adu2021 Refusal Falls off a Cliff: Safety Alignment Failures in Reasoning Models
Core Concept
Large reasoning models maintain refusal intentions during internal reasoning but experience sharp alignment degradation at generation's final tokens, allowing jailbreaks to succeed. This "refusal cliff" is not uniform failure but concentrated in specific attention heads. Mechanistic analysis identifies problematic heads; targeted data curation fixes them efficiently.
Architecture Overview
- Linear Probing: Trace refusal intentions across token positions to locate cliff
- Causal Intervention: Identify specific attention heads causing degradation
- Head Ablation: Minimal (3%) head removal reduces attack success below 10%
- Cliff-as-a-Judge Curation: Automatically select training examples exhibiting largest refusal drop
- Data Efficiency: 1.7% of vanilla safety training data achieves comparable safety
Implementation Steps
1. Linear Probing for Refusal Tracking
Train linear probes to extract refusal intention at each token position.
import torch
import torch.nn as nn
class RefusalProbe:
def __init__(self, hidden_dim=4096, num_layers=48):
"""
Linear probe: map hidden states → refusal probability
"""
self.probes = nn.ModuleList([
nn.Linear(hidden_dim, 1) for _ in range(num_layers)
])
self.num_layers = num_layers
self.hidden_dim = hidden_dim
def extract_refusal_scores(self, model, prompt, target_answer):
"""
Extract refusal intention at each token position.
"""
# Forward pass with hook to capture hidden states
hidden_states_by_layer = {}
def capture_hook(module, input, output):
layer_idx = len(hidden_states_by_layer)
hidden_states_by_layer[layer_idx] = output[0] # [seq_len, batch, hidden]
# Register hooks on transformer layers
hooks = []
for layer_idx, layer in enumerate(model.transformer.h):
h = layer.register_forward_hook(capture_hook)
hooks.append(h)
# Generate tokens and capture states
with torch.no_grad():
tokens = model.tokenize(prompt)
model(tokens)
# Remove hooks
for h in hooks:
h.remove()
# Apply probes to extract refusal at each position
refusal_trajectories = {} # layer_idx → [seq_len] refusal scores
for layer_idx in range(self.num_layers):
hidden = hidden_states_by_layer.get(layer_idx)
if hidden is None:
continue
# Apply linear probe
refusal_logits = self.probes[layer_idx](hidden)
refusal_probs = torch.sigmoid(refusal_logits)
refusal_trajectories[layer_idx] = refusal_probs.squeeze().detach().cpu().numpy()
return refusal_trajectories
def detect_refusal_cliff(self, refusal_trajectories):
"""
Identify where refusal intention drops sharply.
"""
cliff_locations = {}
for layer_idx, trajectory in refusal_trajectories.items():
# Compute first derivative (change in refusal score)
diffs = np.diff(trajectory)
# Find largest negative jump (cliff)
cliff_idx = np.argmin(diffs) # Most negative
cliff_magnitude = diffs[cliff_idx]
if cliff_magnitude < -0.2: # Significant drop
cliff_locations[layer_idx] = {
'position': cliff_idx,
'magnitude': cliff_magnitude,
'pre_cliff_score': trajectory[cliff_idx],
'post_cliff_score': trajectory[cliff_idx + 1]
}
return cliff_locations
2. Causal Intervention Analysis
Identify which attention heads cause refusal degradation via ablation.
class AttentionHeadAnalysis:
def __init__(self, model):
self.model = model
self.num_heads = model.config.num_attention_heads
self.num_layers = model.config.num_hidden_layers
def ablate_attention_head(self, layer_idx, head_idx):
"""
Ablate specific attention head by zeroing its outputs.
"""
def ablation_hook(module, input, output):
# output = (attn_output, attn_weights)
attn_output, attn_weights = output
# Zero out specific head
head_dim = attn_output.shape[-1] // self.num_heads
start = head_idx * head_dim
end = (head_idx + 1) * head_dim
attn_output[:, :, start:end] = 0
return (attn_output, attn_weights)
# Register ablation hook
layer = self.model.transformer.h[layer_idx].self_attn
hook = layer.register_forward_hook(ablation_hook)
return hook
def evaluate_head_importance(self, jailbreak_prompt, safety_loss_fn):
"""
Measure each head's contribution to safety by ablating and measuring loss.
"""
critical_heads = []
for layer_idx in range(self.num_layers):
for head_idx in range(self.num_heads):
# Baseline safety performance
with torch.no_grad():
baseline_output = self.model(jailbreak_prompt)
baseline_loss = safety_loss_fn(baseline_output)
# With head ablated
hook = self.ablate_attention_head(layer_idx, head_idx)
with torch.no_grad():
ablated_output = self.model(jailbreak_prompt)
ablated_loss = safety_loss_fn(ablated_output)
hook.remove()
# Importance: how much does ablation hurt safety?
importance = baseline_loss - ablated_loss # Positive = important for safety
if importance > 0.1: # Threshold for critical heads
critical_heads.append({
'layer': layer_idx,
'head': head_idx,
'importance': importance
})
# Sort by importance
critical_heads.sort(key=lambda x: x['importance'], reverse=True)
return critical_heads
def batch_ablate_critical_heads(self, critical_heads, ablation_ratio=0.03):
"""
Ablate top critical heads (e.g., 3% of total).
