Networks that predict angles, gaze directions, and rotations do not output these quantities directly: they emit raw coordinates that a fixed decoder, such as normalization, maps to the final prediction. Attribution methods are usually applied to the raw coordinates, so they can credit inputs for changes the decoder discards, such as rescaling a vector when only its direction matters. The resulting explanation can highlight inputs that have no effect on the prediction. We instead attribute through the decoder differential, which for normalization decoders reduces to a scaled tangent projection. We characterize invariance to output rescaling, show that every fixed smooth scalar target on a sphere has points of zero tangential sensitivity, and show that an endpoint-projected variant of Integrated Gradients can still attribute to scale-only inputs, whereas direct angular Integrated Gradients cannot. On synthetic tasks with analytically known relevance, ambient attribution assigns up to 72% of its mass to an irrelevant region. On MPIIGaze validation data, tangent Grad-CAM produces 44% greater prediction displacement under 10% pixel deletion. On PASCAL3D+ with a 6D rotation decoder, the full differential improves mean deletion displacement by 14.1 degrees over the ambient coordinate sum and by 13.8 degrees over a decoder-invariant scalar. Results for a two-polarity tangent LayerCAM variant are mixed: it exceeds ambient-sum LayerCAM on circular regression but no difference is detected on quaternion pose. These results separate two failure sources, radial contamination and scalarization loss. The approach uses existing attribution methods with decoder-aware targets, although some variants also require changes to attribution aggregation. An invariant attribution can still miss inputs that drive the prediction, so invariance and deletion performance should each be checked.
Figures & tables
Figure 1: Left: a structured-output model factors as x↦z↦π(z) ; attribution is normally taken with respect to z , but the quantity to be explained is the decoded prediction y . Right: on S1 , z and 2z denote the same prediction, so variation along the ray is discarded by the decoder; tangential variation moves the prediction, radial variation spans kerDπ and cannot. By the normalisation Jacobian (§ 2.1 ) the tangent projector is the decoder differential up to a positive scalar.
Table 1: Oracle AMR, the fraction of absolute attribution mass on the prediction-relevant region (1.0 perfect, 0.5 chance, 0.0 all mass on the irrelevant factor). Rows group the three CAM/saliency variants, with ranges spanning the methods; the per-method values quoted in the text are Grad-CAM. SD ≤0.064 . The path-integral row accumulates tangent-Jacobian magnitudes at 32 path points; it is distinct from direct angular and endpoint-projected IG (Appendix N.4 ).
Task / deletion
Ambient-sum
Decoder-invariant scalar
Full differential
p vs. ambient / invariant scalar
Gaze / 10%
4.52±0.36
–
6.50±0.29
0.001 / –
Gaze / mean
8.02±0.65
–
9.08±0.72
0.017 / –
Quaternion pose / mean
47.77±3.48
–
56.55±4.82
0.013 / –
6D pose / mean
48.9
49.2
63.0
0.008 / <0.001
Table 2: Selected real-task deletion results in degrees across five seeds. Values are means, with seed SD where reported. The first row uses 10% deletion; the remaining rows average the nine deletion fractions. Gaze rows use validation participant p13 (also used for epoch selection); gaze and quaternion rows use ResNet18 Grad-CAM; the 6D row compares attribution targets through the Gram-Schmidt decoder. The final column gives uncorrected paired seed-level p -values for the full differential against the ambient sum and, where evaluated, the decoder-invariant scalar. Detailed gaze and quaternion statistics in Appendix P ; the 6D comparison in Appendix N.6 .
Figure 2: Controlled representational separability. Left: attribution mass on the angle-encoding region against α , the amount of prediction-relevant representation mixed into the decoder-null pathway (5 seeds per point, mean ± SD). Tangent is flat at 1.000 by construction; radial climbs off zero as the null pathway begins reading region-A features; ambient, a mixture, follows. The shaded band is the gap. Right: the same gap against the mean cosine similarity of tangent and radial saliency maps over the first 40 test images per run (pooled Pearson r=−0.76 ; definition in Appendix I.1 ). All 35 runs are shown.
