A neural field can encode the same image with different weights. A reader that recognizes one such representation may therefore rely on parameter coordinates rather than the visual function. We study this distinction through function-preserving transformations, measuring recognition, prediction disagreement, and numerical fidelity together. Across shape, digit, and natural-image fields, high accuracy on the original weights coexists with substantial sensitivity to equivalent representations. On 10,000 official CIFAR10 test origins, one-period hidden-bias shifts change 7.3% of a released HiddenProbe configuration's predictions while barely changing accuracy; three-period shifts change 22.0% and reduce accuracy from 57.8% to 54.4%. Output-only ProbeGen remains unchanged. These effects are smaller than in a separate adapted-reader study, where matched phase and sign controls remove the corresponding sensitivities. Controlled readout studies show why invariance is insufficient: complete canonical hidden features reach 38.7% on MNIST, whereas explicitly composing those activations, outgoing weights, and output bias into field outputs reaches 81.5%. The information is present, but the learned readout does not recover it as effectively. A separate 720-response study of language models reading numerical weights shows why stochastic audits require repeats on both inputs: observed accuracy losses need not yield rejection of response-law equality. Together, these results distinguish invariance, useful visual information, and the evidence required to evaluate stochastic readers.
Figures & tables
Figure 1: One visual function, multiple weight inputs. A verified transformation changes the parameters while preserving the field. Paired predictions reveal coordinate dependence; correctness measures retained visual information. Stochastic readers additionally require fresh responses to each unchanged input.
Shapes
MNIST
Reader
Raw
Joint
Disagree
Raw
Joint
Disagree
Flat ridge
32.5
20.0
75.0
13.5
7.0
89.5
Stats ridge
96.7
24.2
75.0
33.0
11.5
89.0
Norm+Stats
97.5
97.5
0.0
33.5
33.5
0.0
Canon+Flat
71.7
71.7
0.0
33.5
33.5
0.0
Probe ridge
100.0
100.0
0.0
74.0
74.0
0.0
Table 1: Recognition and coordinate sensitivity on ReLU fields. Accuracy and raw/joint disagreement are percentages on 120 shape and 200 digit test origins. Joint transformations combine neuron permutation and positive rescaling. The probe reader queries the represented function; other rows read parameters.
Figure 2: Adapted natural-image readers: orbit sensitivity and targeted controls. All conditions preserve the SIREN function in real arithmetic. (a) Accuracy on 400 CIFAR10 originals, averaged over three fitting bundles and three reader seeds; bars are 95% original-image bootstrap intervals. Phase wrapping removes phase sensitivity; Gram features also remove sign sensitivity. (b) Paired prediction disagreement exposes changes concealed by similar net accuracy. The existing phasor reader and output-only probes are additional invariant controls.
Reader
Raw
Phase 1
Phase 3
Sign
D1
HiddenProbe
57.8
57.6
54.4
57.9
7.3
ProbeGen
57.3
57.3
57.3
57.3
0.0
Table 2: Official CIFAR10 corpus audit. Accuracy (%) on 10,000 test origins, mean of 3 independently trained readers per method. D1 is Phase-1 prediction disagreement (%). Targets/actions use FP64; readouts use FP32. Native clean precision and paired intervals are in the supplement.
Population
Reader features
Raw (%)
Averaged (%)
Change
Go
CIFAR10
Field outputs
24.2
24.2
+0.0
<10−9
CIFAR10
Hidden activations
16.8
18.4
+1.6
0.644
CIFAR10
Hidden preactivations
21.8
11.4
-10.4
0.774
CIFAR10
Weight statistics
17.8
10.6
-7.2
0.916
MNIST
Field outputs
81.1
81.1
+0.0
<10−9
MNIST
Hidden activations
14.3
12.5
-1.8
0.893
Table 3: Feature location changes orbit sensitivity and the effect of averaging. Means over five reader seeds on 100 independent CIFAR10 audit origins or 200 independent MNIST test origins. Seeds and transformed copies are repeated measurements, not additional independent images. Averaged predictions use the recorded 16-entry transformation design; CIFAR10 entries repeat four fixed actions. Input dimensions differ across feature types, so rows are not capacity-matched. Go is the orbit-variance share in Eq. ( 5 ); changes are percentage points.
