File size: 8,927 Bytes
e53f10b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
"""
Direction extraction (rewritten Apr 2026).

Two versions kept:
  v1_raw         — single mean-diff direction (D,)
  v_pca_subspace — top-k subspace from inter-class scatter PCA (k, D)

Earlier v2_ortho_general / v3_ortho_crossdim / v4_pca were removed because:
  - v2/v3 had cosine > 0.95 to v1 in v1 results (no signal added)
  - v4 was conceptually wrong (PCA over all decision points, not over the
    plan-vs-exec contrast)

The new v_pca_subspace performs PCA on the **inter-class scatter**:
  S_b = sum_c (mu_c - mu) (mu_c - mu)^T
where c ∈ {plan, exec}. Top-k eigenvectors form a k-D subspace capturing
the directions of largest plan-vs-exec variation.

Steering with this subspace:
  h_new = h - (1 - alpha) · Q^T Q · h
where Q ∈ R^(k × D) is row-orthonormal.
"""
import torch
import numpy as np
from typing import Dict, List, Optional


def _safe_normalize(v: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
    n = v.norm(dim=-1, keepdim=True).clamp(min=eps)
    return v / n


# ============================================================
# v1_raw: single direction mean-diff
# ============================================================
def compute_mean_diff(
    plan_acts_per_layer: Dict[int, torch.Tensor],
    exec_acts_per_layer: Dict[int, torch.Tensor],
) -> Dict[int, torch.Tensor]:
    """
    v1: raw mean-diff per layer.

    Returns {layer_id: (D,) float32 direction (NOT normalized)}.
    """
    directions = {}
    for li in plan_acts_per_layer:
        h_plan = plan_acts_per_layer[li].to(torch.float32)
        h_exec = exec_acts_per_layer[li].to(torch.float32)
        if h_plan.shape[0] == 0 or h_exec.shape[0] == 0:
            directions[li] = torch.zeros(h_plan.shape[1] if h_plan.shape[0] else
                                          (h_exec.shape[1] if h_exec.shape[0] else 0))
            continue
        mu_plan = h_plan.mean(dim=0)
        mu_exec = h_exec.mean(dim=0)
        directions[li] = mu_plan - mu_exec
    return directions


# ============================================================
# v_pca_subspace: top-k PCA on plan-vs-exec inter-class structure
# ============================================================
def compute_pca_subspace(
    plan_acts_per_layer: Dict[int, torch.Tensor],
    exec_acts_per_layer: Dict[int, torch.Tensor],
    k: int = 3,
) -> Dict[int, torch.Tensor]:
    """
    For each layer, compute a top-k subspace basis Q ∈ R^(k × D) capturing the
    directions of largest variation between plan and exec activations.

    Approach: build a balanced "contrast set" — for each plan token, sample
    the same number of exec tokens. Center each class to its own mean, then
    take PCA on the centered union (variance within both classes around
    their respective means; the leading eigenvectors capture the axes along
    which the two classes most differ).

    More principled than fitting plain PCA on union — this version emphasises
    the inter-class scatter direction.

    Returns:
        {layer_id: (k, D) row-orthonormal basis}
    """
    bases = {}
    for li in plan_acts_per_layer:
        H_plan = plan_acts_per_layer[li].to(torch.float32)
        H_exec = exec_acts_per_layer[li].to(torch.float32)
        if H_plan.shape[0] == 0 or H_exec.shape[0] == 0:
            D = H_plan.shape[1] if H_plan.shape[0] else (H_exec.shape[1] if H_exec.shape[0] else 0)
            bases[li] = torch.zeros(0, D)
            continue

        D = H_plan.shape[1]
        # Balance classes — sample same number from larger class
        n_plan, n_exec = H_plan.shape[0], H_exec.shape[0]
        n_use = min(n_plan, n_exec)
        if n_plan > n_use:
            idx = torch.randperm(n_plan)[:n_use]
            H_plan = H_plan[idx]
        if n_exec > n_use:
            idx = torch.randperm(n_exec)[:n_use]
            H_exec = H_exec[idx]

        # Class means
        mu_plan = H_plan.mean(dim=0)   # (D,)
        mu_exec = H_exec.mean(dim=0)

        # Inter-class scatter contribution: signed displacement of each sample
        # from the OPPOSITE class mean. This captures inter-class variance.
        # Fisher-style: project each sample onto axis spanning the two means,
        # but extract a k-D subspace via eigendecomposition.
        #
        # Build M = [H_plan - mu_exec ; H_exec - mu_plan] (2N, D)
        # then Cov(M) has top eigenvectors aligned with directions of
        # plan-vs-exec separation (not with within-class noise).
        M_plan = H_plan - mu_exec.unsqueeze(0)   # (n_use, D)
        M_exec = H_exec - mu_plan.unsqueeze(0)   # (n_use, D)
        M = torch.cat([M_plan, M_exec], dim=0)   # (2N, D)
        # Center M overall
        M = M - M.mean(dim=0, keepdim=True)

