Advanced Evaluation

This chapter covers the complete methodology for using LLMs as evaluators (LLM-as-a-Judge), including direct scoring, pairwise comparison, rubric generation, and bias mitigation. It's not a single trick — it's a composable set of evaluation strategies.

• Direct scoring works best for objective criteria.
• Pairwise comparison works best for preference-based evaluation.
• You must swap positions to reduce position bias.
• Rubrics can dramatically reduce evaluation variance.
• Confidence calibration matters as much as the scores themselves.

What You'll Learn

• When to use direct scoring vs. pairwise comparison
• How to build reusable rubrics
• How to handle position/length/authority and other biases

When to Activate

Activate this skill when:

• Building automated evaluation pipelines for LLM outputs
• Comparing multiple model responses to select the best one
• Establishing consistent quality standards across evaluation teams
• Debugging evaluation systems that show inconsistent results
• Designing A/B tests for prompt or model changes
• Creating rubrics for human or automated evaluation
• Analyzing correlation between automated and human judgments

Core Concepts

The Evaluation Taxonomy

Evaluation approaches fall into two primary categories:

Direct Scoring: A single LLM scores the response.

• Best for: factual accuracy, instruction following, toxicity
• Failure mode: scale drift, inconsistent interpretation

Pairwise Comparison: Compare two responses and pick the better one.

• Best for: tone, style, persuasiveness
• Failure mode: position bias, length bias

Research (MT-Bench, Zheng et al., 2023) shows that for preference-based tasks, pairwise comparison is more reliable than direct scoring.

The Bias Landscape

Common biases in LLM judges:

Position Bias: Preference for whichever response appears first. Mitigation: swap positions.

Length Bias: Longer responses get higher scores. Mitigation: explicitly instruct to ignore length.

Self-Enhancement Bias: Models rate their own outputs higher. Mitigation: use different models for generation and evaluation.

Verbosity Bias: More detail = higher score. Mitigation: rubric should explicitly state "deduct for irrelevant detail."

Authority Bias: More confident tone = higher score. Mitigation: require evidence or citations.

Metric Selection Framework

One gotcha: systematic bias matters more than raw agreement rate.

Evaluation Approaches

Direct Scoring Implementation

Direct scoring requires: clear criteria, a calibrated scale, and structured output.

Criteria Definition Pattern:

Criterion: [Name]
Description: [What this criterion measures]
Weight: [Relative importance, 0-1]

Scale Calibration:

• 1-3: low cognitive load
• 1-5: best balance
• 1-10: needs detailed rubric

Prompt Structure:

You are an expert evaluator assessing response quality.
...

Requiring justification-before-score improves consistency.

Pairwise Comparison Implementation

Position Bias Mitigation Protocol:

1. A first / B second
2. B first / A second
3. Inconsistent results -> TIE + lower confidence

Rubric Generation

Rubrics reduce evaluation variance by 40-60%.

Rubric Components:

1. Level descriptions
2. Characteristics
3. Examples (optional)
4. Edge cases
5. Scoring guidelines

Practical Guidance

Evaluation Pipeline Design

Criteria Loader -> Primary Scorer -> Bias Mitigation -> Confidence Scoring

Common Anti-Patterns

• Scoring without justification
• Single-pass pairwise
• Overloaded criteria
• Missing edge cases
• Ignoring confidence calibration

Decision Framework: Direct vs. Pairwise

Objective? -> Direct scoring
Preference? -> Pairwise comparison
Reference available? -> Reference-based evaluation

Scaling Evaluation

• PoLL (Panel of LLMs)
• Hierarchical evaluation
• Human-in-the-loop

Examples

Example 1: Direct Scoring for Accuracy

Input:

Prompt: "What causes seasons on Earth?"
Response: "Seasons are caused by Earth's tilted axis. As Earth orbits the Sun,
different hemispheres receive more direct sunlight at different times of year."
Criterion: Factual Accuracy (weight: 1.0)
Scale: 1-5

Output:

{
  "criterion": "Factual Accuracy",
  "score": 5,
  "evidence": [
    "Correctly identifies axial tilt as primary cause",
    "Correctly explains differential sunlight by hemisphere",
    "No factual errors present"
  ],
  "justification": "Response accurately explains the cause of seasons with correct
scientific reasoning. Both the axial tilt and its effect on sunlight distribution
are correctly described.",
  "improvement": "Could add the specific tilt angle (23.5°) for completeness."
}

Example 2: Pairwise Comparison with Position Swap

Input:

Prompt: "Explain machine learning to a beginner"
Response A: [Technical explanation with jargon]
Response B: [Simple analogy-based explanation]
Criteria: ["clarity", "accessibility"]

First Pass (A first):

{ "winner": "B", "confidence": 0.8 }

Second Pass (B first):

{ "winner": "A", "confidence": 0.6 }

Final Result:

{
	"winner": "B",
	"confidence": 0.7,
	"positionConsistency": {
		"consistent": true,
		"firstPassWinner": "B",
		"secondPassWinner": "B"
	}
}

Minimal Rubric Template

Criterion: <name>
Description: <what good looks like>
Scale: 1-5
Levels:

-   1: <poor>
-   3: <adequate>
-   5: <excellent>
    Edge cases: <how to score ambiguous cases>

Guidelines

1. Always require justification before scores
2. Always swap positions in pairwise comparison
3. Match scale granularity to rubric specificity
4. Separate objective and subjective criteria
5. Include confidence scores
6. Define edge cases explicitly
7. Use domain-specific rubrics
8. Validate against human judgments
9. Monitor systematic bias
10. Design for iteration

Practice Task

• Write a 1-5 rubric (e.g., "answer accuracy")
• Use it to evaluate 2 model outputs and record the differences

Related Pages

• Context Engineering Fundamentals
• Context Compression Strategies
• Tool Design for Agents

Integration

• context-fundamentals
• tool-design
• context-optimization
• evaluation

References

• Eugene Yan: Evaluating the Effectiveness of LLM-Evaluators
• Judging LLM-as-a-Judge (Zheng et al., 2023)
• G-Eval: NLG Evaluation using GPT-4 (Liu et al., 2023)
• Large Language Models are not Fair Evaluators (Wang et al., 2023)

---

Skill Metadata

Created: 2024-12-24 Last Updated: 2024-12-24 Author: Muratcan Koylan Version: 1.0.0