Skip to main content

Documentation Index

Fetch the complete documentation index at: https://icrl.dev/docs/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

The preference learning demo shows ICRL’s ability to learn and adapt to individual user preferences, communication styles, and working patterns over time. Vanilla LLMs treat every user the same; ICRL can match verbosity, format, depth, and tone to each user’s past approved interactions.

​
The Problem

Every user has different preferences:
  • Verbosity: Some want detailed explanations, others want terse commands
  • Format: Some prefer bullet points, others want prose
  • Depth: Some want “just tell me what to do”, others want “explain why”
  • Tone: Some prefer formal, others casual
Vanilla LLMs treat every user the same. You end up repeatedly saying “Be more concise” or “Give me the command, not the explanation”.

​
How ICRL Solves This

ICRL stores successful interactions that reflect your preferences. Over time:
  1. Learns your style from past interactions you approved
  2. Retrieves preference-matched examples for similar requests
  3. Adapts responses automatically without re-prompting

​
Demo Structure

examples/preference_learning_demo
README.md
setup_demo.py
run_demo.py
user_profiles
expert_terse.json
learner_detailed.json
manager_summary.json
scenarios
seed_interactions.json
test_requests.json

​
User Profiles

ProfileStyleWantsExample
Expert TerseMinimal explanationCommands, code snippets”How do I rebase?” yields git rebase -i HEAD~3
Learner DetailedThorough explanationsWhy things work, pitfalls, examplesFull explanation with diagrams
Manager SummaryHigh-level overviewImpact, timeline, risks, decisions”Rebasing reorganizes commit history…”

​
Running the Demo

From the project root: 
cd examples/preference_learning_demo

# 1. Setup — seeds trajectories for each user profile
uv run python setup_demo.py

# 2. Run the comparison test
uv run python run_demo.py

# 3. View detailed evaluation
uv run python evaluate_responses.py

​
Expected Results

User ProfileICRL ScoreVanilla ScoreImprovement
Expert Terse~90%~40%+50%
Learner Detailed~85%~50%+35%
Manager Summary~85%~30%+55%

​
Prerequisites

  • OPENAI_API_KEY or ANTHROPIC_API_KEY set
  • uv run from project root, or python with PYTHONPATH including src/

​
Key Insight

This demo proves ICRL’s value for personalization where:
  • One size doesn’t fit all
  • User preferences are implicit in past interactions
  • The same question should have different “right” answers for different users