Skip to content

Latest commit

 

History

History
377 lines (276 loc) · 13.9 KB

File metadata and controls

377 lines (276 loc) · 13.9 KB

6. Experimental Results

This chapter presents the experimental findings from our evaluation of controllable skill forgetting across multiple frontier language models. The results are organized into seven subsections, progressing from prompting strategy comparisons to fine-grained skill-level analyses.


6.1 Prompting Strategy Comparison

Research Question

How effective are different prompting strategies at enforcing skill-level behavioral control?

Methodology

We measured Root Mean Squared Error (RMSE) between the intended binary skill vector and observed per-skill accuracy. Lower RMSE indicates better alignment with the desired student profile.

Key Finding: Combined Strategy Dominance

Prompting Strategy RMSE Range Performance
Combined (Rules + Few-Shot) 0.08 - 0.12 Excellent
Few-Shot Only 0.45 - 0.60 Poor
Rule-Based Only 0.50 - 0.60 Poor

Interpretation

The combined prompting strategy significantly outperforms both individual approaches across both Grade 4 and Grade 5 datasets. This synergy occurs because:

  • Rules explicitly define desired behavior
  • Examples demonstrate how behavior manifests in practice
  • Together, they enable better internalization of skill constraints

Implications

  • Neither explicit instructions nor examples alone are sufficient
  • Prompt design is a necessary but not sufficient condition for controllability
  • All subsequent experiments use the combined prompting strategy

6.2 Per-Skill Accuracy Analysis

Fine-Grained Behavioral Adherence

The per-skill accuracy grids reveal how well models adhere to imposed student profiles at the individual skill level.

Quantitative Results

Condition Accuracy Range Interpretation
Forgotten Skills (K=0) ~0.03 - 0.09 Near-complete suppression (best cases)
Retained Skills (K=1) ~0.85 - 0.94 High preservation

Key Observations

  1. Clear Separation: Strong models demonstrate near-zero accuracy on forgotten skills while maintaining high accuracy on retained skills

  2. Skill-Dependent Variability: Controllability is not uniform across skills:

    • Certain skills are easier to suppress
    • Residual performance observed on some "forgotten" skills
    • Variation likely stems from task difficulty and pretraining biases
  3. Continuous Relaxation: The accuracy vector approximates a continuous relaxation of the binary skill profile:

    • Skills marked as 1 (mastered) → high accuracy values
    • Skills marked as 0 (forgotten) → low accuracy values

Figure Description: Per-Skill Accuracy Heatmaps

(For inclusion: heatmaps showing per-skill accuracy across models and skill domains for Grades 4 and 5)

  • Rows: Different models (Claude, DeepSeek, GPT-4o)
  • Columns: Skill domains
  • Color intensity: Accuracy level
  • Lower values (darker) indicate effective forgetting

6.3 Retained vs. Forgotten Skills Comparison

Research Question

Do models exhibit differential behavior between retained and forgotten skills, and how does this vary across model architectures?

Quantitative Results

Model Retained Skills Forgotten Skills Gap
Claude ~0.85-0.92 ~0.05-0.15 Strong separation
DeepSeek ~0.85-0.92 ~0.30-0.40 Moderate separation
GPT-4o ~0.85-0.92 ~0.20-0.35 Moderate separation

Model-Specific Analysis

Claude (Anthropic)

  • Forgotten Skills: ~5-15% accuracy
  • Behavior: Near-complete suppression of targeted skills
  • Controllability: Strongest among evaluated models

DeepSeek

  • Forgotten Skills: ~30-40% accuracy
  • Behavior: Partial resistance to forgetting instructions
  • Controllability: Weaker - retains substantial knowledge

GPT-4o (OpenAI)

  • Forgotten Skills: ~20-35% accuracy
  • Behavior: Clear forgetting but not to same extent as Claude
  • Controllability: Moderate, variable across skill domains

Key Finding: Localized Degradation

The performance degradation is not global but highly localized to specified skills, confirming:

  • Controlled, selective forgetting behavior
  • Not general performance decline
  • Structured, reproducible response to skill constraints

Figure Description: Retained vs. Forgotten Comparison

(For inclusion: bar chart comparing imperfect student (IS) vs. perfect student (PS) accuracy across retained and forgotten skills)

  • Grouped bars by model
  • Percentage annotations showing relative accuracy (IS/PS)
  • Clear visual separation between retained and forgotten conditions

6.4 Skill Correlation Analysis

Research Question

Are skill performances independent, or does suppressing one skill affect others?

