zaaabik commited on
Commit
205da16
·
verified ·
1 Parent(s): f23cf50

Add files using upload-large-folder tool

Browse files
Files changed (49) hide show
  1. _shadow/ps1/features/electra_yelp_s2/trajdev_test.npz +3 -0
  2. _shadow/ps1/features/electra_yelp_s2/trajdev_train.npz +3 -0
  3. _shadow/ps1/features/electra_yelp_s2/trajdev_validation.npz +3 -0
  4. _shadow/ps2/features/roberta_cola_s1/trajclu_test.npz +3 -0
  5. _shadow/ps2/features/roberta_cola_s1/trajclu_train.npz +3 -0
  6. _shadow/ps2/features/roberta_cola_s1/trajclu_validation.npz +3 -0
  7. _shadow/ps2/features/roberta_cola_s1/trajdev_test.npz +3 -0
  8. _shadow/ps2/features/roberta_cola_s1/trajdev_train.npz +3 -0
  9. _shadow/ps2/features/roberta_cola_s1/trajdev_validation.npz +3 -0
  10. _shadow/ps2/features/roberta_cola_s1/trajpath_test.npz +3 -0
  11. _shadow/ps2/features/roberta_cola_s1/trajpath_train.npz +3 -0
  12. _shadow/ps2/features/roberta_cola_s1/trajpath_validation.npz +3 -0
  13. _shadow/ps2/features/roberta_cola_s2/predictions_test.parquet +3 -0
  14. _shadow/ps2/features/roberta_cola_s2/predictions_train.parquet +3 -0
  15. _shadow/ps2/features/roberta_cola_s2/predictions_validation.parquet +3 -0
  16. _shadow/ps2/features/roberta_cola_s2/trajclu_test.npz +3 -0
  17. _shadow/ps2/features/roberta_cola_s2/trajclu_train.npz +3 -0
  18. _shadow/ps2/features/roberta_cola_s2/trajclu_validation.npz +3 -0
  19. _shadow/ps2/features/roberta_cola_s2/trajdev_test.npz +3 -0
  20. _shadow/ps2/features/roberta_cola_s2/trajpath_test.npz +3 -0
  21. _shadow/ps2/features/roberta_cola_s2/trajpath_train.npz +3 -0
  22. _shadow/ps2/features/roberta_cola_s2/trajpath_validation.npz +3 -0
  23. _shadow/ps2/features/roberta_sst5_s1/predictions_validation.parquet +3 -0
  24. mc_dropout/roberta_cola/DONE +1 -0
  25. output/generation/assets/figs/fig_channel_auc.pdf +0 -0
  26. output/generation/assets/figs/fig_mechanism.pdf +0 -0
  27. output/generation/assets/figs/fig_ood.pdf +0 -0
  28. output/generation/assets/figs/fig_percfg_bars.pdf +0 -0
  29. output/generation/assets/figs/fig_percfg_grid.pdf +0 -0
  30. output/generation/assets/figs/fig_percfg_gridv.pdf +0 -0
  31. output/generation/assets/tables/tab_classaxis.tex +3 -3
  32. output/generation/assets/tables/tab_clustered.tex +8 -11
  33. output/generation/assets/tables/tab_controls.tex +6 -6
  34. output/generation/assets/tables/tab_datasets.tex +5 -5
  35. output/generation/assets/tables/tab_ladder.tex +8 -10
  36. output/generation/assets/tables/tab_main.tex +24 -29
  37. output/generation/assets/tables/tab_mechanism.tex +4 -4
  38. output/generation/assets/tables/tab_ood.tex +16 -22
  39. output/generation/assets/tables/tab_percfg.tex +20 -20
  40. output/generation/assets/tables/tab_percfg_full.tex +26 -26
  41. output/generation/assets/tables/tab_percfg_full_rows.tex +21 -21
  42. output/generation/assets/tables/tab_percfg_full_seeds.tex +40 -0
  43. output/generation/assets/tables/tab_seed_cells.tex +28 -0
  44. output/generation/assets/tables/tab_seed_ladder.tex +15 -0
  45. output/generation/assets/tables/tab_seed_ood.tex +31 -0
  46. output/generation/assets/tables/tab_seed_summary.tex +35 -0
  47. output/generation/assets/tables/tab_trajladder.tex +5 -5
  48. output/generation/clustered_inference.json +65 -103
  49. output/generation/last_build.json +64 -34
_shadow/ps1/features/electra_yelp_s2/trajdev_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a0054107c2f944aac730df388719bfa6755a5265c8bf73222365a5cd542ceb9
3
+ size 9553546
_shadow/ps1/features/electra_yelp_s2/trajdev_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6712f88f2ad4c17406eb6edc50357757e2f9ae9e4783b545dd5f40700928b5d5
3
+ size 8694254
_shadow/ps1/features/electra_yelp_s2/trajdev_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9727838be4cf7035d8414c38f04e37cc45d052d6160de70f82beb6f6b7353a7a
3
+ size 3726475
_shadow/ps2/features/roberta_cola_s1/trajclu_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:970ce59b15cf8ae3c59a929bff1e761383233561c7b846e16ba9c88378a948aa
3
+ size 240084
_shadow/ps2/features/roberta_cola_s1/trajclu_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:84abf4ff635486d2b0cd40935af7c938097285d230a17830baf48fff5e940b39
3
+ size 1553238
_shadow/ps2/features/roberta_cola_s1/trajclu_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:336560126574d77ace9e77589ea8f02b880f7b2769ab66123b994b455e2ddc7c
3
+ size 397819
_shadow/ps2/features/roberta_cola_s1/trajdev_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61503e9cfd2985cd95357d99f42fcbbf9f4eea59a45f61c0a55d9d53146eb7ef
3
+ size 337160
_shadow/ps2/features/roberta_cola_s1/trajdev_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9648d04856b98e2322680e2b73930ff3181bbc31fd7712dae795a4b390d225db
3
+ size 2204241
_shadow/ps2/features/roberta_cola_s1/trajdev_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f6b71e0509424a5e362a27944dc11726984f7b14437853d16d491f7604917214
3
+ size 552185
_shadow/ps2/features/roberta_cola_s1/trajpath_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0bea61ca107372a805f456f36f1876f45f75ecb7a4c39d658c25425e5ea2e7b
3
+ size 262427
_shadow/ps2/features/roberta_cola_s1/trajpath_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9dd0bcb2481324d29594377485eced8b9e4189fc5a28f769529164b0cd302982
3
+ size 1715363
_shadow/ps2/features/roberta_cola_s1/trajpath_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:41e9aee72d3ce46bebe1e3ad86a6adfe990e0074e1459de9c56a98abd72bc6de
3
+ size 429804
_shadow/ps2/features/roberta_cola_s2/predictions_test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b31e1887231eb936c60b9aa1748344450c94c0fe320978101f4cae3239eaf33e
3
+ size 51383
_shadow/ps2/features/roberta_cola_s2/predictions_train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff49dd7f1ad88cc0983640d5cecc6168d0385efd9286142b5fe1a1796a38a09d
3
+ size 308015
_shadow/ps2/features/roberta_cola_s2/predictions_validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:566646cd6788ebf88acba9516b01f516c29cc2ff1990f36f057da09b765dd06a
3
+ size 80061
_shadow/ps2/features/roberta_cola_s2/trajclu_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:354c5f15ee677ff108a26a0d6a37f81cb2c6b1566e61dd0548de6cf2003ea005
3
+ size 245044
_shadow/ps2/features/roberta_cola_s2/trajclu_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:280733c6d996650418b24a450f994d094d1a19863ab690a0532db342bce9ca5a
3
+ size 1584805
_shadow/ps2/features/roberta_cola_s2/trajclu_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7c9919122f3a6a81d7ac24df4238d2fdc20bd0ce955f016a2a25be8e4694784
3
+ size 398472
_shadow/ps2/features/roberta_cola_s2/trajdev_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:942f87a0ddc00bbe8728db6cd7dd3e2585a7eb5cfaf06d55b14100a6efc12ff1
3
+ size 334829
_shadow/ps2/features/roberta_cola_s2/trajpath_test.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2dcb0bef61308e903a4a3bad2f4bcb6836b27dfb2a1e9621a64df1022da269e8
3
+ size 262189
_shadow/ps2/features/roberta_cola_s2/trajpath_train.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:27a829ff6a31a81cdb1f669cd4354063ba226f2e5205df4d232fae68df24c938
3
+ size 1713514
_shadow/ps2/features/roberta_cola_s2/trajpath_validation.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ee1d908c3a30da6255be759b9a8a1615ae59a356579616ca43a37ea2921a5ac
3
+ size 429407
_shadow/ps2/features/roberta_sst5_s1/predictions_validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b24e74e502e9fb6b3dde309e92383c7da7d19eb94a19e0272672a7b644bda9e
3
+ size 106648
mc_dropout/roberta_cola/DONE ADDED
@@ -0,0 +1 @@
 
 
1
+ passes=20 save_states=10
output/generation/assets/figs/fig_channel_auc.pdf ADDED
Binary file (44.2 kB). View file
 
output/generation/assets/figs/fig_mechanism.pdf CHANGED
Binary files a/output/generation/assets/figs/fig_mechanism.pdf and b/output/generation/assets/figs/fig_mechanism.pdf differ
 
output/generation/assets/figs/fig_ood.pdf CHANGED
Binary files a/output/generation/assets/figs/fig_ood.pdf and b/output/generation/assets/figs/fig_ood.pdf differ
 
output/generation/assets/figs/fig_percfg_bars.pdf CHANGED
Binary files a/output/generation/assets/figs/fig_percfg_bars.pdf and b/output/generation/assets/figs/fig_percfg_bars.pdf differ
 
output/generation/assets/figs/fig_percfg_grid.pdf CHANGED
Binary files a/output/generation/assets/figs/fig_percfg_grid.pdf and b/output/generation/assets/figs/fig_percfg_grid.pdf differ
 
output/generation/assets/figs/fig_percfg_gridv.pdf CHANGED
Binary files a/output/generation/assets/figs/fig_percfg_gridv.pdf and b/output/generation/assets/figs/fig_percfg_gridv.pdf differ
 
output/generation/assets/tables/tab_classaxis.tex CHANGED
@@ -5,9 +5,9 @@
5
  \toprule
6
  $C$ & cells & shape & identity \\
7
  \midrule
8
- $2$ & $8$ & $+1.34$ & $-0.24$ \\
9
- $3$ & $4$ & $+0.36$ & $+0.13$ \\
10
- $5$ & $4$ & $+3.49$ & $+0.96$ \\
11
  \bottomrule
12
  \end{tabular}
13
  \end{table}
 
