---
id: wiki-2026-0508-circuit-discovery
title: Circuit Discovery (Mechanistic Interpretability)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [회로 발견, mechanistic interpretability, MI, induction head, activation patching, path patching, ACDC, sparse autoencoder]
duplicate_of: none
source_trust_level: A
confidence_score: 0.88
verification_status: applied
tags: [interpretability, mechanistic, circuits, activation-patching, induction-heads, sparse-autoencoder, anthropic, transformer]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: TransformerLens / nnsight / SAE
---

# Circuit Discovery

## 📌 한 줄 통찰
> **"매 LLM 내부 의 algorithm 의 reverse engineer"**. 매 black box 의 X — 매 specific neuron + attention head 의 연결망 의 algorithmic role. 매 modern Anthropic / OpenAI alignment 의 핵심 capability. 매 SAE (Sparse Autoencoder) 의 2024 의 breakthrough.

## 📖 핵심

### 매 mechanistic interpretability (MI)
- 매 neural network 의 specific algorithm 의 identify.
- 매 saliency map 의 X — 매 mechanism.
- 매 hypothesis-driven.

### 매 핵심 기법

#### Activation Patching
- 매 specific position 의 activation 의 swap.
- 매 output change 의 measure.
- 매 causal effect 의 identify.

#### Path Patching
- 매 specific connection 의 isolate.
- 매 layer-to-layer information flow.

#### Logit Lens
- 매 each layer 의 output 의 unembed.
- 매 layer 의 prediction evolution.

#### Induction Heads (Olsson et al. 2022)
- 매 [A][B] ... [A] → [B] 의 pattern 의 copy.
- 매 in-context learning 의 base.
- 매 layer 2-3 의 typical.

#### IOI (Indirect Object Identification)
- "When Mary and John went to the store, John gave a drink to ___" → Mary.
- 매 26-head circuit 의 mapped (Wang et al. 2022).

#### Automated Circuit Discovery (ACDC, Conmy et al. 2023)
- 매 algorithmic search.
- 매 pruning of insignificant edges.

#### Sparse Autoencoder (SAE, 2023-2024)
- 매 monosemantic feature 의 decompose.
- 매 polysemantic neuron 의 split.
- 매 Anthropic Sonnet/Sonnet 3.5 의 large-scale.

### 매 Anthropic 의 milestone
- "A Mathematical Framework for Transformer Circuits" (2021).
- "In-context Learning and Induction Heads" (2022).
- "Toy Models of Superposition" (2022).
- "Towards Monosemanticity" (2023).
- "Scaling Monosemanticity" (2024) — 매 Claude Sonnet의 30M feature.

### 매 응용
1. **Alignment**: 매 deceptive circuit 의 detect.
2. **Bias**: 매 specific feature 의 identify.
3. **Steering**: 매 feature 의 amplify / suppress.
4. **Debugging**: 매 hallucination cause.
5. **Trust**: 매 capability evaluation.
6. **Capability discovery**: 매 emergent.

### 매 limitation
- 매 scaling: 매 large model 의 expensive.
- 매 polysemanticity: 매 single neuron 의 many concept.
- 매 superposition: 매 high-dim feature 의 low-dim 의 squeeze.
- 매 generalization: 매 single task circuit 의 wider behavior?

### 매 modern tool
- **TransformerLens** (Neel Nanda).
- **nnsight** (NDIF, Bau lab).
- **SAELens**: 매 SAE training.
- **Neuropedia**: 매 feature catalog.
- **Pyvene**: 매 intervention.
- **Inseq**: 매 attribution.

## 💻 패턴

### TransformerLens (basic)
```python
import transformer_lens
import torch

model = transformer_lens.HookedTransformer.from_pretrained('gpt2-small')

# 매 hook
def hook_print(activation, hook):
    print(f'{hook.name}: shape {activation.shape}')

# 매 forward with hook
text = 'When Mary and John went to the store, John gave a drink to'
logits, cache = model.run_with_cache(text, return_type='logits')
print(cache['blocks.5.attn.hook_z'].shape)  # head outputs
```

### Activation Patching
```python
def activation_patching(model, clean_input, corrupt_input, position, layer, head):
    """매 corrupt activation 의 clean run 의 patch."""
    # 매 1. cache corrupt
    _, corrupt_cache = model.run_with_cache(corrupt_input)
    
    # 매 2. patch into clean run
    def patch_hook(activation, hook):
        activation[:, position, head] = corrupt_cache[hook.name][:, position, head]
        return activation
    
    patched_logits = model.run_with_hooks(
        clean_input,
        fwd_hooks=[(f'blocks.{layer}.attn.hook_z', patch_hook)],
    )
    return patched_logits

# 매 effect 의 measure
clean_logits = model(clean_input)
patched = activation_patching(model, clean_input, corrupt_input, pos=10, layer=5, head=3)
effect = (clean_logits - patched)[0, -1, target_token]
```

### Induction head detection
```python
def detect_induction_head(model, text):
    """매 [A][B]...[A] → [B] 의 head 의 identify."""
    # 매 random repeating sequence
    rep = torch.randint(0, model.cfg.d_vocab, (1, 50))
    rep = torch.cat([rep, rep], dim=1)  # 매 50 + 매 같은 50
    
    _, cache = model.run_with_cache(rep)
    
    induction_scores = {}
    for layer in range(model.cfg.n_layers):
        for head in range(model.cfg.n_heads):
            # 매 attention from later half 의 매 earlier half + 1 의 score
            attn = cache[f'blocks.{layer}.attn.hook_pattern'][0, head]
            # 매 diagonal 의 -49 만큼 의 average
            score = attn.diagonal(offset=-49).mean()
            induction_scores[(layer, head)] = score.item()
    
    return sorted(induction_scores.items(), key=lambda x: -x[1])[:5]
```

