Granite-Abstract: Continuous Head Embeddings for Reasoning

A language model with learnable continuous embeddings for abstract reasoning, trained SFT and reinforcement learning.

Overview

Granite-Abstract extends Granite with a continuous head that generates soft embeddings instead of discrete tokens. This enables the model to perform internal reasoning more efficiently while maintaining compatibility with standard generation.

Key characteristics:

• Parallel continuous head for soft embeddings
• Two-phase generation: abstract reasoning → natural output
• Trainable continuous components via reinforcement learning
• 500M parameters total (Granite 350M + Continuous Head 150M)

Architecture

Dual-Head System

The model operates with two parallel heads:

Input → Embedding Layer → Granite Backbone → Hidden State
                                                 ↓
                                    ┌─────────────┴─────────────┐
                                    ↓                           ↓
                            Continuous Head              LM Head
                          (soft embeddings)          (discrete tokens)

Generation Process

Phase 1: Abstract Mode - Model generates continuous embeddings for internal reasoning

cont_logits = continuous_head(hidden_state)
top_logits, top_indices = topk(cont_logits, 256)
next_embedding = softmax(top_logits) @ embed_layer[top_indices]

Phase 2: Transition - At </think> token, switches to output mode

Phase 3: Natural Mode - Generates discrete tokens for human-readable output

logits = lm_head(hidden_state)
next_token = argmax(softmax(logits))

Training Process

Stage 1: Supervised Fine-Tuning (SFT)

• Train Granite on reasoning tasks with <think>...</think>Answer format
• Best checkpoint: epoch 2 (validation loss: 1.2729)

Stage 2: Initialize Continuous Head

• Copy embedding and LM head weights from SFT model
• Freeze Granite backbone
• Only continuous components are trainable

Stage 3: Reinforcement Learning

• Algorithm: Gaussian Group Relative Policy Optimization (GRPO)
• Sample 4 completions per prompt
• Score with reward = answer_correctness + format_compliance
• Update via policy gradient with AdamW (lr=1e-5)
• Gaussian noise (σ=0.1) during training for exploration

Performance

Benchmark Evaluation

Results on three standard benchmarks (1024 samples each):

Model	MMLU	GSM8K	DROP	Overall
Gemma 3 (270M)	19.24%	0.39%	0.98%	6.87%
Granite 4 Nano (350M)	7.13%	4.00%	9.96%	7.03%
Granite 4 Nano SFT (350M)	6.17%	6.54%	10.06%	7.59%
Abstract Granite RL (500M)	5.76%	7.42%	10.45%	7.876%
DeepSeek R1 Qwen (1.5B)	0.20%	0.49%	0.10%	0.26%

Efficiency Analysis

Despite having fewer parameters than larger baselines, Abstract Granite achieves competitive performance through specialized training for reasoning tasks.

Usage

Installation

pip install torch transformers

Basic Inference

from abstract_model import AbstractModel

model = AbstractModel.load_from_directory(
    output_dir='./models/granite-abstract',
    sft_model_path='./models/granite-sft',
    device='cuda'
)

messages = [{"role": "user", "content": "What is 5 + 3?"}]
formatted = model.tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
input_ids = model.tokenizer(formatted, return_tensors='pt')['input_ids'].to('cuda')

result = model.forward(input_ids, max_length=256, temperature=0.7)
response = model.tokenizer.decode(result['generated_tokens'])
print(response)
print("Modes:", result['mode_sequence'])  # A=abstract, N=natural, T=transition