Figure 3: Complete information still requires an effective readout. MNIST follow-up, six seeds on 200 test images; dots are seeds and bars are means. Hidden features are canonical during training and evaluation. v,c denote outgoing weights and bias. Fixed composition reconstructs field outputs before applying the output reader. Input dimensions and parameter counts differ.
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
Reader
Raw
Perm.
Scale
Joint
Dis.
Both
Flat ridge
32.5
38.3
15.8
20.0
75.0
10.8
Stats ridge
96.7
96.7
21.7
24.2
75.0
24.2
Norm+Stats
97.5
97.5
97.5
97.5
0.0
97.5
Canon+Flat
71.7
71.7
71.7
71.7
0.0
71.7
Probe ridge (81 queries)
100.0
100.0
100.0
100.0
0.0
100.0
Set
96.7
96.7
26.7
20.0
80.8
19.2
Appendix
Table 4: Shape readers: accuracy and agreement separate. Percent on 120 test origins; each joint transformation is paired with its original. Dis. is prediction disagreement and Both is both-correct rate. Probe ridge executes the field; all other rows read weights. Confidence intervals are in the supplement.
Reader
Raw
Joint
Dis.
Flat ridge
13.5
7.0
89.5
Stats ridge
33.0
11.5
89.0
Norm+Stats
33.5
33.5
0.0
Canon+Flat
33.5
33.5
0.0
Probe ridge
74.0
74.0
0.0
Set
68.5
20.0
83.5
Appendix
Table 5: MNIST image fields. Percent on 200 held-out images. Same structural controls, ten classes and width 64; language models are not evaluated on this dataset.
Reader
Raw [95% CI]
Joint [95% CI]
Change [95% CI]
C → W
W → C
Flat ridge
32.5 [24.8, 41.3]
20.0 [13.8, 28.0]
-12.5 [-22.5, -3.3]
26
11
Stats ridge
96.7 [91.7, 98.7]
24.2 [17.4, 32.6]
-72.5 [-78.3, -66.7]
87
0
Norm+Stats
97.5 [92.9, 99.1]
97.5 [92.9, 99.1]
0.0 [0.0, 0.0]
0
0
Canon+Flat
71.7 [63.0, 79.0]
71.7 [63.0, 79.0]
0.0 [0.0, 0.0]
0
0
Probe ridge (81 queries)
100.0 [96.9, 100.0]
100.0 [96.9, 100.0]
0.0 [0.0, 0.0]
0
0
Set
96.7 [91.7, 98.7]
20.0 [13.8, 28.0]
-76.7 [-80.8, -71.7]
93
1
Appendix
Table 6: Complete shape reader intervals and paired flip counts. Accuracy intervals are Wilson; change intervals are stratified origin bootstrap, in percentage points.
Reader
Raw [95% CI]
Joint [95% CI]
Change [95% CI]
C → W
W → C
Flat ridge
13.5 [9.4, 18.9]
7.0 [4.2, 11.4]
-6.5 [-12.0, -1.0]
25
12
Stats ridge
33.0 [26.9, 39.8]
11.5 [7.8, 16.7]
-21.5 [-27.5, -16.0]
56
13
Norm+Stats
33.5 [27.3, 40.3]
33.5 [27.3, 40.3]
0.0 [0.0, 0.0]
0
0
Canon+Flat
33.5 [27.3, 40.3]
33.5 [27.3, 40.3]
0.0 [0.0, 0.0]
0
0
Probe ridge (81 queries)
74.0 [67.5, 79.6]
74.0 [67.5, 79.6]
0.0 [0.0, 0.0]
0
0
Set
68.5 [61.8, 74.5]
20.0 [15.0, 26.1]
-48.5 [-56.0, -40.5]
112
15
Appendix
Table 7: Complete digit reader intervals and paired flip counts. Accuracy intervals are Wilson; change intervals are stratified origin bootstrap, in percentage points.
