Skip to main content 首页 创作者 adu2021 skillxiv refusal-falls-off-cliff-safety-reasoning
refusal-falls-off-cliff-safety-reasoning Identify and patch critical safety vulnerabilities in large reasoning models. Via linear probing and causal intervention, locate specific attention heads responsible for alignment degradation at final tokens. Recover safety via 'Cliff-as-a-Judge' data curation targeting examples exhibiting largest refusal decline, achieving comparable improvements using only 1.7% of vanilla safety training data.
跳到安装 Skills Marketplace 发现并探索由社区构建的 Agent Skills
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill refusal-falls-off-cliff-safety-reasoning命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
下载 Zip 下载中... name refusal-falls-off-cliff-safety-reasoning title Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2510.06036 keywords ["safety alignment","reasoning models","mechanistic interpretability","refusal cliff","attention analysis"] description Identify and patch critical safety vulnerabilities in large reasoning models. Via linear probing and causal intervention, locate specific attention heads responsible for alignment degradation at final tokens. Recover safety via 'Cliff-as-a-Judge' data curation targeting examples exhibiting largest refusal decline, achieving comparable improvements using only 1.7% of vanilla safety training data.
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 cliffCausal Intervention : Identify specific attention heads causing degradationHead Ablation : Minimal (3%) head removal reduces attack success below 10%Cliff-as-a-Judge Curation : Automatically select training examples exhibiting largest refusal dropData 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.
"""
hidden_states_by_layer = {}
def
capture_hook
module, input , output
len
0
for
in
enumerate
with
for
in
for
in
range
self
if
is
None
continue
self
return
def
detect_refusal_cliff
self, refusal_trajectories
"""
Identify where refusal intention drops sharply.
"""
for
in
if
0.2
'position'
'magnitude'
'pre_cliff_score'
'post_cliff_score'
1
return
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 ):
attn_output, attn_weights = output
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)
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):
with torch.no_grad():
baseline_output = self .model(jailbreak_prompt)
baseline_loss = safety_loss_fn(baseline_output)
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 = baseline_loss - ablated_loss
if importance > 0.1 :
critical_heads.append({
'layer' : layer_idx,
'head' : head_idx,
'importance' : 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]
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
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' ]
refusal_scores = probe.extract_refusal_scores(model, prompt, target)
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)
cliff_score = abs (cliff_magnitude)
curated_examples.append({
'example' : example,
'cliff_score' : cliff_score
})
curated_examples.sort(key=lambda x: x['cliff_score' ], reverse=True )
if num_examples is None :
num_examples = int (0.017 * len (safety_training_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' ]
logits = model(prompt)
loss = safety_loss_fn(logits, safe_response)
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
print (f"Epoch {epoch+1 } : Safety loss={total_loss/len (curated_examples):.4 f} " )
return model
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
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)
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.