5
  \toprule
6
  $C$ & cells & shape & identity \\
7
  \midrule
8
+ $2$ & $8$ & $+1.20$ & $-0.12$ \\
9
+ $3$ & $4$ & $+0.37$ & $+0.15$ \\
10
+ $5$ & $4$ & $+3.94$ & $+0.79$ \\
11
  \bottomrule
12
  \end{tabular}
13
  \end{table}
output/generation/assets/tables/tab_clustered.tex CHANGED
@@ -1,20 +1,17 @@
1
  \begin{table*}[t]\centering\small
2
- \caption{Every comparison the paper rests on, at both units. The mean difference and the configuration-level $p$ treat the cells as independent, which they are not: the grid is encoders crossed with 4 tasks. $p_{\mathrm{task}}$ is an exact sign-flip test over task means and cannot fall below $2/2^{4}=0.125$, so a configuration-level value beneath it is answering a different question rather than answering the same one more strongly. $r$ is the matched-pairs rank-biserial effect size, and ICC the share of variance the task term carries. The interval is a task-clustered bootstrap; at this many clusters its coverage is well below nominal, so we report it as a measure of spread and rely on the sign-flip test. Rows marked $\dagger$ cover fewer configurations than the rest and are not comparable to them. Rows marked $\ddagger$ pair a five-seed detector mean against a single-run baseline, because the probe baselines are produced once per configuration rather than per seed; the pairing is therefore between quantities with different sampling error on the two sides.}
3
  \label{tab:clustered}
4
  \begin{tabular}{lrrrrrrr}
5
  \toprule
6
  Comparison & $\Delta$ & 95\% CI & wins & $p_{\mathrm{cfg}}$ & $p_{\mathrm{task}}$ & $r$ & ICC \\
7
  \midrule
8
- sorted over collapsed & $+1.63$ & $[+0.67, +2.86]$ & 14/16 & $5.8\times 10^{-4}$ & $0.125$ & $+0.90$ & $0.39$ \\
9
- profile over collapsed & $+1.78$ & $[+0.47, +3.46]$ & 15/16 & $0.002$ & $0.125$ & $+0.82$ & $0.52$ \\
10
- trajectory over profile & $+0.49$ & $[+0.10, +0.75]$ & 13/16 & $0.008$ & $0.250$ & $+0.74$ & $0.28$ \\
11
- profile over raw CLS & $+1.98$ & $[+0.79, +3.65]$ & 14/16 & $4.3\times 10^{-4}$ & $0.125$ & $+0.91$ & $0.71$ \\
12
- MT over raw CLS & $+2.47$ & $[+1.31, +4.16]$ & 15/16 & $6.1\times 10^{-5}$ & $0.125$ & $+0.99$ & $0.72$ \\
13
- profile over attention & $+1.90$ & $[+0.31, +4.14]$ & 13/16 & $0.009$ & $0.250$ & $+0.72$ & $0.70$ \\
14
- profile over strongest probe$^\ddagger$ & $+1.43$ & $[+0.72, +2.15]$ & 12/16 & $0.016$ & $0.125$ & $+0.68$ & $0.00$ \\
15
- MT over strongest probe$^\ddagger$ & $+1.93$ & $[+1.42, +2.67]$ & 14/16 & $0.002$ & $0.125$ & $+0.82$ & $0.00$ \\
16
- profile over final layer in full & $+0.28$ & $[+0.22, +0.39]$ & 13/16 & $0.034$ & $0.125$ & $+0.60$ & $0.00$ \\
17
- MT over final layer in full & $+0.78$ & $[+0.35, +1.06]$ & 14/16 & $5.8\times 10^{-4}$ & $0.125$ & $+0.90$ & $0.19$ \\
18
  \bottomrule
19
  \end{tabular}
20
  \end{table*}
 
1
  \begin{table*}[t]\centering\small
2
+ \caption{Paired comparisons at both units of the grid. $\Delta$ is the mean difference in ROC-AUC over the 16 configurations, wins the number in which it is positive, $p_{\mathrm{cfg}}$ a Wilcoxon signed-rank test over those cells and $p_{\mathrm{task}}$ an exact sign-flip test over the 4 task means, whose floor is $2/2^{4}=0.125$; $r$, ICC and the task-clustered bootstrap interval are defined in Section~\ref{subsec:stats}. Rows marked $\dagger$ cover fewer configurations than the rest; rows marked $\ddagger$ compare against a probe baseline produced once per configuration.}
3
  \label{tab:clustered}
4
  \begin{tabular}{lrrrrrrr}
5
  \toprule
6
  Comparison & $\Delta$ & 95\% CI & wins & $p_{\mathrm{cfg}}$ & $p_{\mathrm{task}}$ & $r$ & ICC \\
7
  \midrule
8
+ sorted over collapsed & $+1.68$ & $[+0.60, +3.24]$ & 16/16 & $3.1\times 10^{-5}$ & $0.125$ & $+1.00$ & $0.74$ \\
9
+ profile over collapsed & $+1.85$ & $[+0.63, +3.73]$ & 15/16 & $4.3\times 10^{-4}$ & $0.125$ & $+0.91$ & $0.77$ \\
10
+ trajectory over profile & $+0.57$ & $[+0.19, +0.91]$ & 13/16 & $0.005$ & $0.250$ & $+0.76$ & $0.23$ \\
11
+ profile over raw CLS & $+1.87$ & $[+0.89, +3.37]$ & 15/16 & $1.5\times 10^{-4}$ & $0.125$ & $+0.96$ & $0.72$ \\
12
+ MT over raw CLS & $+2.44$ & $[+1.62, +3.91]$ & 16/16 & $3.1\times 10^{-5}$ & $0.125$ & $+1.00$ & $0.71$ \\
13
+ profile over strongest probe$^\ddagger$ & $+1.61$ & $[+0.66, +2.27]$ & 14/16 & $5.8\times 10^{-4}$ & $0.125$ & $+0.90$ & $0.27$ \\
14
+ MT over strongest probe$^\ddagger$ & $+2.18$ & $[+1.46, +2.90]$ & 16/16 & $3.1\times 10^{-5}$ & $0.125$ & $+1.00$ & $0.26$ \\
 
 
 
15
  \bottomrule
16
  \end{tabular}
17
  \end{table*}
output/generation/assets/tables/tab_controls.tex CHANGED
@@ -1,15 +1,15 @@
1
  \begin{table*}[t]\centering\small
2
- \caption{Per-sample layer-shuffle ablation. For every example, the interior transformer layers are independently permuted while the embedding and final layers remain fixed; features are rebuilt and detectors refit. Shuffled results average three permutation seeds. The Profile row is the reference without trajectory features; the next columns give what original and shuffled trajectory features add to it on the same 16 configurations. The final two columns compare each shuffled representation with its unshuffled counterpart. Both $p$ columns use two-sided Wilcoxon signed-rank tests over paired configurations.}
3
  \label{tab:controls}
4
  \begin{tabular}{lrrrrrr}
5
  \toprule
6
  Representation & ROC-AUC & $\Delta$ vs Profile & wins & $p$ & $\Delta$ vs original & $p$ \\
7
  \midrule
8
- Profile & 73.76 & --- & --- & --- & --- & --- \\
9
- MT (Profile\,+\,TRAJ) & 74.25 & $+0.49$ & 13/16 & 0.008 & --- & --- \\
10
- \quad per-sample shuffled MT & 73.20 & $-0.56$ & 3/16 & 0.003 & $-1.06$ & $<$0.001 \\
11
- TRAJ & 73.08 & $-0.69$ & 6/16 & 0.175 & --- & --- \\
12
- \quad per-sample shuffled TRAJ & 70.94 & $-2.82$ & 0/16 & $<$0.001 & $-2.14$ & $<$0.001 \\
13
  \bottomrule
14
  \end{tabular}
15
  \end{table*}
 
1
  \begin{table*}[t]\centering\small
2
+ \caption{Per-sample layer-shuffle ablation. For every example, the interior transformer layers are independently permuted while the embedding and final layers remain fixed; features are rebuilt and detectors refit. Shuffled results average three permutation seeds; every entry is the mean over the three fine-tuning seeds. The Profile row is the reference without trajectory features; the next columns give what original and shuffled trajectory features add to it on the same 16 configurations. The final two columns compare each shuffled representation with its unshuffled counterpart. Both $p$ columns use two-sided Wilcoxon signed-rank tests over paired configurations.}
3
  \label{tab:controls}
4
  \begin{tabular}{lrrrrrr}
5
  \toprule
6
  Representation & ROC-AUC & $\Delta$ vs Profile & wins & $p$ & $\Delta$ vs original & $p$ \\
7
  \midrule
8
+ Profile & 73.96 & --- & --- & --- & --- & --- \\
9
+ MT (Profile\,+\,TRAJ) & 74.53 & $+0.57$ & 13/16 & 0.005 & --- & --- \\
10
+ \quad per-sample shuffled MT & 73.49 & $-0.48$ & 4/16 & 0.025 & $-1.04$ & $<$0.001 \\
11
+ TRAJ & 73.28 & $-0.68$ & 6/16 & 0.144 & --- & --- \\
12
+ \quad per-sample shuffled TRAJ & 71.25 & $-2.71$ & 0/16 & $<$0.001 & $-2.03$ & $<$0.001 \\
13
  \bottomrule
14
  \end{tabular}
15
  \end{table*}
output/generation/assets/tables/tab_datasets.tex CHANGED
@@ -1,6 +1,6 @@
1
  \begin{table}[t]\centering\footnotesize
2
  \setlength{\tabcolsep}{3pt}
3
- \caption{Datasets, split sizes and base-classifier accuracy (\%). Splits are shared by every encoder and fixed across every experiment; the detector never sees the test split during fitting or selection.}
4
  \label{tab:datasets}
5
  \resizebox{\columnwidth}{!}{%
6
  \begin{tabular}{lrrrrrrrr}
@@ -9,10 +9,10 @@
9
  \cmidrule(lr){6-9}
10
  Task & $C$ & Train & Val & Test & RoBERTa & DeBERTa & DeBERTa-s & ELECTRA \\
11
  \midrule
12
- CoLA & 2 & 6,840 & 1,711 & 1,043 & 79.58 & 84.18 & 82.26 & 81.88 \\
13
- SST-5 & 5 & 8,544 & 1,101 & 2,210 & 52.71 & 52.85 & 50.23 & 47.74 \\
14
- ToxiGen & 2 & 7,168 & 1,792 & 940 & 79.26 & 78.09 & 76.06 & 78.72 \\
15
- Yelp & 3 & 27,300 & 11,700 & 30,000 & 81.15 & 82.69 & 78.96 & 79.55 \\
16
  \bottomrule
17
  \end{tabular}}
18
  \end{table}
 
1
  \begin{table}[t]\centering\footnotesize
2
  \setlength{\tabcolsep}{3pt}
3
+ \caption{Datasets, split sizes and base-classifier accuracy (\%, mean $\pm$ standard deviation over the three fine-tuning seeds). Splits are shared by every encoder and fixed across every experiment; the detector never sees the test split during fitting or selection.}
4
  \label{tab:datasets}
5
  \resizebox{\columnwidth}{!}{%
6
  \begin{tabular}{lrrrrrrrr}
 
9
  \cmidrule(lr){6-9}
10
  Task & $C$ & Train & Val & Test & RoBERTa & DeBERTa & DeBERTa-s & ELECTRA \\
11
  \midrule
12
+ CoLA & 2 & 6,840 & 1,711 & 1,043 & 79.2$\pm$0.5 & 84.3$\pm$0.1 & 82.1$\pm$0.1 & 82.0$\pm$0.2 \\
13
+ SST-5 & 5 & 8,544 & 1,101 & 2,210 & 52.3$\pm$0.4 & 52.9$\pm$1.2 & 50.8$\pm$1.4 & 47.5$\pm$0.3 \\
14
+ ToxiGen & 2 & 7,168 & 1,792 & 940 & 76.3$\pm$5.5 & 78.5$\pm$0.4 & 77.4$\pm$1.2 & 78.5$\pm$0.4 \\
15
+ Yelp & 3 & 27,300 & 11,700 & 30,000 & 81.1$\pm$0.0 & 82.5$\pm$0.1 & 79.4$\pm$0.6 & 79.1$\pm$1.0 \\
16
  \bottomrule
17
  \end{tabular}}
18
  \end{table}
output/generation/assets/tables/tab_ladder.tex CHANGED
@@ -1,27 +1,25 @@
1
  \begin{table}[t]\centering\small
2
- \caption{The matched-readout ladder over the 16-configuration grid. Group A is the closed-form score as usually reported; every row below it is read by the detector used throughout, so a step changes only what the detector is given. Group C undoes the class collapse, the depth collapse, or both, each measured from the scalar of group B; the class and depth rows are alternatives at one position, not steps in a sequence. Group D gives the same detector everything the final layer offers instead, measured from the profile, as is the last row. }
3
  \label{tab:ladder}
4
  \begin{tabular}{@{}lrr@{}}
5
  \toprule
6
  Representation given to the detector & ROC-AUC & $\Delta$ \\
7
  \midrule
8
  \multicolumn{3}{@{}l}{\emph{A. the standard score}} \\
9
- closed-form RMD, final layer, no detector & 63.56 & \\
10
  \addlinespace[2pt]
11
  \multicolumn{3}{@{}l}{\emph{B. the learned readout ($\Delta$ from A)}} \\
12
- the same scalar, read by the detector & 70.09 & $+6.53$ \\
13
  \addlinespace[2pt]
14
  \multicolumn{3}{@{}l}{\emph{C. the collapses undone ($\Delta$ from B)}} \\
15
- final layer, every class & 72.49 & $+2.40$ \\
16
- every layer, one class & 71.98 & $+1.89$ \\
17
- every layer, every class (\textbf{Profile}) & 73.76 & $+3.67$ \\
18
  \addlinespace[2pt]
19
  \multicolumn{3}{@{}l}{\emph{D. the final layer in full ($\Delta$ from Profile)}} \\
20
- raw \texttt{[CLS]} & 71.78 & $-1.98$ \\
21
- logits\,+\,MD\,+\,RMD (LastOut) & 73.06 & $-0.70$ \\
22
- raw \texttt{[CLS]}\,+\,logits\,+\,distances (LastAll) & 73.48 & $-0.28$ \\
23
  \midrule
24
- profile\,+\,trajectory (\textbf{MT}) & 74.25 & $+0.49$ \\
25
  \bottomrule
26
  \end{tabular}
27
  \end{table}
 