Dataset
Quartile
Raw
Joint
Shapes
Q1
93.3
23.3
Shapes
Q2
96.7
16.7
Shapes
Q3
96.7
20.0
Shapes
Q4
100.0
20.0
MNIST
Q1
66.0
20.0
MNIST
Q2
72.0
18.0
Appendix
Table 8: Post hoc reconstruction-stratified set-reader accuracy (percent). Quartiles are formed within each class, low to high quality. Each shape quartile has 30 origins and each MNIST quartile 50. The coordinate failure persists in well-fitted fields.
Reader
Raw accuracy (%)
Phase-one change (pp)
Phase-one D
Phase-three acc.
Sign acc.
Sign D
HiddenProbe
30.0 [27.1, 33.0]
-8.1 [-10.4, -6.0]
61.6
14.2
29.4
27.0
HiddenPhase
29.3 [26.4, 32.2]
0.0 [0.0, 0.0]
0.0
29.3
28.6
19.1
HiddenAugment
29.4 [26.5, 32.3]
0.8 [-0.1, 1.8]
18.0
27.8
29.3
13.2
HiddenGram
28.7 [25.9, 31.6]
0.0 [0.0, 0.0]
0.0
28.7
28.7
0.0
PhasorW12
24.5 [22.5, 26.6]
0.0 [0.0, 0.0]
0.0
24.5
24.5
0.0
ProbeGen
29.2 [26.3, 32.1]
0.0 [0.0, 0.0]
0.0
29.2
29.2
0.0
Appendix
Table 10: S0: fitted training source. Accuracy and phase-one change have 95% original-image intervals. Remaining entries are means over nine states. D is paired prediction disagreement (%). All permutation disagreements are zero.
Reader
Raw accuracy (%)
Phase-one change (pp)
Phase-one D
Phase-three acc.
Sign acc.
Sign D
HiddenProbe
30.1 [27.2, 33.1]
-9.1 [-11.1, -7.0]
60.2
13.9
30.0
24.7
HiddenPhase
29.4 [26.6, 32.4]
0.0 [0.0, 0.0]
0.0
29.4
29.3
18.9
HiddenAugment
30.0 [27.1, 32.9]
0.4 [-0.6, 1.4]
17.5
28.2
28.8
12.5
HiddenGram
29.7 [26.8, 32.6]
0.0 [0.0, 0.0]
0.0
29.7
29.7
0.0
PhasorW12
20.3 [18.7, 22.0]
0.0 [0.0, 0.0]
0.0
20.3
20.3
0.0
ProbeGen
29.2 [26.4, 32.1]
0.0 [0.0, 0.0]
0.0
29.2
29.2
0.0
Appendix
Table 11: S2: unseen fitting source. Accuracy and phase-one change have 95% original-image intervals. Remaining entries are means over nine states. D is paired prediction disagreement (%). All permutation disagreements are zero.
Bundle 0
Bundle 1
Bundle 2
Reader
Raw
Phase one
Raw
Phase one
Raw
Phase one
HiddenProbe
26.8
17.1
31.4
25.2
31.8
23.3
HiddenPhase
28.0
28.0
27.8
27.8
32.0
32.0
HiddenAugment
27.8
28.8
30.1
31.4
30.2
30.4
HiddenGram
28.8
28.8
29.3
29.3
27.9
27.9
PhasorW12
22.9
22.9
24.1
24.1
26.5
26.5
Appendix
Table 12: Per-bundle S0 accuracy (%). Each cell averages three reader seeds; these are three fitting bundles, not additional independent image populations. Complete per-seed values are included in the aggregate records.
Reader
Parameters
Selected epochs
HiddenProbe
447,908
30–200
HiddenPhase
447,908
40–220
HiddenAugment
447,908
20–220
HiddenGram
447,908
20–240
PhasorW12
1,875,270
20–180
ProbeGen
579,362
10–150
Appendix
Table 13: Reader capacity and model selection. Parameter counts are per state; selected-epoch ranges span nine states per architecture in the natural-image orbit study. The source-transfer pilot uses the same seven corresponding reader architectures.
