-Distillation) is a latent reasoning model based on GPT-2 that extends the base architecture with an optional projector module for enhanced hidden state representations. This model is part of the Parallel Test-Time Scaling for Latent Reasoning Models framework.

Model Details

Base Architecture: GPT-2 Language Model
Model Class: CODIGPT2 (extends GPT2LMHeadModel)
Special Features: Optional projector module for extended hidden states
Latent Tokens: Uses special tokens <|latent|>, <|start-latent|>, <|end-latent|> for latent reasoning
Input Format: Direct input without newline before <|start-latent|> token

Related Models

This repository includes other latent reasoning models that you might find useful:

ModalityDance/latent-tts

Installation

Download the model from HuggingFace:

huggingface-cli download dd101bb/latent-tts-codi --local-dir checkpoints/codi

Quick Start

Basic Usage

python

from transformers import AutoTokenizer
from src.generation_mixin import LatentGenerationMixin, LatentGenerationConfig
from src.paths import MODELS

# Load tokenizer
model_id = "checkpoints/codi"
tokenizer = AutoTokenizer.from_pretrained(model_id)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Get latent token IDs
latent_id = tokenizer.convert_tokens_to_ids("<|latent|>")
start_id = tokenizer.convert_tokens_to_ids("<|start-latent|>")
end_id = tokenizer.convert_tokens_to_ids("<|end-latent|>")

# Create model class with generation mixin
class LatentCODI(MODELS["codi"]["class"], LatentGenerationMixin):
    def __init__(self, config):
        super().__init__(config)

# Load model
model = LatentCODI.from_pretrained(
    model_id,
    latent_id=latent_id,
    latent_start_id=start_id,
    latent_end_id=end_id,
    device_map="auto",
)

# Prepare input (note: no newline before <|start-latent|>)
question = "What is 2 + 2?<|start-latent|>"
inputs = tokenizer(question, return_tensors="pt").to(model.device)

# Configure generation
generation_config = LatentGenerationConfig(
    max_new_tokens=512,
    latent_length=6,
    latent_do_sample=True,
    latent_do_sample_by="dropout",  # or "noise"
    dropout_p=0.1,
    pad_token_id=tokenizer.pad_token_id,
    eos_token_id=tokenizer.eos_token_id,
)

# Generate
output = model.generate(
    **inputs,
    generation_config=generation_config,
    num_return_sequences=1,
)

# Decode result
result = tokenizer.decode(output[0], skip_special_tokens=True)
print(result)

Batch Processing

The model fully supports batch processing with Transformers:

python

# Prepare batch inputs
questions = [
    "What is 2 + 2?<|start-latent|>",
    "What is 5 * 3?<|start-latent|>",
    "What is 10 - 4?<|start-latent|>",
]
inputs = tokenizer(questions, return_tensors="pt", padding=True).to(model.device)

# Generate for batch
outputs = model.generate(
    **inputs,
    generation_config=generation_config,
    num_return_sequences=1,
)

# Decode batch results
results = tokenizer.batch_decode(outputs, skip_special_tokens=True)
for result in results:
    print(result)

Model Architecture

Projector Module

CODI includes an optional projector module that extends hidden states:

python

# Projector configuration (if enabled in model)
projector = nn.Sequential(
    nn.Dropout(projector_dropout),
    nn.Linear(hidden_size, projector_hidden_size),
    nn.GELU(),
    nn.Linear(projector_hidden_size, hidden_size),
    nn.LayerNorm(hidden_size),
)

The projector is used when output_hidden_states=True and config.projector=True.

Generation Parameters

LatentGenerationConfig

max_new_tokens (int): Maximum number of tokens to generate
latent_length (int): Number of latent tokens (default: 6)
latent_do_sample (bool): Whether to use stochastic sampling
latent_do_sample_by (str): Sampling method - "dropout" or "noise"
dropout_p (float): Dropout probability for Monte Carlo Dropout (e.g., 0.1)
noise_std (float): Standard deviation for Additive Gaussian Noise

Sampling Methods

Monte Carlo Dropout: Randomly drops activations during forward passes

python
1generation_config = LatentGenerationConfig( 2 latent_do_sample_by="dropout", 3 dropout_p=0.1, 4 # ... 5)
Additive Gaussian Noise: Injects noise into latent embeddings

python
1generation_config = LatentGenerationConfig( 2 latent_do_sample_by="noise", 3 noise_std=0.1, 4 # ... 5)

Answer Extraction

CODI uses standard number extraction from the generated text:

python

from src.paths import extract_answer_number

# Extract answer from generated text
answer = extract_answer_number(result)
print(f"Answer: {answer}")

Evaluation

Run evaluation using the provided scripts:

bash

# For CODI (GPT-2 based models)
./run_tests.sh

Model Card

Paper: Parallel Test-Time Scaling for Latent Reasoning Models
HuggingFace: dd101bb/latent-tts-codi
Benchmarks: GSM8K Test, GSM8K Hard, MultiArith

Citation

If you use this model, please cite:

bibtex

@misc{you2025paralleltesttimescalinglatent,
      title={Parallel Test-Time Scaling for Latent Reasoning Models}, 
      author={Runyang You and Yongqi Li and Meng Liu and Wenjie Wang and Liqiang Nie and Wenjie Li},
      year={2025},
      eprint={2510.07745},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.07745}, 
}

@misc{shen2025codicompressingchainofthoughtcontinuous,
      title={CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation}, 
      author={Zhenyi Shen and Hanqi Yan and Linhai Zhang and Zhanghao Hu and Yali Du and Yulan He},
      year={2025},
      eprint={2502.21074},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.21074}, 
}