Pairwise Comparison Evaluation

Pairwise evaluation compares two agent versions on the same inputs and determines which produces better outputs. This is especially useful when absolute scoring is difficult but relative comparison is easier.

When to Use Pairwise Evaluation

Use pairwise comparison when:

Comparing agent versions - Which agent is more concise, helpful, or accurate?
Relative quality matters - It’s easier to say “A is better than B” than “A scores 0.73”
Subjective criteria - Conciseness, tone, style
A/B testing decisions - Which version should go to production?

Pairwise evaluation works best for comparing two versions on the same dataset. For evaluating a single agent, use standard evaluators instead.

Basic Pairwise Evaluation

Run experiments on both agents, then compare:

Real-World Example: Conciseness Comparison

This example from the OfficeFlow course compares two agent versions on conciseness:

eval_conciseness_pairwise.py

"""
Pairwise conciseness evaluator for comparing two experiments.

Run this AFTER run_agents.py, using the experiment names it outputs.

Usage:
    uv run python eval_conciseness_pairwise.py <experiment-a> <experiment-b>
    uv run python eval_conciseness_pairwise.py agent-v4-3e016f9c agent-v5-7d7ee287
"""

from openai import OpenAI
from langsmith import evaluate

client = OpenAI()

CONCISENESS_PROMPT = """
You are evaluating two responses to the same customer question.
Determine which response is MORE CONCISE while still providing all crucial information.

**Conciseness** means getting straight to the point, avoiding filler, and not repeating information.
**Crucial information** includes direct answers, necessary context, and required next steps.

A shorter response is NOT automatically better if it omits crucial information.

**Question:** {question}

**Response A:**
{response_a}

**Response B:**
{response_b}

Output your verdict as a single number:
1 if Response A is more concise while preserving crucial information
2 if Response B is more concise while preserving crucial information
0 if they are roughly equal
"""

def conciseness_evaluator(inputs: dict, outputs: list[dict]) -> list[int]:
    response = client.chat.completions.create(
        model="gpt-5-nano",
        messages=[
            {"role": "system", "content": "You are a conciseness evaluator. Respond with only a single number: 0, 1, or 2."},
            {"role": "user", "content": CONCISENESS_PROMPT.format(
                question=inputs["question"],
                response_a=outputs[0].get("answer", "N/A"),
                response_b=outputs[1].get("answer", "N/A"),
            )}
        ],
    )
    
    preference = int(response.choices[0].message.content.strip())
    
    if preference == 1:
        return [1, 0]  # A wins
    elif preference == 2:
        return [0, 1]  # B wins
    else:
        return [0, 0]  # Tie

if __name__ == "__main__":
    import sys
    if len(sys.argv) != 3:
        print("Usage: python eval_conciseness_pairwise.py <experiment-a> <experiment-b>")
        print("Example: python eval_conciseness_pairwise.py agent-v4-3e016f9c agent-v5-7d7ee287")
        sys.exit(1)
    
    evaluate(
        (sys.argv[1], sys.argv[2]),
        evaluators=[conciseness_evaluator],
        randomize_order=True,
    )

randomize_order=True is critical - it prevents position bias where the LLM might favor the first or second response systematically.

Complete Workflow

Here’s how to run both agents and compare them in one script:

run_pairwise_experiment.py

import asyncio
import sys
from pathlib import Path

agent_dir = Path(__file__).resolve().parent.parent.parent / "officeflow-agent"
sys.path.insert(0, str(agent_dir))

from langsmith import aevaluate, evaluate
from agent_v4 import chat as chat_v4, load_knowledge_base as load_kb_v4
from agent_v5 import chat as chat_v5, load_knowledge_base as load_kb_v5
from eval_conciseness_pairwise import conciseness_evaluator
from dotenv import load_dotenv

load_dotenv()

DATASET_NAME = "officeflow-dataset"
KB_DIR = str(agent_dir / "knowledge_base")

async def chat_wrapper_v4(inputs: dict) -> dict:
    question = inputs.get("question", "")
    result = await chat_v4(question)
    return {"answer": result["output"]}

async def chat_wrapper_v5(inputs: dict) -> dict:
    question = inputs.get("question", "")
    result = await chat_v5(question)
    return {"answer": result["output"]}

async def main():
    # Load knowledge bases for both agents
    print("Loading knowledge bases...")
    await load_kb_v4(KB_DIR)
    await load_kb_v5(KB_DIR)
    
    # Step 1: Run experiment for agent v4
    print("\n" + "="*50)
    print("Running experiment for agent_v4...")
    print("="*50)
    v4_results = await aevaluate(
        chat_wrapper_v4,
        data=DATASET_NAME,
        experiment_prefix="agent-v4",
    )
    
