Variation in Verification: Understanding Verification Dynamics in Large Language Models

This is the codebase for Variation in Verification: Understanding Verification Dynamics in Large Language Models

Overview

In this paper, we study generative verifiers, which perform verification by generating chain-of-thought (CoT) reasoning followed by a binary verdict. We systematically analyze verification dynamics across three dimensions – problem difficulty, generator capability, and verifier generation capability – with empirical studies on 12 benchmarks across mathematical reasoning, knowledge, and natural language reasoning tasks using 14 open-source models (2B to 72B parameter range) and GPT-4o. Our experiments reveal three key findings about verification effectiveness: (1) Easy problems allow verifiers to more reliably certify correct responses; (2) Weak generators produce errors that are easier to detect than strong generators; (3) Verification ability is generally correlated with the verifier’s own problem-solving capability, but this relationship varies with problem difficulty.

Update

• Add the verification code
• Published the generator and verifier inference results and released the visualization code

Installation

# Clone the repository
git clone https://github.com/your-username/llm-verify-dynamics.git
cd llm-verify-dynamics
conda create --name llm-verify python=3.10.16 -y
conda activate llm-verify
pip install uv
uv pip install torch==2.5.1
uv pip install -r requirements.txt

# Set up your HuggingFace token for model access
export BASE_PATH="/path/to/your/project"
export HF_TOKEN=your_huggingface_token_here
export OPENAI_API_KEY=your_openai_api_key_here
export TOGETHER_API_KEY=your_toge_api_key_here

Verification results

1. Download preprocessed evaluation data (CoT)

We provided the generator candidate solutions and verification results in huggingface dataset.

2. Run evaluation from scratch

# View help and all parameters
bash serve_eval_verifier.sh
# Single model evaluation using local vLLM
bash examples/single_model_local.sh
# Or use API providers (OpenAI/Together)
bash examples/api_provider.sh

See the examples/ directory for more details. Modify the scripts to customize models, datasets, and evaluation parameters.

Visualization

1. RQ1 Problem difficulty influences the TPR

See the visualization/rq1.ipynb

2. RQ2 Generator capability influences the TNR

Run the visualization/rq2_vis_prepare.sh See the visualization/rq2.ipynb

3. RQ3 Verifier solving capability influences the Verifi Performance

See the visualization/rq3.ipynb

4. RQ4 / RQ5 Implication of dynamics above on TTS

The TTS inference results can downloaded from huggingface dataset. Run the visualization/rq4_5_tts.py to simulate TTS. Then vis by visualization/rq4_5_tts.ipynb