Step-by-Step Teaching
Step 1: What is LLM Conditioning?
Definition: Conditioning means controlling what an LLM outputs — either:
- Injecting a concept → making the model talk about something
- Removing a concept → making the model avoid something (e.g., toxic content)
Real-world example:
You want a chatbot to always respond formally (injection) or never produce harmful content (removal)
Step 2: The Core Trade-Off
HIGH EFFECTIVENESS HIGH FLUENCY
(concept well controlled) (natural, readable output)
↑ ↑
|___________________________|
TENSION EXISTS
Key finding from the paper:
Methods that are very good at steering behavior often damage the natural quality of generated text
Think of it like this:
- Forcing a person to only use formal words → they sound robotic
- The more you constrain → the less natural the output
Step 3: Major Conditioning Methods
| Method | How It Works | Cost |
|---|---|---|
| Prompting | Add instructions in the input text | Very cheap |
| Activation Steering | Directly modify internal model activations | Medium |
| Supervised Fine-Tuning (SFT) | Retrain model on target examples | Expensive |
Step 4: What Works for What?
Concept Injection (adding a behavior):
✅ Prompting — works well, cheap
✅ Fine-tuning (SFT) — works well, expensive
⚠️ Activation Steering — works but hurts fluency
Concept Removal (blocking a behavior):
❌ Prompting — less reliable
❌ Fine-tuning — less reliable
⚠️ Activation Steering — effective but costly to fluency
Key insight: No single method dominates across both scenarios
Step 5: Critical Interaction — Base vs. Instruction-Tuned Models
This is a previously overlooked finding:
Base Model → Activation Steering WORKS WELL
Instruction-Tuned → Activation Steering MUCH LESS EFFECTIVE
Why does this matter?
- Most deployed models are instruction-tuned (e.g., ChatGPT-style)
- Researchers often test on base models
- Results do not transfer reliably
Analogy:
Teaching new habits to someone with no prior training (base) is easier than retraining someone with deeply ingrained habits (instruction-tuned)
Step 6: How Do We Measure Success?
Two types of evaluation metrics:
| Metric Type | Example | Cost |
|---|---|---|
| Textual metrics | Perplexity, n-gram scores | 💚 Cheap |
| LLM-as-judge | Ask another LLM to score quality | 🔴 Expensive |
Important finding:
Cheap textual metrics highly correlate with expensive LLM-judge scores
✅ This means researchers can use affordable metrics without losing evaluation quality
Step 7: Bonus Concept — Dynamic Steering (DSAS)
A problem with standard activation steering:
It applies the same intervention even when unnecessary → degrades performance
Solution — Dynamically Scaled Activation Steering (DSAS):
Input arrives
↓
Does this input NEED steering?
├── YES → Apply steering
└── NO → Don't intervene (preserve fluency)
Key principle: Decouple when to steer from how to steer
Summary Table
| Concept | Key Takeaway |
|---|---|
| Conditioning | Injecting or removing concepts from LLM output |
| Core trade-off | More effective steering = less fluent output |
| Best for injection | Prompting or Fine-tuning |
| Best for removal | Activation steering (with fluency cost) |
| Model type matters | Instruction-tuned models resist activation steering |
| Evaluation | Cheap text metrics ≈ expensive LLM-judge scores |
| DSAS | Steer only when needed to preserve quality |
Quick Self-Check Questions
- What are the two scenarios studied in LLM conditioning?
- Why does activation steering perform worse on instruction-tuned models?
- What is the practical benefit of textual metrics correlating with LLM-judge scores?
- How does DSAS improve upon standard activation steering?