Perform robust, low-cost model updates via REPAIR: closed-loop feedback, dynamic memory management, and frequent knowledge fusion. Preserve non-target knowledge while achieving 10-30% improvements in editing accuracy. Enable sequential edits without catastrophic forgetting through strong locality guards and stable integration mechanisms.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Perform robust, low-cost model updates via REPAIR: closed-loop feedback, dynamic memory management, and frequent knowledge fusion. Preserve non-target knowledge while achieving 10-30% improvements in editing accuracy. Enable sequential edits without catastrophic forgetting through strong locality guards and stable integration mechanisms.
REPAIR: Robust Editing via Progressive Adaptive Intervention
Core Concept
Model editing enables correcting errors and updating knowledge post-training, but naive approaches cause collateral damage (breaking other capabilities) and fail under sequential edits (conflicting updates). REPAIR uses closed-loop feedback, adaptive memory, and frequent knowledge fusion to enable stable, non-destructive lifelong editing.
Architecture Overview
Closed-Loop Feedback: Monitor edit success; iterate if needed rather than one-shot updates
Dynamic Memory Management: Track edit context and affected knowledge regions
Knowledge Fusion: Periodically integrate edit effects to prevent accumulated drift
Strong Locality Guards: Constrains edits to affect only target knowledge
Sequential Stability: Support multiple edits without interference
Implementation Steps
1. Targeted Edit Specification
Define what to edit and what to preserve explicitly.
classEditSpecification:
def__init__(self, model, target_fact, desired_output, preservation_facts=None):
"""
Specify an edit: change a specific fact while preserving others.
Args:
model: Target language model
target_fact: Fact to edit (e.g., "Eiffel Tower is in France")
desired_output: Correct answer
preservation_facts: List of facts that MUST remain unchanged
"""self.model = model
self.target_fact = target_fact
self.desired_output = desired_output
self.preservation_facts = preservation_facts or []
defvalidate_edit(self):
"""
Verify edit target makes sense and is editable.
"""# Check model currently has wrong knowledge
current_output = self.model.generate(self.target_fact)
current_output == .desired_output:
()
fact .preservation_facts:
fact_statement = fact[]
fact_answer = .model.generate(fact_statement)
fact_answer != fact[]:
()
classClosedLoopEditor:
def__init__(self, model, max_iterations=5):
self.model = model
self.max_iterations = max_iterations
self.edit_history = []
defedit_with_feedback(self, edit_spec, learning_rate=1e-5):
"""
Iterative editing with feedback on success.
"""for iteration inrange(self.max_iterations):
print(f"Edit iteration {iteration + 1}/{self.max_iterations}")
# Step 1: Identify neurons/parameters responsible for target fact
target_neurons = self._identify_responsible_neurons(edit_spec.target_fact)
# Step 2: Compute edit direction
edit_direction = self._compute_edit_direction(
edit_spec.target_fact,
edit_spec.desired_output,
target_neurons
)
# Step 3: Apply updateself._apply_constrained_update(edit_direction, learning_rate)
# Step 4: Evaluate success
success_score = self._evaluate_edit(edit_spec)
print(f" Success score: {success_score:.2%}")
# Step 5: Check preservation (side effects)
preservation_loss = self._evaluate_preservation(edit_spec.preservation_facts)
print(f" Preservation loss: {preservation_loss:.4f}")
# Record iterationself.edit_history.append({
'iteration': iteration,
'success': success_score,
'preservation': preservation_loss,
'learning_rate': learning_rate
})
# Early stopping: success thresholdif success_score > 0.9and preservation_loss < 0.05:
print(f"✓ Edit successful at iteration {iteration + 1}")
returnTrue# Adapt learning rateif success_score < 0.3:
learning_rate *= 0.5# Reduce if overshootingprint(f"✗ Edit incomplete after {self.max_iterations} iterations")
returnFalsedef_identify_responsible_neurons(self, target_fact):
"""
Find which neurons encode target fact via gradient analysis.
"""# Simplified: neurons with highest gradient w.r.t. target fact
gradients = self.model.compute_gradients(target_fact)
# Select top-k neurons
top_k = 100
responsible_idx = gradients.argsort()[-top_k:]
return responsible_idx
def_compute_edit_direction(self, target_fact, desired_output, target_neurons):
"""
Compute direction to update neuron values.
"""
current_output = self.model.generate(target_fact)
# Difference vector: what needs to change# Simplified: use gradient of loss w.r.t. neuron outputs
loss = self.model.compute_loss(target_fact, desired_output)
edit_direction = torch.autograd.grad(loss, target_neurons, allow_unused=True)
return edit_direction
def_apply_constrained_update(self, edit_direction, learning_rate):
"""
Apply update with locality constraints.
"""for param_idx, grad inenumerate(edit_direction):
if grad isnotNone:
# Update with learning rate
param = self.model.get_parameter(param_idx)
# Constraint: limit update magnitude (locality guard)
max_update = 0.01 * torch.abs(param.data).mean()
grad = torch.clamp(grad, -max_update, max_update)
param.data -= learning_rate * grad
def_evaluate_edit(self, edit_spec):
"""
Measure how well edit succeeded.