1
  \begin{table}[t]\centering\small
2
+ \caption{The matched-readout ladder over the 16-configuration grid (three fine-tuning seeds $\times$ five detector seeds; every entry is the mean over the three fine-tuning runs). Group A is the closed-form score as usually reported; every row below it is read by the detector used throughout, so a step changes only what the detector is given. Group C undoes the class collapse, the depth collapse, or both, each measured from the scalar of group B; the class and depth rows are alternatives at one position, not steps in a sequence. Group D gives the same detector the final layer instead, as its raw state and as its logits with the distances, measured from the profile, as is the last row. }
3
  \label{tab:ladder}
4
  \begin{tabular}{@{}lrr@{}}
5
  \toprule
6
  Representation given to the detector & ROC-AUC & $\Delta$ \\
7
  \midrule
8
  \multicolumn{3}{@{}l}{\emph{A. the standard score}} \\
9
+ closed-form RMD, final layer, no detector & 63.65 & \\
10
  \addlinespace[2pt]
11
  \multicolumn{3}{@{}l}{\emph{B. the learned readout ($\Delta$ from A)}} \\
12
+ the same scalar, read by the detector & 70.47 & $+6.82$ \\
13
  \addlinespace[2pt]
14
  \multicolumn{3}{@{}l}{\emph{C. the collapses undone ($\Delta$ from B)}} \\
15
+ final layer, every class & 72.70 & $+2.23$ \\
16
+ every layer, one class & 72.11 & $+1.64$ \\
17
+ every layer, every class (\textbf{Profile}) & 73.96 & $+3.49$ \\
18
  \addlinespace[2pt]
19
  \multicolumn{3}{@{}l}{\emph{D. the final layer in full ($\Delta$ from Profile)}} \\
20
+ raw \texttt{[CLS]} & 72.09 & $-1.87$ \\
 
 
21
  \midrule
22
+ profile\,+\,trajectory (\textbf{MT}) & 74.53 & $+0.57$ \\
23
  \bottomrule
24
  \end{tabular}
25
  \end{table}
output/generation/assets/tables/tab_main.tex CHANGED
@@ -1,42 +1,37 @@
1
  \begin{table*}[t]\centering\small
2
- \caption{Misclassification detection over the 16-configuration grid (4 encoders $\times$ 4 tasks, five detector seeds). ROC-AUC is higher-is-better; AURC, AUGRC and risk at 80\% and 90\% coverage are lower-is-better. The pool rows are produced by the same detector and the same splits; the output- and distance-space rows are closed form and the probe rows are logistic regression (Section~\ref{subsec:baselines}). The last row combines MT with the final layer in full.}
3
  \label{tab:main}
4
  \begin{tabular}{lrrrrr}
5
  \toprule
6
  Method & ROC-AUC & AURC & AUGRC & r@80 & r@90 \\
7
  \midrule
8
  \multicolumn{6}{l}{\emph{Output- and distance-space baselines}} \\
9
- MD (final layer) & 58.18 & 0.2304 & 0.1217 & 0.2501 & 0.2604 \\
10
- RMD (final layer) & 63.56 & 0.2188 & 0.1134 & 0.2253 & 0.2449 \\
11
- MD, all layers & 63.80 & 0.2072 & 0.1135 & 0.2428 & 0.2550 \\
12
- RMD, all layers & 71.98 & 0.1732 & 0.0994 & 0.2164 & 0.2423 \\
13
- CLUE-2c & 62.28 & 0.2010 & 0.1164 & 0.2635 & 0.2658 \\
14
- TrustScore & 62.03 & 0.2293 & 0.1160 & 0.2320 & 0.2486 \\
15
- Entropy & 70.30 & 0.1751 & 0.1020 & 0.2220 & 0.2446 \\
16
- SR & 70.76 & 0.1745 & 0.1009 & 0.2212 & 0.2444 \\
17
- DOCTOR & 70.47 & 0.1746 & 0.1017 & 0.2217 & 0.2442 \\
18
- MaxLogit & 70.45 & 0.1739 & 0.1014 & 0.2222 & 0.2444 \\
19
- Energy & 69.06 & 0.1825 & 0.1040 & 0.2256 & 0.2471 \\
20
  \multicolumn{6}{l}{\emph{Supervised probes}} \\
21
- LinProbe\_begin & 62.10 & 0.2115 & 0.1153 & 0.2429 & 0.2554 \\
22
- LinProbe\_mid & 65.20 & 0.1928 & 0.1100 & 0.2402 & 0.2547 \\
23
- LinProbe\_last & 71.78 & 0.1712 & 0.0995 & 0.2212 & 0.2444 \\
24
- AttnProbe\_begin & 64.27 & 0.1974 & 0.1117 & 0.2409 & 0.2561 \\
25
- AttnProbe\_mid & 67.39 & 0.1830 & 0.1063 & 0.2352 & 0.2524 \\
26
- AttnProbe\_last & 72.33 & 0.1676 & 0.0984 & 0.2200 & 0.2438 \\
27
  \multicolumn{6}{l}{\emph{Final layer, same detector as the pools}} \\
28
- Raw \texttt{[CLS]} & 71.78 & 0.1716 & 0.0997 & 0.2184 & 0.2418 \\
29
- Logits\,+\,MD\,+\,RMD (LastOut) & 73.06 & 0.1643 & 0.0967 & 0.2127 & 0.2389 \\
30
- Raw \texttt{[CLS]}\,+\,logits\,+\,distances (LastAll) & 73.48 & 0.1629 & 0.0961 & 0.2133 & 0.2402 \\
31
  \multicolumn{6}{l}{\emph{Representation pools (ours in bold)}} \\
32
- AttnTopo & 71.86 & 0.1717 & 0.0998 & 0.2195 & 0.2427 \\
33
- Mahalanobis profile, no background & 65.47 & 0.1988 & 0.1106 & 0.2390 & 0.2531 \\
34
- MT, no background & 73.23 & 0.1655 & 0.0970 & 0.2127 & 0.2383 \\
35
- \textbf{Mahalanobis trajectory} & 73.08 & 0.1654 & 0.0972 & 0.2133 & 0.2387 \\
36
- \textbf{Mahalanobis profile} & 73.76 & 0.1611 & 0.0956 & 0.2133 & 0.2390 \\
37
- \textbf{MT} (profile\,+\,trajectory) & \textbf{74.25} & 0.1595 & 0.0946 & 0.2108 & 0.2380 \\
38
- \multicolumn{6}{l}{\emph{Combined}} \\
39
- \textbf{MT}\,+\,LastAll & 74.17 & 0.1598 & 0.0948 & 0.2117 & 0.2375 \\
40
  \bottomrule
41
  \end{tabular}
42
  \end{table*}
 
1
  \begin{table*}[t]\centering\small
2
+ \caption{Misclassification detection over the 16-configuration grid (4 encoders $\times$ 4 tasks, three fine-tuning seeds $\times$ five detector seeds; every entry is the mean over the three fine-tuning runs). ROC-AUC is higher-is-better; AURC, AUGRC and risk at 80\% and 90\% coverage are lower-is-better; the best value of each column is bold. The pool rows are produced by the same detector and the same splits; the output- and distance-space rows are closed form and the probe rows are logistic regression (Section~\ref{subsec:baselines}).}
3
  \label{tab:main}
4
  \begin{tabular}{lrrrrr}
5
  \toprule
6
  Method & ROC-AUC & AURC & AUGRC & r@80 & r@90 \\
7
  \midrule
8
  \multicolumn{6}{l}{\emph{Output- and distance-space baselines}} \\
9
+ MD (final layer) & 58.45 & 0.2323 & 0.1221 & 0.2503 & 0.2610 \\
10
+ RMD (final layer) & 63.65 & 0.2195 & 0.1139 & 0.2271 & 0.2466 \\
11
+ MD, all layers & 64.08 & 0.2065 & 0.1135 & 0.2431 & 0.2564 \\
12
+ RMD, all layers & 72.11 & 0.1740 & 0.0997 & 0.2184 & 0.2433 \\
13
+ CLUE-2c & 62.25 & 0.2013 & 0.1169 & 0.2672 & 0.2680 \\
14
+ TrustScore & 63.41 & 0.2248 & 0.1143 & 0.2304 & 0.2496 \\
15
+ Entropy & 70.54 & 0.1752 & 0.1022 & 0.2235 & 0.2465 \\
16
+ SR & 70.99 & 0.1741 & 0.1012 & 0.2230 & 0.2462 \\
17
+ DOCTOR & 70.69 & 0.1743 & 0.1019 & 0.2232 & 0.2461 \\
18
+ MaxLogit & 70.63 & 0.1735 & 0.1018 & 0.2240 & 0.2460 \\
19
+ Energy & 69.37 & 0.1812 & 0.1042 & 0.2283 & 0.2489 \\
20
  \multicolumn{6}{l}{\emph{Supervised probes}} \\
21
+ LinProbe\_begin & 62.00 & 0.2118 & 0.1158 & 0.2450 & 0.2577 \\
22
+ LinProbe\_mid & 65.24 & 0.1941 & 0.1103 & 0.2406 & 0.2552 \\
23
+ LinProbe\_last & 72.16 & 0.1691 & 0.0992 & 0.2221 & 0.2461 \\
24
+ AttnProbe\_begin & 64.35 & 0.1979 & 0.1119 & 0.2427 & 0.2571 \\
25
+ AttnProbe\_mid & 67.70 & 0.1818 & 0.1062 & 0.2363 & 0.2532 \\
26
+ AttnProbe\_last & 72.35 & 0.1676 & 0.0988 & 0.2212 & 0.2458 \\
27
  \multicolumn{6}{l}{\emph{Final layer, same detector as the pools}} \\
28
+ Raw \texttt{[CLS]} & 72.09 & 0.1698 & 0.0996 & 0.2196 & 0.2437 \\
 
 
29
  \multicolumn{6}{l}{\emph{Representation pools (ours in bold)}} \\
30
+ Mahalanobis profile, no background & 65.57 & 0.1987 & 0.1108 & 0.2393 & 0.2548 \\
31
+ MT, no background & 73.45 & 0.1650 & 0.0970 & 0.2138 & 0.2398 \\
32
+ \textbf{Mahalanobis trajectory} & 73.28 & 0.1652 & 0.0973 & 0.2147 & 0.2401 \\
33
+ \textbf{Mahalanobis profile} & 73.96 & 0.1605 & 0.0956 & 0.2138 & 0.2410 \\
34
+ \textbf{MT} (profile\,+\,trajectory) & \textbf{74.53} & \textbf{0.1585} & \textbf{0.0946} & \textbf{0.2122} & \textbf{0.2388} \\
 
 
 
35
  \bottomrule
36
  \end{tabular}
37
  \end{table*}
output/generation/assets/tables/tab_mechanism.tex CHANGED
@@ -5,10 +5,10 @@
5
  \toprule
6
  Quantity & Errors & mean AUC & agree \\
7
  \midrule
8
- Margin at the output, $d_{\mathrm{rival}}-d_{\mathrm{pred}}$ & smaller & 0.693 & 15/16 \\
9
- Commitment depth (relative depth of the last class change) & later & 0.638 & 15/16 \\
10
- Longest run held on one class (fraction of depth) & shorter & 0.601 & 14/16 \\
11
- Class changes per layer & more & 0.591 & 14/16 \\
12
  \bottomrule
13
  \end{tabular}
14
  \end{table*}
 