Figure 3: PASCAL3D+ SO(3) deletion curves, five seeds (mean ± SD, n=4,886 ), tangent Grad-CAM against the two ambient baselines. Left: against ambient-sum Grad-CAM, the mean tangent curve leads at every fraction; shading marks seed-level p<0.05 . Right: against ambient-sum LayerCAM, whose weighting is the stronger baseline, the plain tangent map trails at low fractions (shaded, p<0.05 ) while no difference is detected for the separate two-polarity tangent comparison (text).
Method
R18
R50
EffV2-S
ViT-B
Ambient-sum Grad-CAM
2.10
0.94
0.65
1.10
Tangent Grad-CAM
1.58
1.32
1.94
0.73
Ambient-sum LayerCAM
5.06
4.45
3.53
0.75
Two-polarity tangent LayerCAM
5.24
4.53
4.33
1.44
Scalar-angle LIME
2.00
2.23
2.45
4.02
Circular LIME
2.32
2.36
2.68
4.78
Table 3: Plant S1 mean deletion displacement (hours; single checkpoint per backbone, n=767 validation images also used for checkpoint selection). Each pair gives a standard target, then its prediction-space counterpart. Full table with directional consistency: Table S3 .
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : MPIIGaze S2 deletion on validation participant p13 ( n=2,996 per run, also used for epoch selection), higher means greater deletion displacement. Left: all seven methods on the original unseeded checkpoint (not one of the five seeded models), the only run in which the perturbation baselines were scored, so the occlusion comparison of § 7.1 rests on this panel alone. Right: tangent vs ambient-sum Grad-CAM over five independent initialisations sharing one LOPO split, mean ± SD across seeds. Shading marks fractions 0.1–0.5, where the paired difference reaches p<0.05 over seeds and favours tangent in 5/5 ; the dashed line marks where the ordering reverses, with 1/5 seeds favouring tangent beyond it. Per-fraction values and tests are in Table S4 .
Figure S2 : Oracle attribution on all three manifolds (rows: S1 , S2 , S3 ). Columns: input, ambient-gradient-magnitude saliency, tangent saliency, radial saliency. The dashed line marks the region boundary, region A (left) provably encodes the angle and region B (right) provably encodes only the irrelevant magnitude factor. Radial attribution lands on region B in every row (it recovers exactly the factor that cannot influence the prediction) while ambient-gradient-magnitude attribution spreads across both. Each panel is inset with the AMR for that single test image alongside the saliency dataset mean from the Table 1 runs. The panel models are separate trainings (4,000 images, 12 epochs). The two can diverge substantially: the S2 ambient-gradient-magnitude panel scores 0.66 on this image against a saliency dataset mean of 0.508. The panels illustrate individual examples; the table reports dataset means.
Figure S3 : Tangent–radial decomposition of an output gradient at a point p on S1 . The tangential component ⟨g,t⟩t is the only part that changes the predicted angle, the radial component ⟨g,p⟩p changes the ambient representation without changing the decoded angle θ^ . Ambient-coordinate attribution combines both. Projecting onto TpS1 discards the radial part by construction.
Method
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Mean
Tangent
5.86
7.88
9.17
9.90
10.34
10.55
10.58
10.51
10.36
9.46
Spherical occlusion
7.86
8.88
9.42
9.68
9.79
9.80
9.69
9.61
9.65
9.37
Spherical LIME
4.60
6.29
7.53
8.43
8.98
9.43
9.70
9.94
10.07
8.33
Spherical KernelSHAP
4.38
5.79
6.92
7.88
8.77
9.36
9.77
10.00
9.97
8.09
Ambient-sum
4.21
5.62
6.53
7.10
7.68
8.41
9.30
10.05
10.39
7.70
Endpoint-projected IG
6.46
7.23
7.58
7.74
7.84
7.92
8.03
8.13
8.19
7.68
Appendix
Table S1 : MPIIGaze S2 per-fraction deletion (degrees, higher means greater deletion displacement), ResNet18, n=2,996 . Validation participant p13, also used for epoch selection. This is the original unseeded checkpoint (validation error 5.49∘ , test error 5.20∘ ), not one of the five seeded models. The seed-replicated tangent/standard comparison, which supersedes the Mean column here, is Table S4 . Bold marks the best value per fraction. The Mean column is left unbolded deliberately, as § 7.1 argues averaging over the full range mixes opposite-signed regimes. The spherical occlusion row is the earlier estimator (patch 16, stride 8, bilinear upsampling), also shown in Figure S1 ; the centred estimator is reported below. Both use black fill and unsigned geodesic displacement, differing in grid resolution and patch-to-pixel mapping.
