Prompt Steering vs Activation Steering in LLMs

A controlled experimental comparison of prompt-based behavioral steering and activation steering in transformer language models.

This project investigates whether activation steering provides more stable behavioral control than prompt engineering — particularly under prompt conflict.

---

Motivation

Large Language Models (LLMs) are commonly steered using prompts:

"You are a confident and decisive expert."

However, prompt steering operates at the input level and competes with the model’s internal representations.

This project explores an alternative:

Activation Steering — directly modifying hidden states at inference time without changing model weights.

We test whether activation steering produces more stable behavioral control compared to prompt steering.

---

Research Question

Does activation steering provide more consistent control over model confidence than prompt steering, especially under conflicting instructions?

We evaluate behavior along one axis:

Confidence vs Hedging

---

Experimental Design

We compare three generation modes using the same frozen model:

1. Baseline — No steering
2. Prompt-Steered — Explicit confidence instructions
3. Activation-Steered — Hidden-state modification via steering vector

All comparisons use identical decoding parameters.

---

Model

• google/gemma-3-4b-it
• Loaded via Hugging Face Transformers
• No fine-tuning or weight updates
• Steering performed via forward hooks

---

Activation Steering Method

We construct a steering vector by contrasting hidden states from:

• Confident prompts
• Hedging prompts

The steering vector is defined as:

mean(hidden_confident) - mean(hidden_hedging)

This vector is injected into a selected transformer layer during generation:

modified_hidden = hidden + α * steering_vector

Where:

• α controls steering strength
• Layer choice controls where behavior is influenced

---

Evaluation Metrics

We compute a composite Confidence Score:

confidence_score = assertive_markers - hedge_markers - contrast_markers

Evaluation includes:

• Mean ± standard deviation over multiple runs
• Stress testing with conflicting prompt instructions
• Layer ablation
• Alpha sweep
• Coherence inspection

---

Key Findings

Prompt Steering Is Brittle

Under stress conditions:

• Confidence score decreases
• Variance increases
• Model reverts to balanced hedging

---

Activation Steering Is More Stable

Across multiple generations:

• Similar or higher mean confidence
• Lower variance
• Better resistance to prompt conflict

---

Steering–Coherence Tradeoff

Early-layer steering (e.g., Layer 6):

• Produces extremely high confidence scores
• Causes repetition and structural degeneration

Middle layers provide better balance between:

• Behavioral control
• Output coherence

---

Example Results

Stress Test (Mean ± Std):

Prompt-Steered: 1.4 ± 1.36
Activation-Steered: 2.0 ± 1.10

This demonstrates improved stability under prompt conflict.

---

Limitations

• Effects are modest in instruction-tuned models (RLHF alignment dampens shifts)
• Steering strength must be carefully tuned
• Early-layer steering can degrade coherence
• Confidence metric is heuristic-based

Activation steering is not a replacement for fine-tuning — it is an inference-time control technique.

---

What This Project Demonstrates

This project shows that:

• Activation steering can produce measurable internal bias shifts
• Behavioral control can be achieved without retraining
• There is a clear tradeoff between steering strength and coherence
• Prompt steering and activation steering behave differently under stress

---

Requirements

pip install transformers accelerate torch matplotlib

GPU recommended (tested in Google Colab with Tesla T4 GPU).

---

Future Extensions

• Compare base vs instruction-tuned models
• Add repetition and coherence metrics
• Explore other behavioral axes (e.g., risk aversion, moral caution)
• Evaluate steering transfer across prompts