Methodology

  • Computed Pearson correlation matrix over n simulated students
  • Each student represented by performance vector [acc_skill1, acc_skill2, ...]
  • Analyzed correlations between skill performances

Quantitative Results

Grade Correlation Range Interpretation
Grade 4 ~-0.1 (near zero) Strong independence
Grade 5 ~-0.1 to -0.24 Weak negative correlation

Interpretation

Grade 4: Strong Independence

  • Correlations approximately uniform and near zero
  • Skills operate independently
  • Supports localized controllability

Grade 5: Mild Interference

  • Slightly stronger negative correlations observed
  • Suggests weak interference effects
  • Performance in one skill may weakly decrease as another increases
  • However: Magnitude remains small, indicating limited practical impact

Mathematical Formalization

We model skill performance as a multivariate Gaussian distribution. Under weak correlations:

E[performance_on_ret | forget_i] ≈ E[performance_on_ret]

This confirms: suppressing one skill does not significantly affect performance on others.

Implication for Simulation Design

  • Skills can be controlled independently
  • Enables construction of diverse student profiles with arbitrary skill combinations
  • Validates core assumption of the skill-vector framework

Figure Description: Correlation Matrices

(For inclusion: heatmaps showing Pearson correlation matrices between skills for Grades 4 and 5)

  • Diagonal: Self-correlation (=1)
  • Off-diagonal: Cross-skill correlations
  • Color scale: Correlation strength
  • Near-zero values indicate independence

6.5 Expected vs. Actual Retained-Skill Accuracy

Research Question

Do models preserve retained skills at predicted levels when other skills are suppressed?

Methodology

  • Derived theoretical expectation for retained-skill accuracy from correlation analysis
  • Compared expected vs. actual accuracy across simulated students
  • Scatter plots for Grades 4 and 5

Quantitative Results

Model Alignment with Expectation Deviation Pattern
Claude Strong Clusters near diagonal
DeepSeek Moderate Near diagonal
GPT-4o Weak Consistent deviation below diagonal

Model-Specific Findings

Claude & DeepSeek

  • Points cluster closely around the diagonal
  • Strong agreement between expected and observed performance
  • Successfully preserve retained skills as predicted

GPT-4o

  • Consistently deviates below the diagonal
  • Actual performance lower than expected
  • Exhibits unintended degradation in retained domains
  • Suggests incomplete skill independence

Interpretation

Weak skill correlations enable localized controllability, but realization depends on model:

  • Claude: Highest fidelity to expected behavior
  • DeepSeek: Moderate alignment
  • GPT-4o: Weaker controllability due to residual cross-skill interference

Figure Description: Expected vs. Actual Scatter Plots

(For inclusion: scatter plots comparing expected retained-skill accuracy with actual accuracy)

  • Diagonal line: Perfect agreement
  • Points: Individual simulated students
  • Color/shape: Different models
  • Deviation below diagonal indicates under-performance

6.6 Prediction Score Analysis

Metric Definition

Prediction Score = Actual Retained Accuracy - Expected Retained Accuracy

Interpretation Framework

Score Value Interpretation
PS ≈ 0 Accurate alignment with expectation
PS > 0 Better-than-expected retention
PS < 0 Under-performance, unintended degradation

Quantitative Results

Model Grade 4 Grade 5 Overall
Claude Positive Positive High alignment
DeepSeek Moderate Positive Moderate Positive Partial alignment
GPT-4o Negative Marginal Positive Inconsistent

Key Findings

Claude

  • Consistently positive prediction scores
  • Strong alignment with expected retained-skill performance
  • Confirms high controllability

DeepSeek

  • Moderate positive scores
  • Partial alignment with theoretical model
  • Some variability across runs

GPT-4o

  • Negative scores in Grade 4
  • Only marginal improvement in Grade 5
  • Systematic under-performance relative to expectations
  • Weaker controllability due to consistent degradation

Figure Description: Prediction Score Bar Chart

(For inclusion: mean prediction score per model with error bars indicating variance)

  • Horizontal axis: Models
  • Vertical axis: Prediction score
  • Error bars: Variance across runs
  • Zero line: Perfect alignment reference

6.7 Forgotten Skill Heatmaps

Research Question

How effective is forgetting suppression at the individual skill level across different models and domains?