Figure S4 : Directional consistency with approximate image-effective Wilson intervals, plant S1 (zero-fill deletion; blur IG integration baseline). Image-effective counts are 767 for Grad-CAM and 743 for angular IG; 24 images have no positive IG pixel and are skipped. The dashed line marks chance. Each image contributes multiple pixel deletions, so treating all deletions as independent understates uncertainty. Intervals use the number of contributing images as an approximate effective sample size; they are not calibrated clustered intervals and do not capture training variability. The two tangent Grad-CAM rows show different ResNet18 checkpoints, with DCS 66.1% and 48.7%. Estimates near 52–53% should not be described as established above-chance performance under these intervals.
Backbone
AIG (tan)
AIG (dir)
Std IG (best)
Δ
ResNet18
6.166
6.157
6.164
−0.007
ResNet50
7.078
7.065
7.064
+0.001
EfficientNet
6.333
6.360
6.344
+0.016
ViT
8.397
8.394
8.427
−0.033
IG variant
Target T / units
Mean abs err
Median ratio
Direct
wrapped Δθ (rad)
0.0186
1.008
Appendix
Table S2 : Top: angular versus standard IG deletion AUC (h), blur IG baseline, zero-fill deletion, 50 steps, n=767 ; Δ is direct AIG minus the best standard target selected on this split. Bottom: attribution-sum diagnostics on ResNet18, n=50 . Ratios are medians over images with ∣T∣>10−3 . Direct AIG uses wrapped angular change in radians, equal to lifted change here because no path winds; its mean absolute error is 0.0186 rad (0.071 h). Endpoint-projected AIG uses T=rr′sinΔθ^ , in squared raw-output units. Standard rows sum absolute attributions and divide by their own scalar change, in raw-output units; they are scale diagnostics and test neither completeness nor its failure.
Figure S5 : Radial-penalty controls (ResNet18, plant S1 ), normalised scale-invariant angular metric; bars are means over three initialization seeds with individual seeds overlaid. Every arm except the strong radial penalty (D) lies within seed-level variation of the baseline (A); D has higher rescored error, with checkpoint selection confounded by the earlier norm-dependent metric (Appendix O.2 ).