5
  \toprule
6
  Quantity & Errors & mean AUC & agree \\
7
  \midrule
8
+ Margin at the output, $d_{\mathrm{rival}}-d_{\mathrm{pred}}$ & smaller & 0.695 & 15/16 \\
9
+ Commitment depth (relative depth of the last class change) & later & 0.637 & 15/16 \\
10
+ Longest run held on one class (fraction of depth) & shorter & 0.601 & 15/16 \\
11
+ Class changes per layer & more & 0.594 & 15/16 \\
12
  \bottomrule
13
  \end{tabular}
14
  \end{table*}
output/generation/assets/tables/tab_ood.tex CHANGED
@@ -1,36 +1,30 @@
1
  \begin{table*}[t]\centering\small
2
- \caption{Out-of-distribution detection over 48 scenarios (16 configurations $\times$ the 3 other tasks of the grid). Every readout is one-class: it is fitted on the training split of the in-distribution task only and sees no correctness label and no out-of-distribution example. AUROC and FPR95 use the full in-distribution split; only AUPR uses a size-matched draw, so only its chance level is $50$. Near, mid and far are terciles of a corpus distance measured from raw text alone --- the mean of the Jensen-Shannon divergences between the two length distributions and between the two unigram distributions --- so the axis is a property of the pair rather than of any representation. The six unordered corpus pairs are divided into two pairs per tier. The effective independent units are therefore the corpus pairs, not the 48 encoder--direction scenarios.}
3
  \label{tab:ood}
4
  \begin{tabular}{lrrrrrr}
5
  \toprule
6
  Readout & AUROC & AUPR & FPR95 & near (16) & mid (16) & far (16) \\
7
  \midrule
8
  \multicolumn{7}{l}{\emph{Logit-based}} \\
9
- \quad Energy & 71.45 & 69.38 & 56.91 & 59.53 & 70.80 & 84.03 \\
10
- \quad MaxLogit & 71.04 & 68.99 & 57.64 & 59.02 & 70.45 & 83.64 \\
11
- \quad Entropy & 70.84 & 69.15 & 57.77 & 58.95 & 70.79 & 82.78 \\
12
- \quad MSP & 68.98 & 67.26 & 61.90 & 57.60 & 68.27 & 81.09 \\
13
  \multicolumn{7}{l}{\emph{Final layer}} \\
14
- \quad MD & 81.50 & 80.18 & 40.60 & 77.57 & 79.34 & 87.60 \\
15
- \quad MD background & 81.48 & 80.16 & 40.66 & 77.54 & 79.32 & 87.57 \\
16
- \quad $k$-NN & 80.52 & 78.06 & 46.55 & 74.20 & 79.35 & 88.00 \\
17
- \quad RMD & 67.11 & 65.18 & 71.95 & 62.28 & 68.31 & 70.72 \\
18
  \multicolumn{7}{l}{\emph{All layers, distances only}} \\
19
- \quad MD, all layers & 92.20 & 90.02 & 24.57 & 84.63 & 94.52 & 97.45 \\
20
- \quad MD background, all layers & 92.13 & 89.96 & 24.89 & 84.46 & 94.49 & 97.43 \\
21
- \quad Profile, no background & 91.74 & 89.69 & 26.42 & 84.00 & 93.94 & 97.27 \\
22
- \quad Profile & 77.18 & 74.95 & 58.03 & 70.10 & 77.70 & 83.73 \\
23
- \quad RMD, all layers & 70.46 & 70.69 & 65.17 & 56.00 & 70.38 & 85.00 \\
24
  \multicolumn{7}{l}{\emph{Trajectory only}} \\
25
- \quad Traj (path geometry) & 91.54 & 89.39 & 26.72 & 83.22 & 94.10 & 97.30 \\
26
- \quad Traj (cluster dynamics) & 82.77 & 81.41 & 48.55 & 70.75 & 85.44 & 92.13 \\
27
- \quad Traj (cluster dynamics, soft) & 82.46 & 81.84 & 39.64 & 74.20 & 84.25 & 88.94 \\
28
  \multicolumn{7}{l}{\emph{Distances + trajectory}} \\
29
- \quad \textbf{Profile, no background + Traj (path geometry)} & \textbf{92.83} & 90.71 & 22.35 & 85.21 & 95.25 & 98.03 \\
30
- \quad MD, all layers + Traj (path geometry) & 92.80 & 90.68 & 22.52 & 85.14 & 95.24 & 98.01 \\
31
- \quad Profile + Traj (path geometry) & 90.99 & 88.69 & 27.12 & 82.52 & 93.43 & 97.02 \\
32
- \quad Profile + Traj (cluster dynamics) & 84.53 & 83.12 & 43.43 & 72.59 & 87.18 & 93.82 \\
33
- \quad Profile, no background + Traj (cluster dynamics) & 83.33 & 81.92 & 46.34 & 71.28 & 85.99 & 92.71 \\
34
  \multicolumn{7}{l}{\emph{Input only}} \\
35
  \quad \emph{Document length only} & 82.46 & 82.12 & 57.59 & 66.95 & 86.26 & 94.17 \\
36
  \bottomrule
 
1
  \begin{table*}[t]\centering\small
2
+ \caption{Out-of-distribution detection over 48 scenarios (16 configurations $\times$ the 3 other tasks of the grid). Every readout except the last group is one-class: it is fitted on the training split of the in-distribution task only and sees no out-of-distribution example; the correctness labels of that split enter only through the correctness-deviation family, whose reference centroids are fitted on them. Profile, the trajectory families and MT are the columns of Table~\ref{tab:trajfam} and Section~\ref{sec:method}, built by the same code as for misclassification and scored by a single Gaussian instead of the detector; the shift variant of path geometry takes the goal direction to the label-free training centroid and adds the whitened step length (Appendix~\ref{app:traj}). The supervised row is the MT detector of Table~\ref{tab:main}, trained on in-distribution correctness, with its error probability read as a shift score. AUROC and FPR95 use the full in-distribution split; only AUPR uses a size-matched draw, so only its chance level is $50$. Near, mid and far are terciles of a corpus distance measured from raw text alone --- the mean of the Jensen-Shannon divergences between the two length distributions and between the two unigram distributions --- so the axis is a property of the pair rather than of any representation. The six unordered corpus pairs are divided into two pairs per tier. The effective independent units are therefore the corpus pairs, not the 48 encoder--direction scenarios. The best value of each column is bold.}
3
  \label{tab:ood}
4
  \begin{tabular}{lrrrrrr}
5
  \toprule
6
  Readout & AUROC & AUPR & FPR95 & near (16) & mid (16) & far (16) \\
7
  \midrule
8
  \multicolumn{7}{l}{\emph{Logit-based}} \\
9
+ \quad Energy & 70.97 & 68.86 & 58.06 & 58.13 & 71.92 & 82.86 \\
10
+ \quad MSP & 69.00 & 67.35 & 63.62 & 57.14 & 69.55 & 80.30 \\
 
 
11
  \multicolumn{7}{l}{\emph{Final layer}} \\
12
+ \quad MD & 78.89 & 78.60 & 42.82 & 74.34 & 75.77 & 86.57 \\
13
+ \quad $k$-NN & 77.66 & 76.71 & 49.15 & 69.75 & 75.25 & 87.99 \\
14
+ \quad RMD & 66.35 & 64.64 & 73.32 & 61.14 & 67.71 & 70.19 \\
 
15
  \multicolumn{7}{l}{\emph{All layers, distances only}} \\
16
+ \quad MD, all layers & 92.05 & 89.92 & 24.90 & \textbf{84.40} & 94.24 & 97.51 \\
17
+ \quad Profile, no background & 91.43 & 89.49 & 27.51 & 83.57 & 93.52 & 97.20 \\
18
+ \quad Profile & 77.69 & 75.37 & 57.45 & 70.30 & 77.80 & 84.95 \\
 
 
19
  \multicolumn{7}{l}{\emph{Trajectory only}} \\
20
+ \quad Traj (path geometry, shift variant) & 90.64 & 88.62 & 28.57 & 81.11 & 93.53 & 97.29 \\
21
+ \quad Traj (path geometry) & 88.91 & 87.05 & 32.85 & 78.41 & 91.81 & 96.50 \\
22
+ \quad Traj (all three families) & 80.40 & 80.51 & 51.04 & 66.23 & 82.18 & 92.79 \\
23
  \multicolumn{7}{l}{\emph{Distances + trajectory}} \\
24
+ \quad MD, all layers + Traj (path geometry, shift variant) & \textbf{92.24} & \textbf{90.25} & \textbf{23.95} & 83.79 & \textbf{94.90} & \textbf{98.02} \\
25
+ \quad Profile, no background + Traj (path geometry) & 91.50 & 89.63 & 25.98 & 82.43 & 94.31 & 97.76 \\
26
+ \multicolumn{7}{l}{\emph{Supervised (in-distribution correctness)}} \\
27
+ \quad MT detector, $P(\text{error})$ & 66.27 & 62.47 & 71.27 & 56.25 & 64.73 & 77.84 \\
 
28
  \multicolumn{7}{l}{\emph{Input only}} \\
29
  \quad \emph{Document length only} & 82.46 & 82.12 & 57.59 & 66.95 & 86.26 & 94.17 \\
30
  \bottomrule
output/generation/assets/tables/tab_percfg.tex CHANGED
@@ -1,28 +1,28 @@
1
  \begin{table*}[t]\centering\small
2
- \caption{Misclassification ROC-AUC (\%) per configuration, mean over five detector seeds. Best in each row in bold. Rows where a method is not defined are left blank rather than imputed.}
3
  \label{tab:percfg}
4
- \begin{tabular}{llrrrrrrrrr}
5
  \toprule
6
- Encoder & Task & MD & RMD & RMDml & AttnP$_\text{l}$ & LastAll & AttnTopo & Traj. & Profile & \textbf{MT} \\
7
  \midrule
8
- RoBERTa & CoLA & 46.31 & 47.57 & 74.44 & 76.71 & 76.72 & 77.22 & 77.69 & 76.33 & \textbf{77.81} \\
9
- & SST-5 & 56.19 & 54.64 & 55.41 & 59.47 & 59.97 & 55.29 & 59.23 & 60.52 & \textbf{61.17} \\
10
- & ToxiGen & 60.83 & 62.14 & 68.30 & 69.10 & 69.75 & 70.10 & 71.57 & 70.88 & \textbf{72.00} \\
11
- & Yelp & 68.62 & 80.23 & 81.31 & 79.68 & 81.47 & 80.23 & 80.05 & \textbf{81.97} & 81.58 \\
12
- DeBERTa & CoLA & 49.89 & 59.37 & 78.46 & 77.95 & 78.99 & 77.59 & 79.58 & \textbf{80.14} & 79.86 \\
13
- & SST-5 & 55.71 & 53.36 & 53.85 & 55.88 & 60.80 & 53.84 & 57.49 & 60.93 & \textbf{62.08} \\
14
- & ToxiGen & 50.32 & 62.80 & 73.45 & \textbf{74.18} & 72.20 & 73.79 & 71.94 & 71.00 & 71.61 \\
15
- & Yelp & 64.45 & 81.79 & 82.66 & 82.21 & 82.87 & 81.92 & 81.68 & \textbf{82.97} & 82.73 \\
16
- DeBERTa-s & CoLA & 67.23 & 63.83 & 78.38 & 79.50 & 80.91 & 79.08 & 81.49 & 80.83 & \textbf{81.54} \\
17
- & SST-5 & 54.27 & 57.69 & 60.65 & 61.72 & 64.47 & 60.02 & 62.52 & 64.85 & \textbf{65.09} \\
18
- & ToxiGen & 49.99 & 66.12 & 72.61 & 68.86 & 73.14 & 72.43 & 73.78 & 73.84 & \textbf{74.06} \\
19
- & Yelp & 59.33 & 81.06 & 81.40 & 78.66 & 81.61 & 79.55 & 80.36 & 81.64 & \textbf{81.82} \\
20
- ELECTRA & CoLA & 61.14 & 55.49 & 78.59 & 79.94 & 79.39 & 80.55 & \textbf{80.91} & 79.61 & 80.86 \\
21
- & SST-5 & 55.32 & 48.40 & 60.39 & 61.42 & 61.06 & 57.67 & 58.52 & 61.79 & \textbf{62.39} \\
22
- & ToxiGen & 65.29 & 61.19 & 70.01 & 71.07 & 70.09 & 69.84 & \textbf{71.24} & 70.41 & 70.92 \\
23
- & Yelp & 66.03 & 81.28 & 81.71 & 80.92 & 82.20 & 80.59 & 81.16 & 82.46 & \textbf{82.56} \\
24
  \midrule
25
- \multicolumn{2}{l}{\textbf{Mean}} & 58.18 & 63.56 & 71.98 & 72.33 & 73.48 & 71.86 & 73.08 & 73.76 & \textbf{74.25} \\
26
  \bottomrule
27
  \end{tabular}
28
  \end{table*}
 
