JulianHJR/crest-f

CREST: Cognitive REasoning Steering at Test‑time

TL; DR

CREST is a training-free test-time steering framework that discovers cognitive heads via simple offline calibration and then rotates activations during decoding to guide the model’s reasoning—preserving norms to avoid per-model hyperparameter tuning. This improves accuracy and reduces tokens across models and datasets.

What is CREST?

CREST (Cognitive REasoning Steering at Test-time) identifies attention heads whose activations are predictive of different reasoning modes (“cognitive heads”), then steers those heads at inference to suppress inefficient trajectories and encourage effective reasoning—without further training.

• Token savings with accuracy gains. e.g., R1-7B on MATH500: 92.4% with 34% fewer tokens; R1-1.5B on AMC23: 37.6% token reduction at higher accuracy.
• Generalizes across models/datasets (DeepSeek-R1 1.5B/7B/32B, Qwen3-4B/30B, GPT-OSS-20B; MATH500, AIME, AMC23, GPQA-D, LiveCodeBench, Calendar Planning).
• Head ratio “gold default.” Steering about the top ~38% heads (ranked by linear-probe accuracy) balances accuracy and token reduction; adopted as the default.

Usage

Install

Recommend use vllm docker: vllm/vllm-openai@sha256:d731ee65c044ae0977421eed3d93f931d4b7d79614394184c939db35b8f28fc2 & docker_info.txt

inside the docker:

cd CREST/probing/omni_math_rule/evaluation
pip install evaluate
cd latex2sympy
pip install -e .
cd ..
pip install -r requirements.txt 
cd ../../
pip install lighteval
pip install datasets==3.5.0
pip install emoji

otherwise:

cd CREST
pip install requirements.txt
cd probing/omni_math_rule/evaluation
pip install evaluate
cd latex2sympy
pip install -e .
cd ..
pip install -r requirements.txt 
cd ../../
pip install lighteval
pip install datasets==3.5.0
pip install emoji

Running Baseline Experiments

# Run baseline without steering
bash script/baseline.sh

This script:

• Sets STEERING=False
• Evaluates models across multiple datasets
• Saves results in results/SteeringFalse/ directory

Running Steering Experiments

# Run with steering enabled
bash script/ours.sh

This script:

• Sets STEERING=True
• Configures steering parameters:
  • steering_number=512: Number of top steering vectors to use
  • steering_coef=-4: Steering coefficient (negative for inhibition)
  • steering_mode=after_o_proj_norm_threshold: Steering application mode
• Saves results in results/SteeringTrue_numb512_coef-4_mode[mode]/ directory
• NOW STEERING MUST SET --enforce_eager, CUDA GRAPH AND TORCH COMPILE ARE NOT SUPPORTED YET

Steering Vector Zoo

The Steering Vector Zoo contains pre-trained steering vectors that can be applied to modify model behavior during inference. These vectors are learned through probing techniques and stored as PyTorch (.pt) files.

probing/results/
├── [DATASET]/ # Training dataset (e.g., MATH_train)
│ └── [MODEL]/ # Model name (e.g., Qwen3-30B-A3B-Thinking-2507)
│ └── template-t0-n1-[SIZE]/ # Template and size configuration
│  └── hidden[MODE]/ # Steering mode directory
|   └── mix_others_low_rank_1000/ # Training methods
│    └── probe_best.pt # Steering vector file

The system automatically selects the top-k most accurate steering vectors based on:

1. Probe Accuracy: Classification performance on the validation set
2. Layer-Head Coverage: Distribution across different attention layers and heads
3. Steering Coefficient: Magnitude of influence (configurable via --steering_coef)

Manual Execution

python main_vllm.py \
    --model_name_or_path "Qwen/Qwen3-30B-A3B-Thinking-2507" \
    --dataset "aime25" \
    --save_dir "results/test/" \
    --use_chat_format \
    --temperature 0.6 \
    --max_tokens 32768 \
    --steering True \
    --steering_vector_path "/path/to/steering/vectors/" \
    --steering_number 512 \
    --steering_coef -4 \
    --steering_mode "after_o_proj_norm_threshold"

Required Environment Variables (for steering)

export STEERING=True
export STEERING_VECTOR_PATH="/path/to/steering/vectors/"
export STEERING_NUMBER=512
export STEERING_COEF=-4
export STEERING_MODE="after_o_proj_norm_threshold"
export MODEL_NAME_OR_PATH="Qwen/Qwen3-30B-A3B-Thinking-2507"

Repository overview

The implementation consists of three main components:

1. Main Inference Engine (main_vllm.py)

• Purpose: Core inference script supporting multiple datasets and models
• Features:
  • Support for multiple datasets (MATH, GSM, AIME, GPQA, LiveCodeBench, etc.)
  • Batch processing with vLLM backend
  • Configurable steering parameters
  • Multi-model support with automatic model detection

2. Steering Implementation (probing/modeling_utils/vllm/)

Contains model-specific monkey patches for different architectures:

• qwen2/monkey_patch.py: DeepSeek-R1-Distill-Qwen models
• qwen3/monkey_patch.py: Qwen3-4B-Thinking models
• qwen3_moe/monkey_patch.py: Qwen3-30B-A3B-Thinking models
• gpt_oss/monkey_patch.py: GPT-OSS models

Each monkey patch implements:

• Attention mechanism modifications
• Layer-wise steering vector application
• Dynamic steering flag management
• Multiple steering modes

3. Evaluation System (probing/get_omni_results.py)

• Purpose: Mathematical reasoning evaluation using OmniMath rules
• Features:
  • Parallel evaluation with timeout handling
  • Multiple prediction aggregation
  • Accuracy computation and result saving

Steering Modes

The system supports four distinct steering modes, each applying steering vectors at different points in the attention mechanism:

1. before_o_proj

• Application Point: Before the output projection in attention
• Mechanism: Modifies attention output before linear transformation
• Use Case: Early intervention in attention computation

2. after_o_proj

• Application Point: After the output projection in attention
• Mechanism: Direct modification of projected attention output
• Use Case: Standard steering application

3. after_o_proj_norm

• Application Point: After output projection with norm preservation
• Mechanism: Applies steering while preserving vector norms
• Use Case: Maintaining activation magnitude consistency

4. after_o_proj_norm_threshold

• Application Point: After output projection with thresholded norm preservation
• Mechanism: Applies steering with awareness thresholding and norm preservation
• Use Case: Selective steering based on activation patterns

Supported Datasets

• "aime25"
• "AIME"
• "amc23"
• "cp"
• "lcb"
• "MATH500"
• "GSM"
• "gpqa"

Supported Models

The system automatically detects and applies appropriate steering based on model names:

• DeepSeek-R1-Distill-Qwen: Uses qwen2 monkey patch
• Qwen3-4B: Uses qwen3 monkey patch
• Qwen3-30B: Uses qwen3_moe monkey patch
• gpt-oss: Uses gpt_oss monkey patch

Steering Vector Format

Steering vectors should be saved as PyTorch files with the following structure:

{
    layer_idx: {
        head_idx: {
            'accuracy': float,           # Probe accuracy
            'model_dict': {
                'weight': torch.Tensor   # Hyperplane weights
            },
            'steering_vector': torch.Tensor  # Steering direction
        }
    }
}

Output Structure

Results are saved in the following structure:

results/
├── SteeringFalse/                    # Baseline results
│   └── [model]/[dataset]/seed[X]/
├── SteeringTrue_numb[N]_coef[C]_mode[M]/  # Steering results
│   └── [model]/[dataset]/
└── predictions.jsonl                 # Detailed predictions
└── metrics.json                      # Evaluation metrics

Key Features

• Dynamic Steering: Steering is applied based on token patterns (e.g., double newlines)
• Multi-GPU Support: Tensor parallelism with proper rank handling
• Batch Processing: Efficient batch inference with vLLM
• Comprehensive Evaluation: Multiple evaluation metrics per dataset
• Flexible Configuration: Easy parameter tuning through environment variables

Technical Implementation

Steering Mechanism

1. Vector Loading: Top-k steering vectors are loaded based on probe accuracy
2. Flag Detection: Steering flags are set based on specific token patterns
3. Awareness Computation: Hyperplane-based awareness scores determine steering strength
4. Vector Application: Steering vectors are applied with configurable coefficients
5. Norm Preservation: Optional norm preservation maintains activation magnitudes

Model Patching

The system uses runtime monkey patching to modify:

• forward() methods in attention layers
• forward() methods in decoder layers
• forward() methods in the main model
• Token-based steering flag management

Performance Considerations

• Memory: Steering vectors are loaded onto GPU memory
• Scalability: Supports tensor parallelism for large models
• Efficiency: Conditional application reduces unnecessary computation

Research Applications

This implementation supports research in:

• Mechanistic interpretability of reasoning
• Activation steering and control
• Cognitive enhancement in language models
• Mathematical reasoning improvement
• Code generation optimization

Citation

If you use this code or ideas, please cite the CREST paper:

Understanding and Steering The Cognitive Behaviors of Reasoning Models At Test-Time (CREST)