        # SVD: M = U S V^T, top rows of V^T are the desired basis
        n_comp = min(k, M.shape[0] - 1, D)
        if n_comp <= 0:
            bases[li] = torch.zeros(0, D)
            continue

        try:
            U, S, Vt = torch.linalg.svd(M, full_matrices=False)
            Q = Vt[:n_comp]   # (n_comp, D)
        except Exception:
            cov = (M.T @ M) / max(M.shape[0] - 1, 1)
            eigvals, eigvecs = torch.linalg.eigh(cov)
            idx = torch.argsort(eigvals, descending=True)
            Q = eigvecs[:, idx[:n_comp]].T

        # Row-orthonormalize defensively (already orthonormal from SVD, but...)
        Q = _row_orthonormalize(Q)
        bases[li] = Q

    return bases


def _row_orthonormalize(Q: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
    """Gram-Schmidt row-orthonormalization."""
    if Q.shape[0] == 0:
        return Q
    out = []
    for i in range(Q.shape[0]):
        v = Q[i].clone()
        for u in out:
            v = v - (v @ u) * u
        n = v.norm()
        if n < eps:
            continue
        out.append(v / n)
    if not out:
        return torch.zeros(0, Q.shape[1])
    return torch.stack(out, dim=0)


# ============================================================
# Normalization
# ============================================================
def normalize_directions(
    directions: Dict[int, torch.Tensor],
) -> Dict[int, torch.Tensor]:
    """Return unit vectors. Works for (D,) tensors (single dir).
    For (k, D) bases, returns row-orthonormal (already from SVD)."""
    out = {}
    for li, w in directions.items():
        w32 = w.to(torch.float32)
        if w32.dim() == 1:
            if w32.norm() < 1e-8:
                out[li] = w32
            else:
                out[li] = _safe_normalize(w32)
        elif w32.dim() == 2:
            out[li] = _row_orthonormalize(w32)
        else:
            out[li] = w32
    return out


# ============================================================
# Save / load
# ============================================================
def save_directions(directions: Dict[int, torch.Tensor], path):
    torch.save({str(li): w for li, w in directions.items()}, path)


def load_directions(path) -> Dict[int, torch.Tensor]:
    raw = torch.load(path, map_location="cpu")
    return {int(k): v for k, v in raw.items()}


# ============================================================
# Cosine analysis
# ============================================================
def compute_cosine_similarity_matrix(
    dirs_dict: Dict[str, Dict[int, torch.Tensor]]
) -> Dict[str, Dict[int, float]]:
    """
    For (D,) directions: cosine between unit vectors.
    For (k, D) bases: principal angle (smallest angle between subspaces).

    Returns {(v1, v2): {layer: cos}} per layer.
    """
    versions = list(dirs_dict.keys())
    out = {}
    for i, v1 in enumerate(versions):
        for v2 in versions[i:]:
            key = f"{v1}__VS__{v2}"
            per_layer = {}
            for li in dirs_dict[v1]:
                if li not in dirs_dict[v2]:
                    continue
                a = dirs_dict[v1][li].to(torch.float32)
                b = dirs_dict[v2][li].to(torch.float32)
                per_layer[li] = _subspace_cosine(a, b)
            out[key] = per_layer
    return out


def _subspace_cosine(a: torch.Tensor, b: torch.Tensor) -> float:
    """
    Cosine for (D,) directions or principal-angle cosine for (k,D) bases.
    """
    if a.numel() == 0 or b.numel() == 0:
        return 0.0
    if a.dim() == 1 and b.dim() == 1:
        if a.norm() < 1e-8 or b.norm() < 1e-8:
            return 0.0
        return float((a @ b) / (a.norm() * b.norm()))
    # Subspace case: largest singular value of A^T B (where A, B are row-orthonormal)
    A = a if a.dim() == 2 else a.unsqueeze(0)
    B = b if b.dim() == 2 else b.unsqueeze(0)
    if A.shape[0] == 0 or B.shape[0] == 0:
        return 0.0
    # row-normalize A and B in case
    A = _row_orthonormalize(A)
    B = _row_orthonormalize(B)
    if A.shape[0] == 0 or B.shape[0] == 0:
        return 0.0
    M = A @ B.T  # (k_a, k_b)
    s = torch.linalg.svdvals(M)
    return float(s[0]) if s.numel() > 0 else 0.0