    # Step 2: Run experiment for agent v5
    print("\n" + "="*50)
    print("Running experiment for agent_v5...")
    print("="*50)
    v5_results = await aevaluate(
        chat_wrapper_v5,
        data=DATASET_NAME,
        experiment_prefix="agent-v5",
    )
    
    # Get experiment names from results
    v4_experiment = v4_results.experiment_name
    v5_experiment = v5_results.experiment_name
    
    print(f"\nv4 experiment: {v4_experiment}")
    print(f"v5 experiment: {v5_experiment}")
    
    # Step 3: Run pairwise evaluation
    print("\n" + "="*50)
    print("Running pairwise evaluation...")
    print("="*50)
    evaluate(
        (v4_experiment, v5_experiment),
        evaluators=[conciseness_evaluator],
        randomize_order=True,
    )
    
    print("\n" + "="*50)
    print("Done! Check LangSmith for results.")
    print("="*50)

if __name__ == "__main__":
    asyncio.run(main())

Pairwise Evaluator Patterns

Simple Preference

The evaluator returns [1, 0], [0, 1], or [0, 0]:

preference_scores.py

def pairwise_evaluator(inputs: dict, outputs: list[dict]) -> list[int]:
    # Determine which is better
    if outputs[0]["answer"] is better:
        return [1, 0]  # First wins
    elif outputs[1]["answer"] is better:
        return [0, 1]  # Second wins
    else:
        return [0, 0]  # Tie

Strength of Preference

Return scores showing how much better one is:

preference_strength.py

def pairwise_with_strength(inputs: dict, outputs: list[dict]) -> list[float]:
    # Score each from 0-1
    score_a = evaluate_quality(outputs[0])
    score_b = evaluate_quality(outputs[1])
    
    # Normalize so they sum to 1
    total = score_a + score_b
    if total > 0:
        return [score_a / total, score_b / total]
    else:
        return [0.5, 0.5]  # Equal if both scored 0

Multiple Criteria

Compare on several dimensions:

multi_criteria_pairwise.py

from openai import OpenAI
import json

client = OpenAI()

def multi_criteria_pairwise(inputs: dict, outputs: list[dict]) -> list[dict]:
    """Compare on multiple criteria and return detailed scores."""
    
    prompt = f"""
    Compare these two responses on multiple criteria:
    
    Question: {inputs['question']}
    Response A: {outputs[0]['answer']}
    Response B: {outputs[1]['answer']}
    
    For each criterion, output which response is better (1 for A, 2 for B, 0 for tie):
    - Conciseness
    - Accuracy
    - Helpfulness
    - Professionalism
    
    Return JSON:
    {{
      "conciseness": 1 or 2 or 0,
      "accuracy": 1 or 2 or 0,
      "helpfulness": 1 or 2 or 0,
      "professionalism": 1 or 2 or 0
    }}
    """
    
    response = client.chat.completions.create(
        model="gpt-5-nano",
        messages=[{"role": "user", "content": prompt}],
        response_format={"type": "json_object"},
        temperature=0
    )
    
    results = json.loads(response.choices[0].message.content)
    
    # Calculate overall winner
    a_wins = sum(1 for v in results.values() if v == 1)
    b_wins = sum(1 for v in results.values() if v == 2)
    
    return [
        {"score": 1 if a_wins > b_wins else 0, "details": results},
        {"score": 1 if b_wins > a_wins else 0, "details": results}
    ]

Best Practices

Always Randomize Order

Prevent position bias:

always_randomize.py

evaluate(
    (exp_a, exp_b),
    evaluators=[your_evaluator],
    randomize_order=True  # Critical!
)

Provide Clear Criteria

Define what “better” means:

clear_criteria.py

GOOD_PROMPT = """
Which response is MORE CONCISE while still providing all crucial information?

**Conciseness** means:
- Getting straight to the point
- Avoiding unnecessary filler words
- Not repeating information

**Crucial information** includes:
- Direct answer to the question
- Necessary context
- Required next steps
"""

VAGUE_PROMPT = """
Which response is better?
"""

Use Descriptive Evaluation Names

descriptive_names.py

# Good: Clear what's being compared
evaluate(
    ("agent-v4-concise", "agent-v5-verbose"),
    evaluators=[conciseness_evaluator],
    experiment_prefix="conciseness-comparison"
)

# Bad: Unclear
evaluate(
    ("exp1", "exp2"),
    evaluators=[evaluator]
)

Consider Multiple Judges

Use multiple models or run multiple times for confidence:

multiple_judges.py

def multi_judge_pairwise(inputs: dict, outputs: list[dict]) -> list[int]:
    """Get consensus from multiple evaluation runs."""
    
    votes_a = 0
    votes_b = 0
    
    # Run evaluation 3 times
    for _ in range(3):
        result = single_judge(inputs, outputs)
        if result == [1, 0]:
            votes_a += 1
        elif result == [0, 1]:
            votes_b += 1
    