1
  \begin{table*}[t]\centering\small
2
+ \caption{Misclassification ROC-AUC (\%) per configuration, mean over three fine-tuning seeds $\times$ five detector seeds; every entry is the mean over the three fine-tuning runs. Best in each row in bold. Rows where a method is not defined are left blank rather than imputed.}
3
  \label{tab:percfg}
4
+ \begin{tabular}{llrrrrrrr}
5
  \toprule
6
+ Encoder & Task & MD & RMD & RMDml & AttnP$_\text{l}$ & Traj. & Profile & \textbf{MT} \\
7
  \midrule
8
+ RoBERTa & CoLA & 43.52 & 46.99 & 74.47 & 76.96 & \textbf{77.59} & 76.30 & 77.50 \\
9
+ & SST-5 & 54.16 & 54.13 & 55.50 & 57.84 & 58.89 & 60.72 & \textbf{61.08} \\
10
+ & ToxiGen & 52.37 & 59.43 & 70.59 & 70.57 & 74.54 & 72.97 & \textbf{74.85} \\
11
+ & Yelp & 67.86 & 80.12 & 81.17 & 79.73 & 80.48 & \textbf{81.81} & 81.68 \\
12
+ DeBERTa & CoLA & 55.73 & 60.77 & 79.38 & 78.43 & 79.18 & \textbf{80.35} & 79.66 \\
13
+ & SST-5 & 55.89 & 54.05 & 55.13 & 58.06 & 57.76 & 61.14 & \textbf{62.14} \\
14
+ & ToxiGen & 57.82 & 67.09 & \textbf{74.55} & 73.54 & 73.78 & 73.76 & 74.53 \\
15
+ & Yelp & 67.29 & 81.98 & 82.75 & 81.86 & 81.69 & \textbf{83.09} & 82.88 \\
16
+ DeBERTa-s & CoLA & 55.68 & 62.90 & 76.78 & 78.05 & 79.99 & 78.86 & \textbf{80.05} \\
17
+ & SST-5 & 55.01 & 57.64 & 58.86 & 61.02 & 61.97 & 63.83 & \textbf{64.27} \\
18
+ & ToxiGen & 54.65 & 65.10 & 70.93 & 68.24 & 71.94 & 72.00 & \textbf{72.44} \\
19
+ & Yelp & 61.04 & 81.04 & 81.32 & 79.30 & 80.34 & 81.72 & \textbf{81.80} \\
20
+ ELECTRA & CoLA & 65.16 & 55.97 & 78.74 & 80.61 & 80.70 & 79.55 & \textbf{80.94} \\
21
+ & SST-5 & 53.87 & 47.29 & 59.93 & 61.87 & 58.64 & 62.66 & \textbf{62.69} \\
22
+ & ToxiGen & 67.41 & 62.54 & 71.83 & 70.90 & \textbf{73.63} & 72.13 & 73.26 \\
23
+ & Yelp & 67.79 & 81.37 & 81.81 & 80.62 & 81.35 & 82.49 & \textbf{82.71} \\
24
  \midrule
25
+ \multicolumn{2}{l}{\textbf{Mean}} & 58.45 & 63.65 & 72.11 & 72.35 & 73.28 & 73.96 & \textbf{74.53} \\
26
  \bottomrule
27
  \end{tabular}
28
  \end{table*}
output/generation/assets/tables/tab_percfg_full.tex CHANGED
@@ -1,39 +1,39 @@
1
  \begin{table*}[t]\centering\scriptsize
2
  \setlength{\tabcolsep}{2pt}
3
- \caption{Misclassification ROC-AUC (\%) for each encoder--dataset configuration. Values are the mean over five detector seeds where a method has seeds and the single deterministic score otherwise, $\pm$ a test-set bootstrap standard deviation over $n=$200 resamples of the seed-averaged score, so every bar in the table is the same quantity; the MEAN row averages the configuration means. Best and second-best values in each row are bold and underlined respectively. Probe suffixes b/m/l denote architecture-specific beginning, middle and final encoder depths. Baselines: SR softmax response, Ent output entropy, DCT DOCTOR, Enr Energy, MaxL MaxLogit, MD Mahalanobis, RMD relative MD, TS TrustScore, CLUE CLUE-2c. LastAll is the final layer in full (raw \texttt{[CLS]}, logits, MD and RMD) read by the same detector as the pools. AttnTopo is the layer-local attention-topology pool, Profile the Mahalanobis profile, Traj the Mahalanobis trajectory, and MT the profile and trajectory together.}
4
  \label{tab:big_per_cfg}
5
  \resizebox{\textwidth}{!}{%
6
- \begin{tabular}{l|ccccc|cccc|cccccc|c|cccc}
7
  \toprule
8
- & \multicolumn{5}{c|}{output-space (softmax/logits)} & \multicolumn{4}{c|}{representation, ref.-based} & \multicolumn{6}{c|}{supervised probes} & \multicolumn{1}{c|}{final layer} & \multicolumn{4}{c}{evaluated feature sets} \\
9
- \cmidrule(lr){2-6}\cmidrule(lr){7-10}\cmidrule(lr){11-16}\cmidrule(lr){17-17}\cmidrule(lr){18-21}
10
- &SR&Ent&DCT&Enr&MaxL&MD&RMD&TS&CLUE&LinP$_\text{b}$&LinP$_\text{m}$&LinP$_\text{l}$&AttP$_\text{b}$&AttP$_\text{m}$&AttP$_\text{l}$&LastAll&AttnTopo&\textbf{Profile}&\textbf{Traj}&\textbf{MT} \\
11
  \midrule
12
- \multicolumn{21}{l}{\textsc{roberta}-base} \\
13
- CoLA & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.0\!\pm\!1.5$ & $75.3\!\pm\!1.5$ & $46.3\!\pm\!2.0$ & $47.6\!\pm\!2.3$ & $53.1\!\pm\!2.4$ & $73.1\!\pm\!1.6$ & $63.6\!\pm\!2.2$ & $69.6\!\pm\!1.6$ & $76.2\!\pm\!1.5$ & $67.4\!\pm\!1.7$ & $73.2\!\pm\!1.6$ & $76.7\!\pm\!1.5$ & $76.7\!\pm\!1.4$ & $77.2\!\pm\!1.5$ & $76.3\!\pm\!1.6$ & \underline{$77.7\!\pm\!1.5$} & $\mathbf{77.8\!\pm\!1.5}$ \\
14
- SST-5 & $60.3\!\pm\!1.1$ & $59.4\!\pm\!1.1$ & $60.0\!\pm\!1.0$ & $58.1\!\pm\!1.1$ & $59.8\!\pm\!1.1$ & $56.2\!\pm\!1.2$ & $54.6\!\pm\!1.3$ & $54.8\!\pm\!1.2$ & $52.1\!\pm\!1.3$ & $55.5\!\pm\!1.3$ & $58.5\!\pm\!1.2$ & $57.8\!\pm\!1.2$ & $54.4\!\pm\!1.3$ & $58.0\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.0\!\pm\!1.1$ & $55.3\!\pm\!1.2$ & \underline{$60.5\!\pm\!1.1$} & $59.2\!\pm\!1.1$ & $\mathbf{61.2\!\pm\!1.1}$ \\
15
- Tox & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.5\!\pm\!2.2$ & $68.8\!\pm\!2.2$ & $60.8\!\pm\!2.4$ & $62.1\!\pm\!2.4$ & $66.9\!\pm\!2.3$ & $57.2\!\pm\!2.1$ & $61.9\!\pm\!2.1$ & $67.0\!\pm\!2.2$ & $68.5\!\pm\!2.0$ & $65.7\!\pm\!2.1$ & $64.2\!\pm\!2.2$ & $69.1\!\pm\!2.0$ & $69.8\!\pm\!2.2$ & $70.1\!\pm\!2.0$ & $70.9\!\pm\!2.1$ & \underline{$71.6\!\pm\!2.1$} & $\mathbf{72.0\!\pm\!2.1}$ \\
16
- Yelp & $80.3\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $80.3\!\pm\!0.3$ & $76.4\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $68.6\!\pm\!0.4$ & $80.2\!\pm\!0.3$ & $78.0\!\pm\!0.3$ & $67.8\!\pm\!0.3$ & $66.9\!\pm\!0.4$ & $75.0\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $75.7\!\pm\!0.3$ & $79.7\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $\mathbf{82.0\!\pm\!0.3}$ & $80.0\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} \\
17
  \midrule
18
- \multicolumn{21}{l}{\textsc{deberta}-v3-base} \\
19
- CoLA & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $76.2\!\pm\!1.9$ & $76.0\!\pm\!1.9$ & $49.9\!\pm\!2.7$ & $59.4\!\pm\!2.7$ & $54.0\!\pm\!2.9$ & $69.9\!\pm\!2.0$ & $62.7\!\pm\!2.7$ & $66.5\!\pm\!2.2$ & $79.2\!\pm\!1.8$ & $64.2\!\pm\!2.2$ & $68.1\!\pm\!2.1$ & $77.9\!\pm\!1.9$ & $79.0\!\pm\!1.7$ & $77.6\!\pm\!1.8$ & $\mathbf{80.1\!\pm\!1.7}$ & $79.6\!\pm\!1.7$ & \underline{$79.9\!\pm\!1.7$} \\
20
- SST-5 & $59.5\!\pm\!1.4$ & $58.8\!\pm\!1.3$ & $59.3\!\pm\!1.3$ & $58.1\!\pm\!1.3$ & $59.6\!\pm\!1.3$ & $55.7\!\pm\!1.2$ & $53.4\!\pm\!1.3$ & $53.9\!\pm\!1.2$ & $52.2\!\pm\!1.3$ & $53.0\!\pm\!1.2$ & $56.2\!\pm\!1.2$ & $55.8\!\pm\!1.2$ & $56.1\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $60.8\!\pm\!1.2$ & $53.8\!\pm\!1.1$ & \underline{$60.9\!\pm\!1.3$} & $57.5\!\pm\!1.3$ & $\mathbf{62.1\!\pm\!1.2}$ \\
21
- Tox & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.8\!\pm\!2.1$ & $67.7\!\pm\!2.1$ & $50.3\!\pm\!2.2$ & $62.8\!\pm\!2.3$ & $52.6\!\pm\!2.5$ & $62.8\!\pm\!2.0$ & $63.3\!\pm\!2.2$ & $66.3\!\pm\!2.0$ & $69.5\!\pm\!2.2$ & $63.7\!\pm\!1.9$ & $66.7\!\pm\!1.8$ & $\mathbf{74.2\!\pm\!2.1}$ & $72.2\!\pm\!2.1$ & \underline{$73.8\!\pm\!2.0$} & $71.0\!\pm\!2.1$ & $71.9\!\pm\!2.0$ & $71.6\!\pm\!2.1$ \\
22
- Yelp & $81.6\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $80.0\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $64.4\!\pm\!0.4$ & $81.8\!\pm\!0.3$ & $75.3\!\pm\!0.4$ & $67.0\!\pm\!0.3$ & $65.7\!\pm\!0.4$ & $71.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $76.8\!\pm\!0.3$ & $82.2\!\pm\!0.2$ & \underline{$82.9\!\pm\!0.3$} & $81.9\!\pm\!0.3$ & $\mathbf{83.0\!\pm\!0.3}$ & $81.7\!\pm\!0.3$ & $82.7\!\pm\!0.3$ \\
23
  \midrule
24
- \multicolumn{21}{l}{\textsc{deberta}-v3-small} \\
25
- CoLA & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $75.6\!\pm\!1.9$ & $75.9\!\pm\!1.9$ & $67.2\!\pm\!1.9$ & $63.8\!\pm\!2.5$ & $61.9\!\pm\!2.4$ & $70.1\!\pm\!1.9$ & $65.1\!\pm\!2.4$ & $66.8\!\pm\!2.0$ & $79.7\!\pm\!1.7$ & $62.9\!\pm\!2.2$ & $68.0\!\pm\!2.1$ & $79.5\!\pm\!1.6$ & $80.9\!\pm\!1.6$ & $79.1\!\pm\!1.6$ & $80.8\!\pm\!1.6$ & \underline{$81.5\!\pm\!1.4$} & $\mathbf{81.5\!\pm\!1.4}$ \\
26
- SST-5 & $59.0\!\pm\!1.2$ & $59.1\!\pm\!1.2$ & $58.9\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.9\!\pm\!1.2$ & $54.3\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $50.4\!\pm\!1.2$ & $51.2\!\pm\!1.2$ & $57.2\!\pm\!1.3$ & $58.2\!\pm\!1.2$ & $61.8\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $61.0\!\pm\!1.2$ & $61.7\!\pm\!1.3$ & $64.5\!\pm\!1.2$ & $60.0\!\pm\!1.2$ & \underline{$64.9\!\pm\!1.2$} & $62.5\!\pm\!1.1$ & $\mathbf{65.1\!\pm\!1.1}$ \\
27
- Tox & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $66.4\!\pm\!1.9$ & $67.9\!\pm\!1.9$ & $50.0\!\pm\!2.3$ & $66.1\!\pm\!2.2$ & $57.5\!\pm\!2.2$ & $59.8\!\pm\!2.0$ & $61.7\!\pm\!2.2$ & $59.6\!\pm\!2.2$ & $70.4\!\pm\!1.9$ & $62.4\!\pm\!2.0$ & $61.8\!\pm\!2.0$ & $68.9\!\pm\!1.8$ & $73.1\!\pm\!2.0$ & $72.4\!\pm\!1.9$ & \underline{$73.8\!\pm\!1.9$} & $73.8\!\pm\!1.9$ & $\mathbf{74.1\!\pm\!1.9}$ \\
28
- Yelp & $79.2\!\pm\!0.3$ & $78.8\!\pm\!0.3$ & $79.1\!\pm\!0.3$ & $77.0\!\pm\!0.3$ & $78.3\!\pm\!0.3$ & $59.3\!\pm\!0.4$ & $81.1\!\pm\!0.3$ & $71.8\!\pm\!0.4$ & $70.5\!\pm\!0.4$ & $67.3\!\pm\!0.3$ & $69.4\!\pm\!0.4$ & $79.3\!\pm\!0.3$ & $70.8\!\pm\!0.3$ & $75.1\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} & $80.4\!\pm\!0.3$ & $\mathbf{81.8\!\pm\!0.3}$ \\
29
  \midrule
30
- \multicolumn{21}{l}{\textsc{electra}-base} \\
31
- CoLA & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $72.7\!\pm\!1.8$ & $75.5\!\pm\!1.8$ & $61.1\!\pm\!2.4$ & $55.5\!\pm\!2.5$ & $67.3\!\pm\!2.2$ & $71.4\!\pm\!1.7$ & $63.7\!\pm\!2.4$ & $68.5\!\pm\!2.0$ & $79.3\!\pm\!1.8$ & $67.1\!\pm\!2.1$ & $69.9\!\pm\!2.0$ & $79.9\!\pm\!1.7$ & $79.4\!\pm\!1.8$ & $80.6\!\pm\!1.6$ & $79.6\!\pm\!1.8$ & $\mathbf{80.9\!\pm\!1.7}$ & \underline{$80.9\!\pm\!1.8$} \\
32
- SST-5 & $58.7\!\pm\!1.2$ & $55.0\!\pm\!1.2$ & $55.1\!\pm\!1.2$ & $55.1\!\pm\!1.3$ & $58.7\!\pm\!1.2$ & $55.3\!\pm\!1.2$ & $48.4\!\pm\!1.3$ & $54.2\!\pm\!1.3$ & $47.0\!\pm\!1.2$ & $56.8\!\pm\!1.2$ & $58.3\!\pm\!1.2$ & $59.8\!\pm\!1.3$ & $57.5\!\pm\!1.2$ & $60.7\!\pm\!1.2$ & $61.4\!\pm\!1.2$ & $61.1\!\pm\!1.2$ & $57.7\!\pm\!1.2$ & \underline{$61.8\!\pm\!1.2$} & $58.5\!\pm\!1.1$ & $\mathbf{62.4\!\pm\!1.1}$ \\
33
- Tox & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.1\!\pm\!2.2$ & $66.5\!\pm\!2.2$ & $65.3\!\pm\!2.1$ & $61.2\!\pm\!2.3$ & $65.2\!\pm\!2.2$ & $57.3\!\pm\!2.1$ & $62.5\!\pm\!2.2$ & $60.4\!\pm\!2.2$ & $69.1\!\pm\!1.9$ & $64.3\!\pm\!2.0$ & $67.1\!\pm\!2.1$ & \underline{$71.1\!\pm\!1.9$} & $70.1\!\pm\!2.1$ & $69.8\!\pm\!1.9$ & $70.4\!\pm\!2.1$ & $\mathbf{71.2\!\pm\!2.0}$ & $70.9\!\pm\!2.0$ \\
34
- Yelp & $77.0\!\pm\!0.3$ & $75.5\!\pm\!0.3$ & $76.6\!\pm\!0.3$ & $72.4\!\pm\!0.4$ & $76.8\!\pm\!0.3$ & $66.0\!\pm\!0.4$ & $81.3\!\pm\!0.3$ & $75.4\!\pm\!0.3$ & $67.0\!\pm\!0.3$ & $66.7\!\pm\!0.4$ & $71.6\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $70.9\!\pm\!0.3$ & $75.9\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $82.2\!\pm\!0.3$ & $80.6\!\pm\!0.3$ & \underline{$82.5\!\pm\!0.3$} & $81.2\!\pm\!0.3$ & $\mathbf{82.6\!\pm\!0.3}$ \\
35
  \midrule
36
- \textbf{MEAN (16 configurations)} & $70.8$ & $70.3$ & $70.5$ & $69.1$ & $70.4$ & $58.2$ & $63.6$ & $62.0$ & $62.3$ & $62.1$ & $65.2$ & $71.8$ & $64.3$ & $67.4$ & $72.3$ & $73.5$ & $71.9$ & \underline{$73.8$} & $73.1$ & $\mathbf{74.3}$ \\
37
  \bottomrule\end{tabular}%
38
  }
39
  \end{table*}
 