Features
Train/test inputs
Accuracy (%)
Averaging change (pp)
Go
Field outputs
Original
81.25
+0.00
<10−9
Hidden activations
Original
15.17
-0.67
0.87217
Hidden + outgoing
Original
14.17
-0.50
0.87735
Hidden activations
Canonical
38.50
+0.00
0.00050
Hidden + outgoing
Canonical
40.08
+0.00
0.00054
Preactivation + outgoing
Canonical
29.75
+0.00
0.00119
Appendix
Table 14: Canonical training on MNIST. Six-seed means on the same 200 test images. Hidden profiles are unprojected here; this is separate from the main-paper feature-location study. Small nonzero Go values remain after finite-precision normalization and sorting.
Reader features
Go
Test (%)
Averaged (%)
Field outputs
<10−9
33.3
33.3
Hidden activations
0.778
8.4
8.2
Hidden preactivations
0.858
11.6
10.1
Weight statistics
0.835
18.1
11.5
Appendix
Table 15: Feature-location results on the larger held-out split. 1,600 CIFAR10 origins (1,000 train / 200 validation / 400 test), fitted twice independently. Between-fit PSNR is 34.8 dB and fit PSNR is 31.8 dB, both in RGB units. Accuracy averages five seeds on the 400 test origins, never used for selection. Head depth, hidden width, optimiser, budget, and split are shared; unequal input dimensions mean parameter counts and information differ.
Features
Input size
Parameters
Accuracy (%)
Go
Direct outputs (original)
169
111,882
81.33
<10−9
Canonical hidden
10816
2,837,514
37.75
0.00035
Hidden + v
10880
2,853,898
39.92
0.00060
Hidden + v,c
10881
2,854,154
38.67
0.00076
Fixed output composition
169
111,882
81.50
<10−9
Appendix
Table 16: Complete-input control on MNIST. Six-seed means on 200 independent test origins; canonical hidden inputs are used in training and evaluation. v,c are outgoing weights and bias. Fixed composition reconstructs the queried field outputs from the complete hidden vector. Heads share depth and hidden width; input sizes and parameter counts differ.
Reader
Training / evaluation inputs
Accuracy (%)
Disagree (%)
Stats ridge
Original / original
96.67 ± 0.00
74.38
Stats ridge
Canonical / canonical
97.50 ± 0.00
0.00
Norm+Stats
Original / original
97.50 ± 0.00
0.00
Norm+Stats
Canonical / canonical
97.50 ± 0.00
0.00
Set
Original / original
96.25 ± 0.42
80.42
Set
Canonical / canonical
95.62 ± 0.91
0.00
Appendix
Table 17: Matching training and evaluation coordinates on shapes. Four independent seed offsets; accuracy is mean ± population standard deviation across offsets. Disagreement is the mean paired prediction-change rate. Augmentation is present only for Set+Aug; the generic source protocol label does not override the reader configuration.
Reader
Gf range
Accuracy change (pp)
HiddenProbe
0.925–0.984
-1.50 to +2.25
MPGMN
0.848–0.893
-4.00 to +2.25
WeightStats
0.926–0.985
-7.25 to -1.75
ProbeGen
<10−9
+0.00 to +0.00
Grid64
<10−9
+0.00 to +0.00
Grid256
<10−9
+0.00 to +0.00
Appendix
Table 18: Orbit variation relative to fitting-source variation. Each row spans three seeds in two training regimes (six states), evaluated on 100 origins and four fitting sources. Gf uses between-fit variation in its denominator; it is not Go . The averaging change uses this eight-entry design.
Figure 4: Smoothing trades image fidelity for fitting repeatability. Two fits of each of 192 images. Agreement between fits improves as bandwidth decreases, while fidelity to the original image falls. The horizontal positions denote the four tested settings.