    # Return majority vote
    if votes_a > votes_b:
        return [1, 0]
    elif votes_b > votes_a:
        return [0, 1]
    else:
        return [0, 0]

Interpreting Results

Pairwise results show:

Win rate: How often each agent won
Tie rate: How often outputs were equivalent
Per-example comparison: Which specific inputs show differences

Statistical Significance

For small datasets, a few wins might not be meaningful:

significance.py

from scipy.stats import binomtest

def is_significant(wins_a: int, wins_b: int, ties: int, alpha=0.05) -> bool:
    """Check if difference is statistically significant."""
    total_decisive = wins_a + wins_b
    
    if total_decisive == 0:
        return False
    
    # Binomial test: is win rate significantly different from 50%?
    result = binomtest(wins_a, total_decisive, 0.5)
    return result.pvalue < alpha

# Example
wins_a = 15
wins_b = 8
ties = 2

if is_significant(wins_a, wins_b, ties):
    print("Agent A is significantly better!")
else:
    print("Difference is not statistically significant.")

Common Use Cases

Prompt Engineering

Compare prompt variations:

prompt_comparison.py

# Two agents with different prompts
agent_a = ChatAgent(prompt="You are a helpful assistant.")
agent_b = ChatAgent(prompt="You are a concise assistant who answers in under 50 words.")

# Compare conciseness
evaluate(
    (experiment_a, experiment_b),
    evaluators=[conciseness_evaluator]
)

Model Selection

Compare different models:

model_comparison.py

# Agent with GPT-5-nano
agent_gpt5 = ChatAgent(model="gpt-5-nano")

# Agent with GPT-3.5
agent_gpt35 = ChatAgent(model="gpt-3.5-turbo")

# Compare quality vs cost
evaluate(
    (experiment_gpt5, experiment_gpt35),
    evaluators=[quality_evaluator]
)

Configuration Tuning

Compare temperature, max_tokens, etc.:

config_comparison.py

agent_temp0 = ChatAgent(temperature=0)
agent_temp1 = ChatAgent(temperature=1)

evaluate(
    (experiment_temp0, experiment_temp1),
    evaluators=[creativity_evaluator]
)

Limitations

Only Compares Two at a Time

For more than two versions, run multiple pairwise comparisons:

multi_version.py

versions = ["v1", "v2", "v3"]

# Compare all pairs
for i, v1 in enumerate(versions):
    for v2 in versions[i+1:]:
        print(f"Comparing {v1} vs {v2}")
        evaluate((v1, v2), evaluators=[your_evaluator])

Subject to LLM Biases

Mitigate by:

Always randomizing order
Using clear rubrics
Running multiple evaluations
Using multiple judge models

Next Steps

Create Datasets

Build test datasets for evaluation

LLM-as-Judge

Deep dive into LLM evaluation

Get Started

Core Concepts

Building Agents

Evaluation

Production

When to Use Pairwise Evaluation

Basic Pairwise Evaluation

Real-World Example: Conciseness Comparison

Complete Workflow

Pairwise Evaluator Patterns

Simple Preference

Strength of Preference

Multiple Criteria

Best Practices

Always Randomize Order

Provide Clear Criteria

Use Descriptive Evaluation Names

Consider Multiple Judges

Interpreting Results

Statistical Significance

Common Use Cases

Prompt Engineering

Model Selection

Configuration Tuning

Limitations

Only Compares Two at a Time

Subject to LLM Biases

Next Steps

Create Datasets

LLM-as-Judge

​When to Use Pairwise Evaluation

​Basic Pairwise Evaluation

​Real-World Example: Conciseness Comparison

​Complete Workflow

​Pairwise Evaluator Patterns

​Simple Preference

​Strength of Preference

​Multiple Criteria

​Best Practices

​Always Randomize Order

​Provide Clear Criteria

​Use Descriptive Evaluation Names

​Consider Multiple Judges

​Interpreting Results

​Statistical Significance

​Common Use Cases

​Prompt Engineering

​Model Selection

​Configuration Tuning

​Limitations

​Only Compares Two at a Time

​Subject to LLM Biases

​Next Steps

Create Datasets

LLM-as-Judge

When to Use Pairwise Evaluation

Basic Pairwise Evaluation

Real-World Example: Conciseness Comparison

Complete Workflow

Pairwise Evaluator Patterns

Simple Preference

Strength of Preference

Multiple Criteria

Best Practices

Always Randomize Order

Provide Clear Criteria

Use Descriptive Evaluation Names

Consider Multiple Judges

Interpreting Results

Statistical Significance

Common Use Cases

Prompt Engineering

Model Selection

Configuration Tuning

Limitations

Only Compares Two at a Time

Subject to LLM Biases

Next Steps