File size: 7,325 Bytes
3646c00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""J-SPACE analyzer — advanced (after Anthropic 'A global workspace in language models', Jul 2026).

Beyond the logit-lens primitive. Implements the real machinery to identify & score a model's reasoning
'global workspace', as a DATA-FREE genetic-merge fitness:

  1. FUTURE-INFLUENCE JACOBIAN (the actual J-lens signal, not next-token logit-lens):
     G_l = d(total future NLL)/d(hidden_l)  [seq,hid].  Row i = the direction in which position i's
     activation steers ALL future outputs (causal). This is "the pattern that makes the model likely to
     say things later" — the workspace read/write signal.

  2. J-SPACE SUBSPACE via SVD of G_l per layer:
       - energy_conc   : fraction of future-influence energy in the top-k singular directions.
       - eff_rank      : exp(entropy of normalized singular values) — a LOW effective rank = a small,
                         coherent workspace (<10% signature). Reported as frac = eff_rank/hidden.
       - The top singular vectors ARE the J-space directions at that layer.

  3. PLANNING via the lens: how EARLY the model resolves future tokens (front-loaded competence),
     base=layer1 (skip embeddings). A rich workspace reasons in mid layers, not just at the end.

  4. CAUSAL INTERVENTION: project the J-space subspace OUT at a mid layer (hook) and measure the
     reasoning-NLL increase vs a random-subspace-of-equal-rank control. If ablating the workspace hurts
     reasoning MORE than a random subspace, the identified subspace is causally the reasoning substrate.

  5. Composite FITNESS = f(front_load, workspace concentration, causal necessity).  All data-free
     (fixed reasoning probe set). CPU / float32.
"""
import sys, math, torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL = "/Users/kikocisneros/coco_ppl/attn_longctx/model"; DEV = "cpu"
tok = AutoTokenizer.from_pretrained(MODEL)
def load(): return AutoModelForCausalLM.from_pretrained(MODEL, dtype=torch.float32).to(DEV).eval()

PROBES = [
 "Q: If a train goes 60 km in 1 hour, how far in 3 hours? 60 times 3 equals",
 "Q: Ana has 5 apples, buys 7 more, gives 3 away. 5 plus 7 is 12, minus 3 is",
 "Q: A spider has 8 legs. How many legs do 3 spiders have? 8 times 3 is",
 "The opposite of hot is cold. The opposite of up is down. The opposite of fast is",
 "Paris is to France as Rome is to",
]

# ── hooks to capture per-layer hidden states with grad, and to intervene ──
class Tap:
    def __init__(self, model):
        self.model = model; self.L = model.config.num_hidden_layers
        self.h = {}; self.handles = []; self.intervene = None
    def __enter__(self):
        for l, lyr in enumerate(self.model.model.layers):
            self.handles.append(lyr.register_forward_hook(self._mk(l)))
        return self
    def _mk(self, l):
        def hook(m, i, o):
            hs = o[0] if isinstance(o, tuple) else o
            if self.intervene and self.intervene[0] == l:            # project OUT a subspace at this layer
                Q = self.intervene[1]                                 # [hid,k] orthonormal
                hs2 = hs - (hs @ Q) @ Q.transpose(-1, -2)
                hs = hs2
                return (hs2,) + o[1:] if isinstance(o, tuple) else hs2
            if hs.requires_grad:                                       # skip under torch.no_grad (reasoning_nll)
                hs.retain_grad(); self.h[l] = hs
        return hook
    def __exit__(self, *a):
        for h in self.handles: h.remove()

def influence_jacobian(model, ids):
    """G_l = d(total next-token NLL)/d(hidden_l) for each layer -> {l: [seq,hid]} (future-influence dirs)."""
    with Tap(model) as tap:
        out = model(ids)
        loss = torch.nn.functional.cross_entropy(out.logits[0, :-1], ids[0, 1:], reduction="sum")
        loss.backward()
        G = {l: tap.h[l].grad[0].detach().clone() for l in range(tap.L) if tap.h[l].grad is not None}
    model.zero_grad(set_to_none=True)
    return G