Method
Invariant
ResNet18
ResNet50
EfficientNetV2- S
ViT- B/16
Ambient-sum Grad-CAM
×
2.10
0.94
0.65
1.10
Ambient-sum HiResCAM
×
2.11
0.94
0.65
0.74
Tangent Grad-CAM
✓
1.58 (0.661)
1.32 (0.493)
1.94 (0.454)
0.73 (0.519)
Tangent HiResCAM (= tangent LayerCAM)
✓
1.58 (0.661)
1.32 (0.493)
1.94 (0.454)
1.46 (0.535)
Ambient-sum LayerCAM (rectified)
×
5.06
4.45
3.53
0.75
Rectified tangent LayerCAM
✓
5.24
4.53
4.33
1.44
Appendix
Table S3 : Plant S1 deletion AUC (hours; mean geodesic displacement of the decoded prediction over the nine deletion fractions under the normalised metric of § 4 , zero-fill deletion, blur IG baseline, single checkpoint per backbone, n=767 validation images also used for checkpoint selection; scalar-angle and circular LIME with the same von Mises weights ( κ=5 , 1000 perturbations), angular IG with 50 steps). Rows are grouped by CAM construction. The two-polarity tangent LayerCAM variant changes both the target and polarity aggregation relative to standard ambient-sum LayerCAM. The ’Invariant’ column marks methods that satisfy the decoder-null invariance condition of Proposition 2.1 ( ✓ * = invariant to a global rescaling of z only, § 3.2 ; tangent angular IG is not invariant, Proposition 3.3 ). Values in parentheses are directional consistency (DCS) for the signed methods: the fraction of the top-100 positively attributed single-pixel deletions that move the prediction against the attributed direction, as removal should (chance 0.5). Tangent Grad-CAM, HiResCAM and LayerCAM coincide on the three GAP CNNs (Appendix H ); on S1 the tangent LayerCAM entry is the tangent HiResCAM map under another name (§ 3.1 ). The rectified tangent row is unsigned and was scored in separate runs with ambient-sum LayerCAM as an in-run control (in-run controls 5.06/4.45/3.53/0.75h , reproducing the ambient-sum LayerCAM row). Occlusion rows use patch 16, stride 8, black fill and centred overlap-add accumulation (729 forward passes per image). The signed geodesic row is the map that carries a direction and a DCS; the unsigned geodesic row ranks by the same displacement without sign cancellation across overlapping patches; unwrapped angular occlusion ranks by the unwrapped decoded-angle difference, which is decoder-invariant but branch-dependent. KernelSHAP and RISE were scored on S1 only before the metric correction and are not reported here; their spherical forms are in Table S1 .
Deletion (deg)
Seed-level (n=5)
Bootstrap (per-image)
Fraction
Tangent
Ambient-sum
Δ± SD
p
Δ [95% CI]
0.1
6.50±0.29
4.52±0.36
+1.97±0.47
0.001
+1.97[+1.88,+2.07]
0.2
8.07±0.28
5.80±0.39
+2.28±0.47
<0.001
+2.28[+2.16,+2.40]
0.3
8.77±0.36
6.73±0.52
+2.03±0.49
0.001
+2.03[+1.90,+2.17]
0.4
9.15±0.69
7.43±0.59
+1.72±0.58
0.003
+1.72[+1.59,+1.86]
0.5
9.49±1.00
8.06±0.61
+1.43±0.75
0.013
+1.43[+1.31,+1.55]
Appendix
Table S4 : MPIIGaze S2 deletion, tangent vs ambient-sum Grad-CAM, five independent initialisations (ResNet18, n=2,996 images per run, identical LOPO split, validation participant p13 (also used for epoch selection), degrees, higher means greater deletion displacement). The Tangent and Ambient-sum columns give the mean across the five runs ± the SD across runs (not within-run image scatter), and Δ is the paired difference. Two uncertainty estimates are reported because they answer different questions. Seed columns give mean ± SD and an uncorrected paired t -test over the five runs (df=4), capturing training variability. The bootstrap column gives a 95% percentile interval from resampling images (20,000 resamples); each image’s paired difference is first averaged over the five seeds, so an image enters once and the five appearances of one image are not treated as independent. The seed and bootstrap point estimates coincide by construction; their intervals measure different sources of variability. They differ a great deal: at fraction 0.1 the seed-level 95% t -interval half-width ( t0.975,4×0.47/5≈0.58∘ ) is about six times the bootstrap half-width (≈0.10∘) . Between-run variation dominates image-sampling variation here, so the seed columns are the conservative and controlling estimate. Deletion fractions are nested subsets, so these tests are dependent and descriptive rather than confirmatory. Relative improvements quoted in the text are means of per-seed ratios, not ratios of means.
Backbone
Test mean
Test med.