Band
Fit (dB)
Pair (dB)
Acc. (%)
1.00
32.09
35.00
27.0
0.50
26.06
39.87
25.7
0.35
22.98
46.29
26.7
0.25
20.95
50.80
24.3
Appendix
Table 19: Repeatable fitting and recognition are distinct outcomes. Fit is PSNR to the original image and Pair is between-fit PSNR, both in unclipped RGB units on 192 images (CPU replay). Accuracy uses only the 100 audit images, averaged over three reader seeds; the validation images used for selection are excluded. The original target is unfiltered at bandwidth 1.00.
Reader
Training
S0
S1
S2
I0
HiddenProbe
S0
25.0 (31/23/21)
25.3 (29/27/20)
23.3 (30/20/20)
24.3 (31/23/19)
HiddenProbe
S0/S1
22.3 (21/24/22)
24.0 (23/27/22)
21.7 (19/24/22)
23.3 (24/23/23)
ProbeGen
S0
25.7 (27/27/23)
26.0 (27/28/23)
26.3 (27/29/23)
26.0 (26/29/23)
ProbeGen
S0/S1
22.7 (24/20/24)
23.3 (26/20/24)
23.0 (25/20/24)
22.7 (24/20/24)
MPGMN
S0
24.7 (34/23/17)
12.7 (11/17/10)
16.3 (19/14/16)
18.0 (17/21/16)
MPGMN
S0/S1
20.7 (24/18/20)
23.0 (22/23/24)
20.0 (24/18/18)
16.7 (18/17/15)
Appendix
Table 20: Complete pilot source matrix (%). Each cell is mean accuracy over reader seeds 11/22/33, with the three individual accuracies in parentheses. All conditions use the same 100 audit images. The Dense CNN second run repeats original-image training; its differences are not an effect of source-balanced inputs.
Reader
S2 − S0
I0 − S0
Balanced gain on S2
HiddenProbe
-1.7 [-3.7,+0.3]
-0.7 [-3.7,+2.3]
-1.7 [-6.3,+3.0]
ProbeGen
+0.7 [-0.3,+2.0]
+0.3 [-1.7,+2.3]
-3.3 [-7.3,+0.7]
MPGMN
-8.3 [-15.0,-1.7]
-6.7 [-14.3,+0.7]
+3.7 [-2.3,+9.7]
Grid64
+1.0 [-0.3,+2.3]
+0.7 [-0.7,+2.0]
-0.7 [-2.7,+1.0]
Grid256
-0.7 [-1.7,+0.0]
-0.3 [-1.0,+0.0]
+1.3 [-1.7,+4.3]
WeightStats
-3.0 [-8.7,+2.7]
-3.7 [-10.0,+2.7]
+1.7 [-2.3,+5.7]
Appendix
Table 21: Paired changes (percentage points, 95% image-bootstrap intervals). Single-source rows compare fixed S0-trained readers across sources. Balanced gain compares S0/S1-trained with S0-trained readers on S2 at equal presentations and updates. Intervals condition on the fitted source bundle and trained seeds; these secondary comparisons are descriptive, without multiplicity-adjusted discovery claims.
Source
Median
10th pct.
Minimum
Below 30
S0
31.92
29.79
27.25
69
S1
31.37
29.26
26.79
118
S2
31.39
29.22
26.15
125
I0
31.70
29.65
26.89
93
Appendix
Table 22: Pilot reconstruction quality for all 600 images per source. No fits were dropped. PSNR uses RGB intensity range one, without clipping.
Model
Rule
Raw [95% CI]
Joint [95% CI]
Canon. [95% CI]
Dis.
Both
Qwen2.5 0.5B
original
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
0.0
21.7
Qwen2.5 0.5B
calibrated
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
0.0
21.7
Qwen2.5 1.5B
original
18.3 [10.6, 29.9]
16.7 [9.3, 28.0]
16.7 [9.3, 28.0]
6.7
16.7
Qwen2.5 1.5B
calibrated
21.7 [13.1, 33.6]
23.3 [14.4, 35.4]
23.3 [14.4, 35.4]
3.3
21.7
SmolLM2 0.36B
original
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
21.7 [13.1, 33.6]
1.7
21.7
SmolLM2 0.36B
calibrated
23.3 [14.4, 35.4]
23.3 [14.4, 35.4]
23.3 [14.4, 35.4]
5.0
23.3
Appendix
Table 23: Complete original and contextual-calibration results on the same 60 origins. Intervals are Wilson intervals; disagreement and both-correct are percentages.