def subspace_stats(G_l, k=8):
    """SVD of the future-influence matrix [seq,hid] -> workspace concentration + effective rank + top dirs."""
    U, S, Vh = torch.linalg.svd(G_l.float(), full_matrices=False)       # V rows = influence directions
    s = S / (S.sum() + 1e-9)
    energy_topk = s[:k].sum().item()
    ent = -(s * (s + 1e-12).log()).sum().item()
    eff_rank = math.exp(ent)
    return {"energy_topk": energy_topk, "eff_rank_frac": eff_rank / G_l.shape[1],
            "dirs": Vh[:k].transpose(0, 1)}                             # [hid,k] top-k directions

def frontload_and_final(model, ids):
    L = model.config.num_hidden_layers; norm, head = model.model.norm, model.lm_head
    with torch.no_grad():
        hs = model(ids, output_hidden_states=True).hidden_states
    tgt = ids[0, 1:]; lp = []
    for l in range(L+1):
        ll = head(norm(hs[l][0][:-1])).float()
        lp.append(torch.log_softmax(ll, -1).gather(1, tgt.unsqueeze(1)).squeeze(1).mean().item())
    base, final = lp[1], lp[-1]
    c = [max(0., min(1.2, (v - base) / (final - base + 1e-6))) for v in lp]
    return sum(c[1:]) / L, final                                        # frontload AUC, final competence

@torch.no_grad()
def reasoning_nll(model, ids, intervene=None):
    with Tap(model) as tap:
        tap.intervene = intervene
        out = model(ids)
        return torch.nn.functional.cross_entropy(out.logits[0, :-1], ids[0, 1:]).item()

def causal_necessity(model, ids, layer, k=8):
    """Ablate the J-space subspace at `layer` vs a random subspace of equal rank -> reasoning-NLL delta."""
    G = influence_jacobian(model, ids)
    if layer not in G: return 0.0
    Q = subspace_stats(G[layer], k)["dirs"]                            # [hid,k] workspace dirs
    Qr, _ = torch.linalg.qr(torch.randn_like(Q))                       # random orthonormal control
    base = reasoning_nll(model, ids)
    js = reasoning_nll(model, ids, intervene=(layer, Q))
    rnd = reasoning_nll(model, ids, intervene=(layer, Qr))
    return (js - base) - (rnd - base)                                  # extra damage from ablating the workspace

def fitness(model, verbose=False):
    L = model.config.num_hidden_layers; mid = L // 2
    fl, fc, et, er, cn = [], [], [], [], []
    for p in PROBES:
        ids = tok(p, return_tensors="pt").input_ids.to(DEV)
        a, b = frontload_and_final(model, ids); fl.append(a); fc.append(b)
        G = influence_jacobian(model, ids)
        st = subspace_stats(G[mid]); et.append(st["energy_topk"]); er.append(st["eff_rank_frac"])
        cn.append(causal_necessity(model, ids, mid))
    r = {"frontload": sum(fl)/len(fl), "final_comp": sum(fc)/len(fc),
         "ws_energy_topk": sum(et)/len(et), "ws_effrank_frac": sum(er)/len(er),
         "causal_necessity": sum(cn)/len(cn)}
    # composite: reward front-loading + a concentrated workspace (high top-k energy, low eff-rank) + causal necessity
    r["JSPACE_FITNESS"] = (r["frontload"] + r["ws_energy_topk"] - r["ws_effrank_frac"]
                           + 0.1 * r["causal_necessity"] + 0.1 * r["final_comp"])
    if verbose:
        for k, v in r.items(): print(f"  {k:18s} = {v:+.4f}")
    return r

if __name__ == "__main__":
    print("=== J-SPACE (advanced) on clean 0.5B ==="); fitness(load(), verbose=True)
    print("JSPACE_ADV_DONE")