1
  \begin{table*}[t]\centering\scriptsize
2
  \setlength{\tabcolsep}{2pt}
3
+ \caption{Misclassification ROC-AUC (\%) for each encoder--dataset configuration. Values are the mean over five detector seeds where a method has seeds and the single deterministic score otherwise, $\pm$ a test-set bootstrap standard deviation over $n=$200 resamples of the seed-averaged score, so every bar in the table is the same quantity; the MEAN row averages the configuration means. Best and second-best values in each row are bold and underlined respectively. Probe suffixes b/m/l denote architecture-specific beginning, middle and final encoder depths. Baselines: SR softmax response, Ent output entropy, DCT DOCTOR, Enr Energy, MaxL MaxLogit, MD Mahalanobis, RMD relative MD, TS TrustScore, CLUE CLUE-2c. Profile is the Mahalanobis profile, Traj the Mahalanobis trajectory, and MT the profile and trajectory together.}
4
  \label{tab:big_per_cfg}
5
  \resizebox{\textwidth}{!}{%
6
+ \begin{tabular}{l|ccccc|cccc|cccccc|c|ccc}
7
  \toprule
8
+ & \multicolumn{5}{c|}{output-space (softmax/logits)} & \multicolumn{4}{c|}{representation, ref.-based} & \multicolumn{6}{c|}{supervised probes} & \multicolumn{1}{c|}{final layer} & \multicolumn{3}{c}{evaluated feature sets} \\
9
+ \cmidrule(lr){2-6}\cmidrule(lr){7-10}\cmidrule(lr){11-16}\cmidrule(lr){17-17}\cmidrule(lr){18-20}
10
+ &SR&Ent&DCT&Enr&MaxL&MD&RMD&TS&CLUE&LinP$_\text{b}$&LinP$_\text{m}$&LinP$_\text{l}$&AttP$_\text{b}$&AttP$_\text{m}$&AttP$_\text{l}$&\textbf{Profile}&\textbf{Traj}&\textbf{MT} \\
11
  \midrule
12
+ \multicolumn{19}{l}{\textsc{roberta}-base} \\
13
+ CoLA & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.0\!\pm\!1.5$ & $75.3\!\pm\!1.5$ & $46.3\!\pm\!2.0$ & $47.6\!\pm\!2.3$ & $53.1\!\pm\!2.4$ & $73.1\!\pm\!1.6$ & $63.6\!\pm\!2.2$ & $69.6\!\pm\!1.6$ & $76.2\!\pm\!1.5$ & $67.4\!\pm\!1.7$ & $73.2\!\pm\!1.6$ & $76.7\!\pm\!1.5$ & $76.3\!\pm\!1.6$ & \underline{$77.7\!\pm\!1.5$} & $\mathbf{77.8\!\pm\!1.5}$ \\
14
+ SST-5 & $60.3\!\pm\!1.1$ & $59.4\!\pm\!1.1$ & $60.0\!\pm\!1.0$ & $58.1\!\pm\!1.1$ & $59.8\!\pm\!1.1$ & $56.2\!\pm\!1.2$ & $54.6\!\pm\!1.3$ & $54.8\!\pm\!1.2$ & $52.1\!\pm\!1.3$ & $55.5\!\pm\!1.3$ & $58.5\!\pm\!1.2$ & $57.8\!\pm\!1.2$ & $54.4\!\pm\!1.3$ & $58.0\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & \underline{$60.5\!\pm\!1.1$} & $59.2\!\pm\!1.1$ & $\mathbf{61.2\!\pm\!1.1}$ \\
15
+ Tox & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.5\!\pm\!2.2$ & $68.8\!\pm\!2.2$ & $60.8\!\pm\!2.4$ & $62.1\!\pm\!2.4$ & $66.9\!\pm\!2.3$ & $57.2\!\pm\!2.1$ & $61.9\!\pm\!2.1$ & $67.0\!\pm\!2.2$ & $68.5\!\pm\!2.0$ & $65.7\!\pm\!2.1$ & $64.2\!\pm\!2.2$ & $69.1\!\pm\!2.0$ & $70.9\!\pm\!2.1$ & \underline{$71.6\!\pm\!2.1$} & $\mathbf{72.0\!\pm\!2.1}$ \\
16
+ Yelp & $80.3\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $80.3\!\pm\!0.3$ & $76.4\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $68.6\!\pm\!0.4$ & $80.2\!\pm\!0.3$ & $78.0\!\pm\!0.3$ & $67.8\!\pm\!0.3$ & $66.9\!\pm\!0.4$ & $75.0\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $75.7\!\pm\!0.3$ & $79.7\!\pm\!0.3$ & $\mathbf{82.0\!\pm\!0.3}$ & $80.0\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} \\
17
  \midrule
18
+ \multicolumn{19}{l}{\textsc{deberta}-v3-base} \\
19
+ CoLA & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $76.2\!\pm\!1.9$ & $76.0\!\pm\!1.9$ & $49.9\!\pm\!2.7$ & $59.4\!\pm\!2.7$ & $54.0\!\pm\!2.9$ & $69.9\!\pm\!2.0$ & $62.7\!\pm\!2.7$ & $66.5\!\pm\!2.2$ & $79.2\!\pm\!1.8$ & $64.2\!\pm\!2.2$ & $68.1\!\pm\!2.1$ & $77.9\!\pm\!1.9$ & $\mathbf{80.1\!\pm\!1.7}$ & $79.6\!\pm\!1.7$ & \underline{$79.9\!\pm\!1.7$} \\
20
+ SST-5 & $59.5\!\pm\!1.4$ & $58.8\!\pm\!1.3$ & $59.3\!\pm\!1.3$ & $58.1\!\pm\!1.3$ & $59.6\!\pm\!1.3$ & $55.7\!\pm\!1.2$ & $53.4\!\pm\!1.3$ & $53.9\!\pm\!1.2$ & $52.2\!\pm\!1.3$ & $53.0\!\pm\!1.2$ & $56.2\!\pm\!1.2$ & $55.8\!\pm\!1.2$ & $56.1\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & \underline{$60.9\!\pm\!1.3$} & $57.5\!\pm\!1.3$ & $\mathbf{62.1\!\pm\!1.2}$ \\
21
+ Tox & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.8\!\pm\!2.1$ & $67.7\!\pm\!2.1$ & $50.3\!\pm\!2.2$ & $62.8\!\pm\!2.3$ & $52.6\!\pm\!2.5$ & $62.8\!\pm\!2.0$ & $63.3\!\pm\!2.2$ & $66.3\!\pm\!2.0$ & $69.5\!\pm\!2.2$ & $63.7\!\pm\!1.9$ & $66.7\!\pm\!1.8$ & $\mathbf{74.2\!\pm\!2.1}$ & $71.0\!\pm\!2.1$ & \underline{$71.9\!\pm\!2.0$} & $71.6\!\pm\!2.1$ \\
22
+ Yelp & $81.6\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $80.0\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $64.4\!\pm\!0.4$ & $81.8\!\pm\!0.3$ & $75.3\!\pm\!0.4$ & $67.0\!\pm\!0.3$ & $65.7\!\pm\!0.4$ & $71.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $76.8\!\pm\!0.3$ & $82.2\!\pm\!0.2$ & $\mathbf{83.0\!\pm\!0.3}$ & $81.7\!\pm\!0.3$ & \underline{$82.7\!\pm\!0.3$} \\
23
  \midrule
24
+ \multicolumn{19}{l}{\textsc{deberta}-v3-small} \\
25
+ CoLA & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $75.6\!\pm\!1.9$ & $75.9\!\pm\!1.9$ & $67.2\!\pm\!1.9$ & $63.8\!\pm\!2.5$ & $61.9\!\pm\!2.4$ & $70.1\!\pm\!1.9$ & $65.1\!\pm\!2.4$ & $66.8\!\pm\!2.0$ & $79.7\!\pm\!1.7$ & $62.9\!\pm\!2.2$ & $68.0\!\pm\!2.1$ & $79.5\!\pm\!1.6$ & $80.8\!\pm\!1.6$ & \underline{$81.5\!\pm\!1.4$} & $\mathbf{81.5\!\pm\!1.4}$ \\
26
+ SST-5 & $59.0\!\pm\!1.2$ & $59.1\!\pm\!1.2$ & $58.9\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.9\!\pm\!1.2$ & $54.3\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $50.4\!\pm\!1.2$ & $51.2\!\pm\!1.2$ & $57.2\!\pm\!1.3$ & $58.2\!\pm\!1.2$ & $61.8\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $61.0\!\pm\!1.2$ & $61.7\!\pm\!1.3$ & \underline{$64.9\!\pm\!1.2$} & $62.5\!\pm\!1.1$ & $\mathbf{65.1\!\pm\!1.1}$ \\
27
+ Tox & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $66.4\!\pm\!1.9$ & $67.9\!\pm\!1.9$ & $50.0\!\pm\!2.3$ & $66.1\!\pm\!2.2$ & $57.5\!\pm\!2.2$ & $59.8\!\pm\!2.0$ & $61.7\!\pm\!2.2$ & $59.6\!\pm\!2.2$ & $70.4\!\pm\!1.9$ & $62.4\!\pm\!2.0$ & $61.8\!\pm\!2.0$ & $68.9\!\pm\!1.8$ & \underline{$73.8\!\pm\!1.9$} & $73.8\!\pm\!1.9$ & $\mathbf{74.1\!\pm\!1.9}$ \\
28
+ Yelp & $79.2\!\pm\!0.3$ & $78.8\!\pm\!0.3$ & $79.1\!\pm\!0.3$ & $77.0\!\pm\!0.3$ & $78.3\!\pm\!0.3$ & $59.3\!\pm\!0.4$ & $81.1\!\pm\!0.3$ & $71.8\!\pm\!0.4$ & $70.5\!\pm\!0.4$ & $67.3\!\pm\!0.3$ & $69.4\!\pm\!0.4$ & $79.3\!\pm\!0.3$ & $70.8\!\pm\!0.3$ & $75.1\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} & $80.4\!\pm\!0.3$ & $\mathbf{81.8\!\pm\!0.3}$ \\
29
  \midrule
30
+ \multicolumn{19}{l}{\textsc{electra}-base} \\
31
+ CoLA & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $72.7\!\pm\!1.8$ & $75.5\!\pm\!1.8$ & $61.1\!\pm\!2.4$ & $55.5\!\pm\!2.5$ & $67.3\!\pm\!2.2$ & $71.4\!\pm\!1.7$ & $63.7\!\pm\!2.4$ & $68.5\!\pm\!2.0$ & $79.3\!\pm\!1.8$ & $67.1\!\pm\!2.1$ & $69.9\!\pm\!2.0$ & $79.9\!\pm\!1.7$ & $79.6\!\pm\!1.8$ & $\mathbf{80.9\!\pm\!1.7}$ & \underline{$80.9\!\pm\!1.8$} \\
32
+ SST-5 & $58.7\!\pm\!1.2$ & $55.0\!\pm\!1.2$ & $55.1\!\pm\!1.2$ & $55.1\!\pm\!1.3$ & $58.7\!\pm\!1.2$ & $55.3\!\pm\!1.2$ & $48.4\!\pm\!1.3$ & $54.2\!\pm\!1.3$ & $47.0\!\pm\!1.2$ & $56.8\!\pm\!1.2$ & $58.3\!\pm\!1.2$ & $59.8\!\pm\!1.3$ & $57.5\!\pm\!1.2$ & $60.7\!\pm\!1.2$ & $61.4\!\pm\!1.2$ & \underline{$61.8\!\pm\!1.2$} & $58.5\!\pm\!1.1$ & $\mathbf{62.4\!\pm\!1.1}$ \\
33
+ Tox & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.1\!\pm\!2.2$ & $66.5\!\pm\!2.2$ & $65.3\!\pm\!2.1$ & $61.2\!\pm\!2.3$ & $65.2\!\pm\!2.2$ & $57.3\!\pm\!2.1$ & $62.5\!\pm\!2.2$ & $60.4\!\pm\!2.2$ & $69.1\!\pm\!1.9$ & $64.3\!\pm\!2.0$ & $67.1\!\pm\!2.1$ & \underline{$71.1\!\pm\!1.9$} & $70.4\!\pm\!2.1$ & $\mathbf{71.2\!\pm\!2.0}$ & $70.9\!\pm\!2.0$ \\
34
+ Yelp & $77.0\!\pm\!0.3$ & $75.5\!\pm\!0.3$ & $76.6\!\pm\!0.3$ & $72.4\!\pm\!0.4$ & $76.8\!\pm\!0.3$ & $66.0\!\pm\!0.4$ & $81.3\!\pm\!0.3$ & $75.4\!\pm\!0.3$ & $67.0\!\pm\!0.3$ & $66.7\!\pm\!0.4$ & $71.6\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $70.9\!\pm\!0.3$ & $75.9\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & \underline{$82.5\!\pm\!0.3$} & $81.2\!\pm\!0.3$ & $\mathbf{82.6\!\pm\!0.3}$ \\