Representation
Acc.
Changed /60
Original
21.7
0
Permutation
21.7
0
Scaling
21.7
0
Joint
21.7
0
Joint then normalize
21.7
0
Joint then sort
21.7
0
Appendix
Table 24: Serializer components, Qwen2.5 0.5B. Percent accuracy and the number of predictions changing from the original among 60 origins. Component ablations process the same joint weights. Zero changes are an observed sample result, not a population-risk certificate. The main paper reports its raw-accuracy interval.
Requested model
Condition
Correct /60
Accuracy [95% CI]
Invalid /60
Astra 6
raw
52/60
86.7 [75.8, 93.1]
0
Astra 6
raw_repeat
54/60
90.0 [79.9, 95.3]
0
Astra 6
perm
51/60
85.0 [73.9, 91.9]
0
Astra 6
joint
47/60
78.3 [66.4, 86.9]
0
Astra 6
canonical
51/60
85.0 [73.9, 91.9]
0
Astra 6
canonical_repeat
55/60
91.7 [81.9, 96.4]
0
Appendix
Table 25: Every generated-answer condition. All completed invalid outputs are included in the denominator. The two repeats are fresh calls, not copied answers. No condition is selected by its score.
Requested model
First/second
Dis.
Harm
Help
Wrong flip
Δ acc. [95% CI]
Astra 6
Raw/Perm.
9
4
3
2
-1.7 [-10.0, +6.7]
Astra 6
Raw/Joint
12
8
3
1
-8.3 [-18.3, +1.7]
Astra 6
Raw/Repeat
5
1
3
1
+3.3 [-1.7, +10.0]
Astra 6
Canon./Repeat
7
1
5
1
+6.7 [+0.0, +13.3]
Astra 6
Raw/Canon.
9
5
4
0
-1.7 [-11.7, +8.3]
Astra 6
Prior/Raw
47
0
42
5
+70.0 [+61.7, +78.3]
Appendix
Table 26: Paired generated-answer diagnostics. Counts out of 60 origins; Δ is second minus first accuracy in percentage points. Intervals are class-stratified origin-bootstrap percentiles. Any invalid member is a disagreement, even when both members are invalid. Raw/canonical and prior/raw contrasts are descriptive follow-up summaries; the full output is retained.
Requested model
Dj−Dr
95% CI
Astra 6
+11.7
[+1.7, +20.0]
Sol 6.1
+3.3
[-5.0, +11.7]
Terra 6.1
+8.3
[-5.0, +23.3]
Luna 6
+3.3
[-10.0, +18.3]
Appendix
Table 27: Joint disagreement above the identical-raw baseline. Percentage-point differences, computed within each origin before class-stratified bootstrap resampling. Negative values are possible. An interval containing zero does not establish equivalence or absence of representation sensitivity.
Figure 5: Repeated-input variation supplies the stochastic reference. Disagreement rates on the same 60 origins, with Wilson intervals. Each repeat is a fresh answer to an identical prompt. Differences between rates are evaluated as paired origin-level contrasts.
Configuration
Raw/Joint
Raw repeat/Joint
Raw/Canonical
Joint excess
Astra 6
0.2266
0.0654
1.0000
0.0654
Sol 6.1
1.0000
0.5078
0.3877
0.7266
Terra 6.1
0.0044
0.1892
1.0000
0.4049
Luna 6
0.8145
1.0000
1.0000
0.8318
Appendix
Table 28: Exact paired sensitivity checks. Two-sided, unadjusted values from the retained 60 origins. The first three columns compare accuracy; the last compares paired disagreement indicators. All comparisons are retrospective and exploratory.