≤0.5h
≤1h
≤2h
ResNet18
1.48
0.81
34.5%
60.7%
82.8%
ResNet50
1.33
0.69
39.5%
62.9%
86.5%
EfficientNetV2- S
1.30
0.64
41.1%
65.5%
83.8%
ViT- B/16
1.42
0.73
39.0%
60.7%
84.4%
Random guess
6.00
n/a
n/a
n/a
n/a
Constant (mean time)
2.32
n/a
n/a
n/a
n/a
Appendix
Table S5 : Plant S1 regression accuracy on the held-out test split (angular error in hours, n=377 , normalised metric). Best-epoch selection for these checkpoints used the earlier unnormalised validation metric (§ O.4 ), so validation-best values are not reported; all test values are scored with the normalised metric.
Method
S1
S2
Model angular error
0.364∘
4.667∘
Ambient-gradient-magnitude Grad-CAM
0.804 (0.081)
0.988 (0.006)
Tangent Grad-CAM
0.802 (0.079)
0.988 (0.007)
Radial Grad-CAM
0.807 (0.083)
0.985 (0.014)
Ambient-gradient-magnitude HiResCAM
0.663 (0.059)
0.721 (0.020)
Tangent HiResCAM
0.667 (0.060)
0.722 (0.022)
Appendix
Table S6 : Entangled single-stream oracles ( n=200 each). Mean AMR (std). Grad-CAM and HiResCAM variants are numerically close within each manifold; Saliency varies more across projections.
Kernel
Surrogate
Deletion AUC (h)
std
Euclidean
scalar
2.501
1.978
von Mises
scalar
2.362
1.952
Combined
scalar
2.387
1.892
Euclidean
circular
2.573
1.764
von Mises
circular
2.575
1.847
Combined
circular
2.590
1.854
Appendix
Table S7 : Corrected-metric LIME ablation, ResNet18, n=50 validation images, κ=5 , 1,000 perturbations with a shared per-image random stream, mean-colour superpixel masking, zero-fill deletion, normalised metric. Values are means and per-image SD. The main comparison in Table S3 uses 767 images; absolute scores are not directly interchangeable.
Table S9 : Direction-specificity across backbones (validation error, hours, mean ± SD over 3 init-seeds, normalised scale-invariant metric; checkpoints were selected under the earlier unnormalised metric).
Predictor
Pooled r
Condition-mean r
Seed t(4) , p
Permutation p
α→ gap
−0.86
−0.997
−7.49 , 0.0017
5.0×10−6 ( k=0 )
kerDπ energy → gap
+0.83
+0.991
+5.01 , 0.0075
5.0×10−6 ( k=0 )
map overlap → gap
−0.76
−0.989
−5.28 , 0.0062
5.0×10−6 ( k=0 )
probe R2→ gap
−0.69
−0.941
−5.64 , 0.0049
5.0×10−6 ( k=0 )
Appendix
Table S10 : Repeated-measures statistics for representation-sharing predictors. The table below reports α and the three model-measured predictors introduced in § I.1 . For each it gives the pooled correlation (as in the predictor analysis), the correlation across the 7 condition means, a seed-level slope test ( t , df=4), and a permutation test preserving seed structure (200,000 within-seed relabelings). Condition-mean correlations are consistently stronger than pooled, since averaging over seeds removes between-seed intercept noise.
Figure S6 : Alignment and scalar attribution. Left: S3 oracle AMR against alignment a(q)=∥Pq1∥/2 . Alignment zero identifies a critical point of the coordinate sum; bins have edges 0,0.2,0.4,0.6,0.8,0.95,1 and are plotted at midpoints. The lowest bin contains 52 of 2,000 images and has AMR 0.495, while the tangent map remains near 1. Right: PASCAL3D+ validation data averaged per image over five seeds before binning, showing the tangent-over-sum deletion gap against alignment quintiles. The real-data association is confounded with pose difficulty; it does not measure the gap at an exact blind point.