"""
output = self.model.generate(edit_spec.target_fact)
success = 1.0if output == edit_spec.desired_output else0.0return success
def_evaluate_preservation(self, preservation_facts):
"""
Measure knowledge preservation (side effects).
"""
total_loss = 0for fact in preservation_facts:
output = self.model.generate(fact['statement'])
loss = 0.0if output == fact['expected'] else1.0
total_loss += loss
return total_loss / max(len(preservation_facts), 1)
3. Dynamic Memory and Knowledge Fusion
Track edits and periodically fuse to prevent drift.
classDynamicMemoryManager:
def__init__(self, model, fusion_interval=10):
self.model = model
self.edits = [] # Track all editsself.edit_counter = 0self.fusion_interval = fusion_interval
self.knowledge_buffer = {} # Snapshot of model knowledgedefrecord_edit(self, edit_spec, success):
"""
Record completed edit in memory.
"""self.edits.append({
'spec': edit_spec,
'success': success,
'timestamp': self.edit_counter
})
self.edit_counter += 1# Periodic knowledge fusionifself.edit_counter % self.fusion_interval == 0:
self.fuse_edits()
deffuse_edits(self):
"""
Integrate accumulated edits to prevent drift.
Resets baseline and re-applies edits cleanly.
"""print(f"Fusing {len(self.edits)} accumulated edits...")
# Snapshot current state
pre_fusion_state = {
name: param.detach().clone()
for name, param inself.model.named_parameters()
}
# Reapply successful edits
successful_edits = [e for e inself.edits if e['success']]
for edit in successful_edits:
# Reapply edit
editor = ClosedLoopEditor(self.model, max_iterations=2)
editor.edit_with_feedback(edit['spec'])
# Verify no drift
drift = self._compute_parameter_drift(pre_fusion_state)
print(f"Post-fusion parameter drift: {drift:.4f}")
def_compute_parameter_drift(self, reference_state):
"""
Measure divergence from reference state.
"""
total_drift = 0for (ref_name, ref_param), (name, param) inzip(
reference_state.items(),
self.model.named_parameters()
):
if ref_name == name:
drift = torch.norm(param.data - ref_param) / (torch.norm(ref_param) + 1e-8)
total_drift += drift.item()
return total_drift / len(reference_state)
4. Sequential Editing Without Catastrophic Forgetting
Enable multiple edits with conflict detection.
classSequentialEditManager:
def__init__(self, model):
self.model = model
self.memory_manager = DynamicMemoryManager(model)
self.editor = ClosedLoopEditor(model)
self.edit_graph = {} # Track edit dependencies/conflictsdefbatch_edits(self, edit_specs, detect_conflicts=True):
"""
Apply multiple edits sequentially with conflict management.
"""
results = []
for idx, edit_spec inenumerate(edit_specs):
print(f"\nEdit {idx+1}/{len(edit_specs)}: {edit_spec.target_fact}")
# Check for conflicts with prior editsif detect_conflicts:
conflicts = self._detect_conflicts(edit_spec, results)
if conflicts:
print(f" ⚠ Potential conflicts detected: {conflicts}")
# Apply edit
success = self.editor.edit_with_feedback(edit_spec)
results.append({'edit': edit_spec, 'success': success})
# Record in memoryself.memory_manager.record_edit(edit_spec, success)
# Track in edit graphself.edit_graph[idx] = {
'spec': edit_spec,
'conflicts': conflicts if detect_conflicts else []
}
return results
def_detect_conflicts(self, edit_spec, prior_edits):
"""
Identify if new edit conflicts with prior edits.
"""
conflicts = []
for prior_result in prior_edits:
prior_spec = prior_result['edit']
# Simple conflict: edits affect overlapping factsifself._edits_overlap(edit_spec, prior_spec):
conflicts.append(f"Overlaps with: {prior_spec.target_fact}")
return conflicts
def_edits_overlap(self, spec1, spec2):
"""
Check if two edits potentially conflict.
"""# Overlap if target facts share knowledge regions
overlap_score = self.model.compute_semantic_similarity(
spec1.target_fact,
spec2.target_fact
)
return overlap_score > 0.5
Practical Guidance
Edit Granularity: Single, focused facts edit better than broad generalizations. "Eiffel Tower is in Paris" edits better than "All facts about Eiffel Tower."
Preservation Facts: Specify 3-5 related facts that must be preserved. More preservation facts slow editing but improve robustness.
Learning Rate: Start at 1e-5, reduce if overshooting. Adaptive rates based on success score work well.
Fusion Interval: Every 5-10 edits fuse to prevent accumulated drift. More frequent fusion is safer but slower.
When to Use / When NOT to Use
Use When:
Correcting factual errors post-deployment
Updating knowledge as new information arrives
Models need graceful degradation when edits fail
Sequential edits on diverse facts
NOT For:
Bulk retraining (dedicated fine-tuning more efficient)
This skill synthesizes findings from "REPAIR: Robust Editing via Progressive Adaptive Intervention and Reintegration" (arXiv:2510.01879). Closed-loop feedback and knowledge fusion enable stable, low-cost editing.