Input
Accuracy (%)
Calibrated (%)
Mass (%)
raw
11.7
21.7
99.93
joint
15.0
18.3
99.93
canonical
15.0
26.7
99.93
redacted
15.0
–
99.93
textcue
100.0
–
99.99
Appendix
Table 29: The 27B candidate-score control uses the fixed 60-origin menu. Mass is the median full-vocabulary probability assigned to A–F; calibration subtracts menu-matched redacted scores.
Repeats per side
Null rejection
Alternative rejection
2
3.8 [2.8, 5.2]
7.0 [5.6, 8.8]
3
3.7 [2.7, 5.1]
11.5 [9.7, 13.6]
Appendix
Table 30: Known candidate-law simulation: rejection rate (%, 95% binomial interval) over 1,000 trials per design and law. Nominal level is 5%; this is a separate, single-test simulation.
Reader
Raw (%)
Joint (%)
Change (pp), 95% interval
Sn
pH
Astra
88.3
81.1
−7.2[−12.8,−2.2]
0.0185
0.162
Sol
86.7
79.4
−7.2[−12.8,−1.7]
0.0259
0.162
Appendix
Table 31: Recognition loss and distributional evidence in 720 fresh responses. Three raw and three joint responses for each of 60 origins per configuration. Accuracy changes have descriptive 95% origin-bootstrap intervals. Sn estimates the mean half-squared response-law distance; pH is the exact distribution-test value after the pre-specified two-test Holm adjustment.
Config.
Sn
I
p
H4
H16
Astra
0.0208
5
0.2263
0.6024
1.0000
Sol
0.0750
7
0.0087
0.0348
0.1303
Terra
-0.0417
12
0.9615
0.9615
1.0000
Luna
0.0375
16
0.2008
0.6024
1.0000
Appendix
Table 32: Raw/canonical distribution comparison. Retrospective analysis on 60 origins. I counts origins with nonconstant conditional permutation terms. p is exact and one-sided; H4/H16 denote Holm adjustment in the four-test and 16-test families. Negative estimates are retained.
Config.
4
3+1
2+2
2+1+1
1+1+1+1
pmin
Astra
49
6
2
3
0
3−5
Sol
41
12
5
2
0
3−7
Terra
37
11
4
8
0
3−12
Luna
25
17
8
8
2
3−16
Appendix
Table 33: Conditional information in retained repeats. Counts of pooled four-response multiplicities. Only (2,2) and (2,1,1) have nonconstant permutation terms; pmin=3−(A+B) .
Reader
Seed
Native
Raw
Phase 1
Phase 3
Sign
D1
D3
Ds
Dp
HiddenProbe
0
57.73
57.83
57.42
53.82
57.82
7.03
22.98
10.14
2.19
HiddenProbe
1
58.25
58.16
58.01
54.76
58.38
7.99
22.90
11.16
2.37
HiddenProbe
2
57.47
57.52
57.38
54.69
57.60
6.81
20.22
10.25
2.24
ProbeGen
0
57.36
57.36
57.36
57.36
57.36
0.00
0.00
0.00
0.00
ProbeGen
1
56.62
56.62
56.62
56.62
56.62
0.00
0.00
0.00
0.00
ProbeGen
2
57.86
57.86
57.86
57.86
57.86
0.00
0.00
0.00
0.00
Appendix
Table 34: Official CIFAR10 audit by reader seed. All rates are percentages on the same 10,000 test origins. Native uses released inference precision; Raw and transformed accuracy use FP64 target evaluation with FP32 readouts. D1,D3,Ds are disagreements with the audit Raw prediction; Dp compares native and FP32 clean inference.
Reader
Action
Change (pp)
95% interval (pp)
Disagreement (%)
HiddenProbe
Perm.
+0.00
[+0.00,+0.00]
0.00
HiddenProbe
Phase 1
-0.23
[−0.48,+0.02]
7.28
HiddenProbe
Phase 3
-3.41
[−3.88,−2.95]
22.03
HiddenProbe
Sign
+0.10
[−0.20,+0.40]
10.52
ProbeGen
Perm.