Figure S7 : Attribution on a representative masked plant image (ResNet18). Panels: input, ambient-sum Grad-CAM, tangent Grad-CAM, ambient-sum LayerCAM, tangent LayerCAM. The outlined band is the segmentation boundary region used in Appendix I . Plant pixels occupy about 1.3% of a masked image on average and 4.9% of this image ( triton032_6mm_20250601_215235_420 ) at 224×224 , and no method concentrates attribution on plant interiors. Ambient-sum LayerCAM differs from ambient-sum Grad-CAM because LayerCAM rectifies each gradient before combining channels, which the GAP collapse of Appendix H does not remove. The tangent LayerCAM panel is the tangent HiResCAM map under another name (no rectification is applied to the projected field, § 3.1 ), and under GAP that map coincides with tangent Grad-CAM, so the two tangent panels are identical by construction.
Attribution target
Decoder-invariant?
Collapse-free?
Mean AUC (deg)
Ambient sum over raw z∈R6
no
no
48.9
Sum over decoded R entries
yes
no
49.2
Tangent differential through Dπ
yes
yes
63.0
Appendix
Table S11 : 6D (Gram-Schmidt) pose, five seeds: mean deletion AUC (degrees) for the three-way decoder comparison of § 7.3 .
Target
Deletion AUC (h)
std
cos only
2.946
1.514
Output norm ∥y∥
2.654
2.195
sin+cos
2.571
1.663
sin only
1.981
1.618
Decoded angle
1.655
1.205
Tangent projection
1.643
1.200
Appendix
Table S12 : Grad-CAM scalar target target, plant S1 , ResNet18, n=100 , deletion AUC ( h , higher means greater deletion displacement) under the normalised metric.
Arm
Description
Val err (h)
10-seed
ρrad
A
Baseline (tanh head, no penalty)
1.076
1.021
0.200
B
L2 -normalized head
1.069
n/a
7×10−15
C
Radial penalty, λ=0.1
1.087
1.019
0.167
D
Radial penalty, λ=1.0
1.780
n/a
0.143
E
Generic grad-norm penalty
1.083
n/a
0.194
F
Tangential penalty, λ=0.1
1.131
1.025
0.261
Appendix
Table S13 : Radial-penalty controls, ResNet18, plant S1 , 10 epochs, normalised angular metric. Val err: mean over 3 initialisation seeds; 10-seed: ten fresh seeds per arm with pre-specified tests; ρrad : radial gradient fraction.
Deletion (deg)
Seed-level (n=5)
Bootstrap
Fraction
Tangent
Ambient-sum
Δ± SD
p
Δ [95% CI]
0.1
6.50±0.29
4.52±0.36
+1.97±0.47
0.001
+1.97[+1.88,+2.07]
0.4
9.15±0.69
7.43±0.59
+1.72±0.58
0.003
+1.72[+1.59,+1.86]
0.7
9.88±1.23
9.81±1.00
+0.06±0.83
0.870
+0.06[−0.01,+0.14]
0.9
10.16±1.24
10.51±1.29
−0.35±0.88
0.416
−0.35[−0.39,−0.32]
Mean
9.08±0.72
8.02±0.65
+1.06±0.60
0.017
n/a
Appendix
Table S14 : MPIIGaze S2 deletion (degrees, higher means greater deletion displacement), tangent vs ambient-sum Grad-CAM, five independent initialisations (ResNet18, validation participant p13, n=2,996 per run, also used for epoch selection). Seed columns: mean ± SD over runs, uncorrected paired t -test (df =4 ); bootstrap: paired per-image differences averaged over seeds. Fractions are nested, so tests are descriptive. Full version in Table S4 .
Fraction
Tangent
Ambient-sum
Δ± SD
p
Δ [95% CI]
0.1
26.42±3.13
21.68±3.88
+4.74±2.59
0.015
+4.74[+4.39,+5.10]
0.4
56.92±4.19
44.90±5.92
+12.02±7.56
0.024
+12.02[+11.30,+12.71]
0.7
67.40±6.91
58.47±4.80
+8.93±5.96
0.029
+8.93[+8.19,+9.66]
Mean
56.55±4.82
47.77±3.48
+8.77±4.65
0.013
+8.77[+8.31,+9.24]
Appendix
Table S15 : PASCAL3D+ SO(3) deletion (degrees of geodesic distance, higher means greater deletion displacement), tangent vs ambient-sum Grad-CAM, five independent initialisations (ResNet18, n=4,886 per run). Seed columns report mean ± SD with an uncorrected paired t -test (df =4 ); the bootstrap column resamples paired per-image differences.