"""
num_total_heads = self.num_layers * self.num_heads
num_to_ablate = max(1, int(num_total_heads * ablation_ratio))
heads_to_ablate = critical_heads[:num_to_ablate]
# Zero out these heads permanently
for head_info in heads_to_ablate:
layer_idx = head_info['layer']
head_idx = head_info['head']
layer = self.model.transformer.h[layer_idx].self_attn
head_dim = layer.hidden_size // self.num_heads
start = head_idx * head_dim
end = (head_idx + 1) * head_dim
# Permanently zero weight
with torch.no_grad():
layer.dense.weight[:, start:end] = 0
if layer.dense.bias is not None:
layer.dense.bias[start:end] = 0
print(f"Ablated {len(heads_to_ablate)} critical heads")
return self.model
3. Cliff-as-a-Judge Data Curation
Automatically select training examples exhibiting largest refusal degradation.
def cliff_as_a_judge_curation(model, safety_training_pool, num_examples=None):
"""
Data curation: select examples exhibiting largest refusal cliff.
These examples are most important for safety training.
"""
probe = RefusalProbe()
curated_examples = []
for example in safety_training_pool:
prompt = example['harmful_prompt']
target = example['safe_refusal']
# Extract refusal trajectory
refusal_scores = probe.extract_refusal_scores(model, prompt, target)
# Find cliff magnitude
cliff_magnitude = 0
for layer_idx, trajectory in refusal_scores.items():
diffs = np.diff(trajectory)
worst_diff = np.min(diffs)
cliff_magnitude = min(cliff_magnitude, worst_diff)
# Score: larger cliff = more important
cliff_score = abs(cliff_magnitude) # 0-1 range
curated_examples.append({
'example': example,
'cliff_score': cliff_score
})
# Sort by cliff score (largest cliffs first)
curated_examples.sort(key=lambda x: x['cliff_score'], reverse=True)
# Select top examples
if num_examples is None:
num_examples = int(0.017 * len(safety_training_pool)) # 1.7% of pool
selected = curated_examples[:num_examples]
print(f"Selected {len(selected)} examples (1.7% of pool) with largest refusal cliffs")
return [ex['example'] for ex in selected]
4. Safety Training with Curated Data
Retrain on curated examples for efficient safety recovery.
def train_safety_with_curated_data(model, curated_examples, num_epochs=3):
"""
Train on cliff-detected examples for efficient safety improvement.
"""
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)
safety_loss_fn = nn.CrossEntropyLoss()
for epoch in range(num_epochs):
total_loss = 0
for example in curated_examples:
prompt = example['harmful_prompt']
safe_response = example['safe_refusal']
# Forward pass
logits = model(prompt)
# Loss: predict safe response
loss = safety_loss_fn(logits, safe_response)
# Backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
print(f"Epoch {epoch+1}: Safety loss={total_loss/len(curated_examples):.4f}")
return model
# Experimental results
results = {
'head_ablation': {
'heads_ablated': '3%',
'attack_success_rate_before': '80%',
'attack_success_rate_after': '<10%',
},
'data_curation': {
'vanilla_safety_training': {
'data_size': '100%',
'safety_improvement': 'Baseline'
},
'cliff_as_a_judge': {
'data_size': '1.7%',
'safety_improvement': 'Comparable to vanilla',
'token_cost': '~60x reduction'
}
}
}
Practical Guidance
Probe Training: Train refusal probes on clean refusal examples (high safety score) vs jailbreak attempts (low score). Use held-out validation for probe quality.
Head Identification: Ablate iteratively; stop when safety improves sufficiently. 3% ablation (3-5 heads on 48-layer models) is typical sweet spot.
Data Curation: Cliff score correlates with retraining importance. Top 1-2% of examples by cliff score provide 80% of safety benefit.
Training Efficiency: Use smaller learning rate (1e-5 vs 1e-4) to avoid destabilizing base model while focusing on safety pathways.
When to Use / When NOT to Use
Use When:
- Deploying reasoning models with safety requirements
- Attack vectors exploit final-token refusal degradation
- Data efficiency is critical (limited retraining budget)
- You need interpretable safety improvements (ablate specific heads)
NOT For:
- Non-reasoning models without clear refusal cliff
- Scenarios where broad retraining is feasible
- Domains requiring complete alignment review
Reference
This skill synthesizes findings from "Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning" (arXiv:2510.06036). Mechanistic analysis reveals concentrated safety vulnerabilities fixable via targeted intervention.