35
  \midrule
36
+ \textbf{MEAN (16 configurations)} & $70.8$ & $70.3$ & $70.5$ & $69.1$ & $70.4$ & $58.2$ & $63.6$ & $62.0$ & $62.3$ & $62.1$ & $65.2$ & $71.8$ & $64.3$ & $67.4$ & $72.3$ & \underline{$73.8$} & $73.1$ & $\mathbf{74.3}$ \\
37
  \bottomrule\end{tabular}%
38
  }
39
  \end{table*}
output/generation/assets/tables/tab_percfg_full_rows.tex CHANGED
@@ -1,25 +1,25 @@
1
- \multicolumn{21}{l}{\textsc{roberta}-base} \\
2
- CoLA & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.0\!\pm\!1.5$ & $75.3\!\pm\!1.5$ & $46.3\!\pm\!2.0$ & $47.6\!\pm\!2.3$ & $53.1\!\pm\!2.4$ & $73.1\!\pm\!1.6$ & $63.6\!\pm\!2.2$ & $69.6\!\pm\!1.6$ & $76.2\!\pm\!1.5$ & $67.4\!\pm\!1.7$ & $73.2\!\pm\!1.6$ & $76.7\!\pm\!1.5$ & $76.7\!\pm\!1.4$ & $77.2\!\pm\!1.5$ & $76.3\!\pm\!1.6$ & \underline{$77.7\!\pm\!1.5$} & $\mathbf{77.8\!\pm\!1.5}$ \\
3
- SST-5 & $60.3\!\pm\!1.1$ & $59.4\!\pm\!1.1$ & $60.0\!\pm\!1.0$ & $58.1\!\pm\!1.1$ & $59.8\!\pm\!1.1$ & $56.2\!\pm\!1.2$ & $54.6\!\pm\!1.3$ & $54.8\!\pm\!1.2$ & $52.1\!\pm\!1.3$ & $55.5\!\pm\!1.3$ & $58.5\!\pm\!1.2$ & $57.8\!\pm\!1.2$ & $54.4\!\pm\!1.3$ & $58.0\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.0\!\pm\!1.1$ & $55.3\!\pm\!1.2$ & \underline{$60.5\!\pm\!1.1$} & $59.2\!\pm\!1.1$ & $\mathbf{61.2\!\pm\!1.1}$ \\
4
- Tox & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.5\!\pm\!2.2$ & $68.8\!\pm\!2.2$ & $60.8\!\pm\!2.4$ & $62.1\!\pm\!2.4$ & $66.9\!\pm\!2.3$ & $57.2\!\pm\!2.1$ & $61.9\!\pm\!2.1$ & $67.0\!\pm\!2.2$ & $68.5\!\pm\!2.0$ & $65.7\!\pm\!2.1$ & $64.2\!\pm\!2.2$ & $69.1\!\pm\!2.0$ & $69.8\!\pm\!2.2$ & $70.1\!\pm\!2.0$ & $70.9\!\pm\!2.1$ & \underline{$71.6\!\pm\!2.1$} & $\mathbf{72.0\!\pm\!2.1}$ \\
5
- Yelp & $80.3\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $80.3\!\pm\!0.3$ & $76.4\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $68.6\!\pm\!0.4$ & $80.2\!\pm\!0.3$ & $78.0\!\pm\!0.3$ & $67.8\!\pm\!0.3$ & $66.9\!\pm\!0.4$ & $75.0\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $75.7\!\pm\!0.3$ & $79.7\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $\mathbf{82.0\!\pm\!0.3}$ & $80.0\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} \\
6
  \midrule
7
- \multicolumn{21}{l}{\textsc{deberta}-v3-base} \\
8
- CoLA & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $76.2\!\pm\!1.9$ & $76.0\!\pm\!1.9$ & $49.9\!\pm\!2.7$ & $59.4\!\pm\!2.7$ & $54.0\!\pm\!2.9$ & $69.9\!\pm\!2.0$ & $62.7\!\pm\!2.7$ & $66.5\!\pm\!2.2$ & $79.2\!\pm\!1.8$ & $64.2\!\pm\!2.2$ & $68.1\!\pm\!2.1$ & $77.9\!\pm\!1.9$ & $79.0\!\pm\!1.7$ & $77.6\!\pm\!1.8$ & $\mathbf{80.1\!\pm\!1.7}$ & $79.6\!\pm\!1.7$ & \underline{$79.9\!\pm\!1.7$} \\
9
- SST-5 & $59.5\!\pm\!1.4$ & $58.8\!\pm\!1.3$ & $59.3\!\pm\!1.3$ & $58.1\!\pm\!1.3$ & $59.6\!\pm\!1.3$ & $55.7\!\pm\!1.2$ & $53.4\!\pm\!1.3$ & $53.9\!\pm\!1.2$ & $52.2\!\pm\!1.3$ & $53.0\!\pm\!1.2$ & $56.2\!\pm\!1.2$ & $55.8\!\pm\!1.2$ & $56.1\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $60.8\!\pm\!1.2$ & $53.8\!\pm\!1.1$ & \underline{$60.9\!\pm\!1.3$} & $57.5\!\pm\!1.3$ & $\mathbf{62.1\!\pm\!1.2}$ \\
10
- Tox & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.8\!\pm\!2.1$ & $67.7\!\pm\!2.1$ & $50.3\!\pm\!2.2$ & $62.8\!\pm\!2.3$ & $52.6\!\pm\!2.5$ & $62.8\!\pm\!2.0$ & $63.3\!\pm\!2.2$ & $66.3\!\pm\!2.0$ & $69.5\!\pm\!2.2$ & $63.7\!\pm\!1.9$ & $66.7\!\pm\!1.8$ & $\mathbf{74.2\!\pm\!2.1}$ & $72.2\!\pm\!2.1$ & \underline{$73.8\!\pm\!2.0$} & $71.0\!\pm\!2.1$ & $71.9\!\pm\!2.0$ & $71.6\!\pm\!2.1$ \\
11
- Yelp & $81.6\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $80.0\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $64.4\!\pm\!0.4$ & $81.8\!\pm\!0.3$ & $75.3\!\pm\!0.4$ & $67.0\!\pm\!0.3$ & $65.7\!\pm\!0.4$ & $71.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $76.8\!\pm\!0.3$ & $82.2\!\pm\!0.2$ & \underline{$82.9\!\pm\!0.3$} & $81.9\!\pm\!0.3$ & $\mathbf{83.0\!\pm\!0.3}$ & $81.7\!\pm\!0.3$ & $82.7\!\pm\!0.3$ \\
12
  \midrule
13
- \multicolumn{21}{l}{\textsc{deberta}-v3-small} \\
14
- CoLA & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $75.6\!\pm\!1.9$ & $75.9\!\pm\!1.9$ & $67.2\!\pm\!1.9$ & $63.8\!\pm\!2.5$ & $61.9\!\pm\!2.4$ & $70.1\!\pm\!1.9$ & $65.1\!\pm\!2.4$ & $66.8\!\pm\!2.0$ & $79.7\!\pm\!1.7$ & $62.9\!\pm\!2.2$ & $68.0\!\pm\!2.1$ & $79.5\!\pm\!1.6$ & $80.9\!\pm\!1.6$ & $79.1\!\pm\!1.6$ & $80.8\!\pm\!1.6$ & \underline{$81.5\!\pm\!1.4$} & $\mathbf{81.5\!\pm\!1.4}$ \\
15
- SST-5 & $59.0\!\pm\!1.2$ & $59.1\!\pm\!1.2$ & $58.9\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.9\!\pm\!1.2$ & $54.3\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $50.4\!\pm\!1.2$ & $51.2\!\pm\!1.2$ & $57.2\!\pm\!1.3$ & $58.2\!\pm\!1.2$ & $61.8\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $61.0\!\pm\!1.2$ & $61.7\!\pm\!1.3$ & $64.5\!\pm\!1.2$ & $60.0\!\pm\!1.2$ & \underline{$64.9\!\pm\!1.2$} & $62.5\!\pm\!1.1$ & $\mathbf{65.1\!\pm\!1.1}$ \\
16
- Tox & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $66.4\!\pm\!1.9$ & $67.9\!\pm\!1.9$ & $50.0\!\pm\!2.3$ & $66.1\!\pm\!2.2$ & $57.5\!\pm\!2.2$ & $59.8\!\pm\!2.0$ & $61.7\!\pm\!2.2$ & $59.6\!\pm\!2.2$ & $70.4\!\pm\!1.9$ & $62.4\!\pm\!2.0$ & $61.8\!\pm\!2.0$ & $68.9\!\pm\!1.8$ & $73.1\!\pm\!2.0$ & $72.4\!\pm\!1.9$ & \underline{$73.8\!\pm\!1.9$} & $73.8\!\pm\!1.9$ & $\mathbf{74.1\!\pm\!1.9}$ \\
17
- Yelp & $79.2\!\pm\!0.3$ & $78.8\!\pm\!0.3$ & $79.1\!\pm\!0.3$ & $77.0\!\pm\!0.3$ & $78.3\!\pm\!0.3$ & $59.3\!\pm\!0.4$ & $81.1\!\pm\!0.3$ & $71.8\!\pm\!0.4$ & $70.5\!\pm\!0.4$ & $67.3\!\pm\!0.3$ & $69.4\!\pm\!0.4$ & $79.3\!\pm\!0.3$ & $70.8\!\pm\!0.3$ & $75.1\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} & $80.4\!\pm\!0.3$ & $\mathbf{81.8\!\pm\!0.3}$ \\
18
  \midrule
19
- \multicolumn{21}{l}{\textsc{electra}-base} \\
20
- CoLA & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $72.7\!\pm\!1.8$ & $75.5\!\pm\!1.8$ & $61.1\!\pm\!2.4$ & $55.5\!\pm\!2.5$ & $67.3\!\pm\!2.2$ & $71.4\!\pm\!1.7$ & $63.7\!\pm\!2.4$ & $68.5\!\pm\!2.0$ & $79.3\!\pm\!1.8$ & $67.1\!\pm\!2.1$ & $69.9\!\pm\!2.0$ & $79.9\!\pm\!1.7$ & $79.4\!\pm\!1.8$ & $80.6\!\pm\!1.6$ & $79.6\!\pm\!1.8$ & $\mathbf{80.9\!\pm\!1.7}$ & \underline{$80.9\!\pm\!1.8$} \\
21
- SST-5 & $58.7\!\pm\!1.2$ & $55.0\!\pm\!1.2$ & $55.1\!\pm\!1.2$ & $55.1\!\pm\!1.3$ & $58.7\!\pm\!1.2$ & $55.3\!\pm\!1.2$ & $48.4\!\pm\!1.3$ & $54.2\!\pm\!1.3$ & $47.0\!\pm\!1.2$ & $56.8\!\pm\!1.2$ & $58.3\!\pm\!1.2$ & $59.8\!\pm\!1.3$ & $57.5\!\pm\!1.2$ & $60.7\!\pm\!1.2$ & $61.4\!\pm\!1.2$ & $61.1\!\pm\!1.2$ & $57.7\!\pm\!1.2$ & \underline{$61.8\!\pm\!1.2$} & $58.5\!\pm\!1.1$ & $\mathbf{62.4\!\pm\!1.1}$ \\
22
- Tox & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.1\!\pm\!2.2$ & $66.5\!\pm\!2.2$ & $65.3\!\pm\!2.1$ & $61.2\!\pm\!2.3$ & $65.2\!\pm\!2.2$ & $57.3\!\pm\!2.1$ & $62.5\!\pm\!2.2$ & $60.4\!\pm\!2.2$ & $69.1\!\pm\!1.9$ & $64.3\!\pm\!2.0$ & $67.1\!\pm\!2.1$ & \underline{$71.1\!\pm\!1.9$} & $70.1\!\pm\!2.1$ & $69.8\!\pm\!1.9$ & $70.4\!\pm\!2.1$ & $\mathbf{71.2\!\pm\!2.0}$ & $70.9\!\pm\!2.0$ \\
23
- Yelp & $77.0\!\pm\!0.3$ & $75.5\!\pm\!0.3$ & $76.6\!\pm\!0.3$ & $72.4\!\pm\!0.4$ & $76.8\!\pm\!0.3$ & $66.0\!\pm\!0.4$ & $81.3\!\pm\!0.3$ & $75.4\!\pm\!0.3$ & $67.0\!\pm\!0.3$ & $66.7\!\pm\!0.4$ & $71.6\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $70.9\!\pm\!0.3$ & $75.9\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $82.2\!\pm\!0.3$ & $80.6\!\pm\!0.3$ & \underline{$82.5\!\pm\!0.3$} & $81.2\!\pm\!0.3$ & $\mathbf{82.6\!\pm\!0.3}$ \\
24
  \midrule
25
- \textbf{MEAN (16 configurations)} & $70.8$ & $70.3$ & $70.5$ & $69.1$ & $70.4$ & $58.2$ & $63.6$ & $62.0$ & $62.3$ & $62.1$ & $65.2$ & $71.8$ & $64.3$ & $67.4$ & $72.3$ & $73.5$ & $71.9$ & \underline{$73.8$} & $73.1$ & $\mathbf{74.3}$ \\
 