+0.00
[+0.00,+0.00]
0.00
ProbeGen
Phase 1
+0.00
[+0.00,+0.00]
0.00
Appendix
Table 35: Paired accuracy changes in the official corpus. Means over 3 reader seeds per method. Intervals are descriptive 95% class-stratified origin bootstrap intervals (10,000 resamples); all seed predictions remain paired within each image. They condition on the selected trained states and fixed actions.
Neural network interpretability studies how models represent information and make predictions, and how to explain, test, or control their internal computations.
Large language models (LLMs) have recently been reported to read and write the raw weights of neural networks directly, a capability termed weight-space programming. We ask whether the reading side of this capability reflects reasoning about weight flows or a combination of local arithmetic at the network's edges and memorized priors over public datasets. We test Claude Opus 5.5 on small convolutional networks of about 125,000 parameters trained on five image datasets, providing every weight as plain text with no tools, no forward passes, and no dataset or class names. The model performs accurate local calculations: batch-normalization statistics lead it to infer about three optimizer steps on networks stopped after three updates, and output-layer sign patterns recover the vehicle/animal split of CIFAR-10. The sessions report reading the first and last layers closely and sampling the middle, and several claims built on partial tracing fail verification, although we did not run a direct forward-pass test. Semantic judgments lean on familiar datasets: the model names CIFAR-10 for CIFAR-10 networks but also for all three fast-food networks, relabels KMNIST as MNIST, reads a 99.57%-accurate SAT-6 network as CIFAR-10 animal photos, and calls "frog" the output favored by a feature that fires most on trees. Permuting CIFAR-10's labels preserves the output split at the new indices, but every class pair the model interprets is then wrong, indicating reliance on remembered label order. These are exploratory results from mostly single runs. We conclude that zero-shot weight reading by this model is better described as local arithmetic plus dataset priors than as weight-flow tracing, and we provide a protocol for auditing future weight-space claims.
Interpretability methods for neural networks are predominantly reactive: they analyse activations produced during specific forward passes, requiring known inputs to find hidden capabilities such as backdoors. We propose Weight Oracles, fine-tuned language models that diagnose properties of a target network by reading its raw weights directly, without behavioural testing. We investigate this paradigm in two phases. Phase I establishes feasibility: through a staged curriculum and an external chain-of-computation that delegates parameter-free operations to deterministic code, an explainer LLM learns to simulate the forward pass of small transformers from their weights, achieving 99% holdout accuracy on unseen targets. Phase II repurposes this infrastructure for safety auditing. We train an oracle on natural language diagnostic questions about weight anomalies using only benign pathologies as training signal, and evaluate it zero-shot on backdoors absent from training. The oracle achieves AUROC 0.93 on attention-routed backdoors and 0.81 across a diversified threat distribution including stealth and adversarially regularized variants. Hand-crafted statistical detectors are sharp on the threat models they implicitly target but collapse on threat-model shift, while the oracle remains uniformly competent across attack types. Scaling to realistic model sizes remains the principal open challenge.
Krishna Kabra, Constantin Venhoff, Christian Schroeder de Witt
Visual question answering requires models to assign attributes and relations to the correct objects, a problem known as visual binding. Structured prompts make these assignments explicit through labeled fields. Yet an improvement from the full prompt does not reveal which labels help or how they interact. First, we isolate field-label effects across ten models and six tasks by holding the image and claim fixed. These comparisons reveal differences that overall accuracy can hide. With a relation label present, adding object-role labels helps Qwen3 but hurts Qwen2 on Spatial. Second, we compare auxiliary hints that preserve or swap the question's role assignments against a no-hint baseline. Across two image sets, Qwen3's advantage with preserving hints mainly reflects errors caused by swapped hints. Preserving hints remain near baseline. Field order further affects the harm. Third, we test clearer hint wording and direct selection of the relation shown in the image. Clearer hints improve text matching but do not consistently improve visual judgments. In contrast, relation selection partly reduces errors for some hints in Qwen3 and GPT-4.1-mini.