Feature attribution is central to diagnosing and trusting deep neural networks, and Integrated Gradients (IG) is widely used due to its axiomatic properties. However, IG can yield unreliable explanations when the integration path between a baseline and the input passes through regions with noisy gradients. While Guided Integrated Gradients reduces this sensitivity by adaptively updating low-gradient-magnitude features, input-space guidance still produces intermediate inputs that deviate from the data manifold. To address this limitation, we propose Manifold-Aligned Guided Integrated Gradients (MA-GIG), which constructs attribution paths in the latent space of a pre-trained variational autoencoder. By decoding intermediate latent states, MA-GIG biases the path toward the learned generative manifold and reduces exposure to implausible input-space regions. Through qualitative and quantitative evaluations, we demonstrate that MA-GIG produces faithful explanations by aggregating gradients on path features proximal to the input. Consequently, our method reduces off-manifold noise and outperforms prior path-based attribution methods across multiple datasets and classifiers. Our code is available at https://github.com/leekwoon/ma-gig/.
Soyeon Kim, Seongwoo Lim, Kyowoon Lee +1
Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea · INEEJI, Seongnam, Korea · KAIST InnoCORE LLM, KAIST, Daejeon, Korea
Feature attribution methods explain predictions by assigning importance scores to input features. Path-based methods such as Integrated Gradients are especially appealing because they satisfy completeness: attributions sum to the change in model output between a reference state and the input. Yet most path methods define this trajectory in input space, explaining a model through pointwise perturbed inputs along a chosen path. An input-space path integrates the model's raw response at each point it passes through, with no control over the resolution at which a feature is queried; the early, baseline-adjacent part of the trajectory contributes to the explanation on equal footing with the input itself. Here, we lift path attribution from input space to a space of structured probe distributions around the example of interest, and call our method Reveal-IG. Rather than traversing raw input values, Reveal-IG progressively reveals information about the input and attributes changes in the model's expected output along this distributional path. The result is a path-attribution framework that retains completeness with respect to the expected model response, and naturally accommodates multiscale image probes and feature-wise uncertainty in tabular data. Synthetic diagnostics show that Reveal-IG avoids path artifacts that affect input-space methods, and across ImageNet classification and tabular regression it produces stable, signed attributions -- leading on metrics that use attribution sign while remaining competitive on the rest.
We ask a simple question about decoder-only transformers: between which two layers is the probability of a predicted token actually produced? Existing layer-wise readout tools answer only approximately. The logit lens and its trained variant report a per-layer level of probability but give no additive decomposition; their estimates are biased and non-monotone across depth. Direct Logit Attribution and related residual-stream methods are additive, but only in logit space, the softmax nonlinearity breaks additivity in probability space, precisely the quantity one usually cares about. Layer Conductance integrates gradients per layer, but attributes each to its own baseline and so does not sum to the total change in prediction. We introduce IG-Lens, a telescoping application of Integrated Gradients along a single path through the hidden states from a baseline to the final layer. Crediting each segment to the layer it terminates at yields a layer-wise attribution whose sum is exactly the change in target probability, with the softmax inside the integration path rather than linearized away. Our default estimator credits each integration step its observed change in target probability (a prediction-aware reweighting in the spirit of IDGI) rather than its raw gradient. Because the readout is a one-dimensional probability, this collapses each segment to a telescoping sum of endpoint values, so completeness holds exactly (to floating point) at any step count, removing Riemann discretization error while suppressing steps that show gradient sensitivity without a change in output. We give the telescoping identity and its proof, verify completeness to floating point, and describe a single-pass batched implementation computing the full token-by-layer map without any backward call. Code: https://github.com/anhnda/IGLens.