### Logit Lens
```python
def logit_lens(model, text):
    _, cache = model.run_with_cache(text)
    
    layer_predictions = []
    for layer in range(model.cfg.n_layers):
        residual = cache[f'blocks.{layer}.hook_resid_post']
        # 매 unembed
        logits = model.unembed(model.ln_final(residual))
        top_token = logits[0, -1].argmax().item()
        layer_predictions.append((layer, model.tokenizer.decode([top_token])))
    
    return layer_predictions
```

### ACDC (Automated Circuit Discovery)
```python
# 매 simplified ACDC
def acdc_prune(model, task_data, threshold=0.01):
    """매 edge 의 effect 가 threshold 이하 의 prune."""
    edges = list_all_edges(model)
    kept = []
    
    baseline_loss = evaluate(model, task_data)
    
    for edge in edges:
        # 매 edge 의 zero-ablate
        with ablated_edge(model, edge):
            ablated_loss = evaluate(model, task_data)
        
        effect = ablated_loss - baseline_loss
        if effect > threshold:
            kept.append((edge, effect))
    
    return sorted(kept, key=lambda x: -x[1])
```

### Sparse Autoencoder (SAE, 2024)
```python
import torch.nn as nn

class SAE(nn.Module):
    """매 dictionary learning 의 transformer activation."""
    def __init__(self, d_in=4096, d_dict=131072, l1_coeff=1e-3):
        super().__init__()
        self.encoder = nn.Linear(d_in, d_dict)
        self.decoder = nn.Linear(d_dict, d_in, bias=False)
        self.l1_coeff = l1_coeff
    
    def forward(self, x):
        features = F.relu(self.encoder(x))
        x_hat = self.decoder(features)
        
        recon_loss = (x - x_hat).pow(2).mean()
        l1_loss = features.abs().mean()  # 매 sparsity
        
        return x_hat, features, recon_loss + self.l1_coeff * l1_loss

# 매 train on 매 model activation cache
sae = SAE(d_in=4096, d_dict=4096*32)
for batch in activation_loader:
    x_hat, features, loss = sae(batch)
    loss.backward()
    optimizer.step()

# 매 features 의 monosemantic 의 inspect.
```

### Steering with SAE feature
```python
def steer(model, sae, text, feature_idx, scale=5.0):
    """매 specific SAE feature 의 amplify."""
    def hook(activation, hook):
        # 매 SAE encode → 매 modify → 매 decode
        features = F.relu(sae.encoder(activation))
        features[:, :, feature_idx] += scale
        return sae.decoder(features)
    
    return model.run_with_hooks(text, fwd_hooks=[('blocks.16.hook_resid_post', hook)])
```

### Feature visualization (Neuronpedia-style)
```python
def find_max_activating(sae, feature_idx, dataset, top_k=10):
    """매 feature 의 max-activate 매 input."""
    activations = []
    for text in dataset:
        with torch.no_grad():
            _, cache = model.run_with_cache(text)
            f = F.relu(sae.encoder(cache['blocks.16.hook_resid_post']))
            max_act = f[0, :, feature_idx].max().item()
            activations.append((text, max_act))
    
    return sorted(activations, key=lambda x: -x[1])[:top_k]
```

## 🤔 결정 기준
| 목적 | Method |
|---|---|
| Behavior cause | Activation patching |
| Information flow | Path patching |
| Layer evolution | Logit lens |
| In-context learning | Induction head analysis |
| Feature discovery | SAE |
| Algorithmic search | ACDC |
| Steering | SAE feature manipulation |
| Capability eval | Probe |

**기본값**: TransformerLens + activation patching 의 baseline. 매 large model = SAE.

## 🔗 Graph
- 부모: [[Interpretability]] · [[AI Safety]] · [[Mechanistic-Interpretability]]
- 변형: [[Activation-Patching]] · [[Path-Patching]] · [[ACDC]]
- 응용: [[Steering]] · [[Induction-Head]]
- Adjacent: [[Anthropic]] · [[AI_Safety_and_Alignment|AI-Alignment]]

## 🤖 LLM 활용
**언제**: 매 alignment research. 매 model debugging. 매 capability discovery. 매 steering. 매 trust evaluation.
**언제 X**: 매 specific task fine-tune. 매 production inference (research domain).

## ❌ 안티패턴
- **Correlation 만**: 매 causal patching 의 X.
- **Single circuit 의 generalize**: 매 narrow finding.
- **No baseline**: 매 random 의 effect 의 detect X.
- **Polysemanticity ignore**: 매 single neuron 의 many concept.
- **No SAE on large model**: 매 superposition 의 miss.
- **Manual only at scale**: 매 100B+ 의 ACDC 의 필요.

## 🧪 검증 / 중복
- Verified (Anthropic transformer-circuits.pub, Olsson induction heads, Wang IOI, ACDC paper).
- 신뢰도 A.
- Related: [[AI_Safety_and_Alignment|AI-Alignment]] · [[AI Safety]] · [[Anthropic-Principle]] · [[Sparse-Autoencoder]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — MI techniques + SAE + 매 TransformerLens / ACDC / steering code |