Methodology

  • Examined per-skill accuracy on forgotten skills (K=0) across all domains
  • Heatmaps for Grades 4 and 5

Quantitative Results

Model Grade 4 Grade 5 Consistency
Claude 0.00 - 0.16 0.09 - 0.20 Consistent low
DeepSeek 0.10 - 0.39 0.31 - 0.34 Consistently high
GPT-4o 0.10 - 0.38 Variable Highly variable

Model-Specific Findings

Claude: Effective Suppression

  • Consistently low accuracy across all forgotten skills
  • Strong, stable forgetting behavior
  • Minimal cross-skill interference

DeepSeek: Residual Knowledge

  • Significantly higher residual accuracy
  • Retains substantial knowledge despite instructions
  • Partial resistance to constraints

GPT-4o: Variable Behavior

  • Sometimes achieves low accuracy (0.10-0.11)
  • Other cases show high residual performance (up to 0.38)
  • Inconsistent suppression across skill domains

Figure Description: Forgotten Skill Heatmaps

(For inclusion: heatmaps showing imperfect-student accuracy on forgotten skills)

  • Rows: Models (Claude, DeepSeek, GPT-4o)
  • Columns: Skill domains
  • Cell values: Accuracy when skill is designated forgotten
  • Color intensity: Lower = more effective forgetting

6.8 Summary of Key Empirical Findings

Based on the comprehensive experimental evaluation, we identify five key empirical findings:

Finding 1: Selective Forgetting is Achievable and Measurable

All evaluated models demonstrated some degree of selective forgetting when instructed through appropriately designed prompts. The controllability metrics successfully quantify this behavior.

Finding 2: Retention Remains Stable Under Skill Removal

For retained skills, imperfect students maintained accuracy levels close to perfect student baselines (~85-92%), confirming that knowledge is largely preserved when required.

Finding 3: Skills Are Predominantly Independent

Weak correlations (~-0.1 to -0.24) between skill performances support the assumption that targeted forgetting can be localized without widespread degradation.

Finding 4: Model Architecture Strongly Affects Controllability

Substantial variation in controllability observed across models:

  • Best: Claude (Anthropic)
  • Moderate: GPT-4o (OpenAI)
  • Weakest: DeepSeek

Finding 5: Claude Provides the Most Reliable Controllable Student Simulation

Across all evaluation dimensions, Claude demonstrated:

  • Near-complete skill suppression
  • Minimal cross-skill interference
  • Highest prediction score alignment
  • Most consistent forgetting behavior

6.9 Results Tables Summary

Table 1: Prompting Strategy Comparison

Strategy Grade 4 RMSE Grade 5 RMSE Recommendation
Combined 0.08-0.10 0.10-0.12 Best
Few-Shot 0.45-0.55 0.50-0.60 Insufficient
Rule-Based 0.50-0.55 0.55-0.60 Insufficient

Table 2: Model Controllability Summary

Model Forgotten Skill Accuracy Retained Skill Accuracy Prediction Score Overall
Claude 0.05-0.15 0.85-0.92 Positive Best
DeepSeek 0.30-0.40 0.85-0.92 Moderate Positive Moderate
GPT-4o 0.20-0.35 0.85-0.92 Negative (G4) Weak

Table 3: Skill Independence Analysis

Grade Correlation Range Independence Level
Grade 4 ~-0.1 Strong
Grade 5 -0.1 to -0.24 Moderate
Both Sufficiently weak for independent control

6.10 Figures for Inclusion

The following figures should be included in the final manuscript:

  1. Figure 1: Controllable Generative Student Architecture (pipeline diagram)
  2. Figure 2: Main Experiment Pipeline (workflow diagram)
  3. Figure 3: Prompting Strategy Comparison (RMSE bar chart)
  4. Figure 4: Retained vs. Forgotten Skills Comparison (grouped bar chart)
  5. Figure 5: Skill Correlation Matrices (heatmap for Grades 4 and 5)
  6. Figure 6: Expected vs. Actual Accuracy Scatter Plots
  7. Figure 7: Prediction Score Comparison (bar chart with error bars)
  8. Figure 8: Forgotten Skill Heatmaps (Grades 4 and 5)
  9. Figure 9: Per-Skill Accuracy Grids