1
+ \multicolumn{19}{l}{\textsc{roberta}-base} \\
2
+ CoLA & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.2\!\pm\!1.5$ & $75.0\!\pm\!1.5$ & $75.3\!\pm\!1.5$ & $46.3\!\pm\!2.0$ & $47.6\!\pm\!2.3$ & $53.1\!\pm\!2.4$ & $73.1\!\pm\!1.6$ & $63.6\!\pm\!2.2$ & $69.6\!\pm\!1.6$ & $76.2\!\pm\!1.5$ & $67.4\!\pm\!1.7$ & $73.2\!\pm\!1.6$ & $76.7\!\pm\!1.5$ & $76.3\!\pm\!1.6$ & \underline{$77.7\!\pm\!1.5$} & $\mathbf{77.8\!\pm\!1.5}$ \\
3
+ SST-5 & $60.3\!\pm\!1.1$ & $59.4\!\pm\!1.1$ & $60.0\!\pm\!1.0$ & $58.1\!\pm\!1.1$ & $59.8\!\pm\!1.1$ & $56.2\!\pm\!1.2$ & $54.6\!\pm\!1.3$ & $54.8\!\pm\!1.2$ & $52.1\!\pm\!1.3$ & $55.5\!\pm\!1.3$ & $58.5\!\pm\!1.2$ & $57.8\!\pm\!1.2$ & $54.4\!\pm\!1.3$ & $58.0\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & \underline{$60.5\!\pm\!1.1$} & $59.2\!\pm\!1.1$ & $\mathbf{61.2\!\pm\!1.1}$ \\
4
+ Tox & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.9\!\pm\!2.1$ & $68.5\!\pm\!2.2$ & $68.8\!\pm\!2.2$ & $60.8\!\pm\!2.4$ & $62.1\!\pm\!2.4$ & $66.9\!\pm\!2.3$ & $57.2\!\pm\!2.1$ & $61.9\!\pm\!2.1$ & $67.0\!\pm\!2.2$ & $68.5\!\pm\!2.0$ & $65.7\!\pm\!2.1$ & $64.2\!\pm\!2.2$ & $69.1\!\pm\!2.0$ & $70.9\!\pm\!2.1$ & \underline{$71.6\!\pm\!2.1$} & $\mathbf{72.0\!\pm\!2.1}$ \\
5
+ Yelp & $80.3\!\pm\!0.3$ & $80.2\!\pm\!0.3$ & $80.3\!\pm\!0.3$ & $76.4\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & $68.6\!\pm\!0.4$ & $80.2\!\pm\!0.3$ & $78.0\!\pm\!0.3$ & $67.8\!\pm\!0.3$ & $66.9\!\pm\!0.4$ & $75.0\!\pm\!0.3$ & $79.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $75.7\!\pm\!0.3$ & $79.7\!\pm\!0.3$ & $\mathbf{82.0\!\pm\!0.3}$ & $80.0\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} \\
6
  \midrule
7
+ \multicolumn{19}{l}{\textsc{deberta}-v3-base} \\
8
+ CoLA & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $75.4\!\pm\!1.9$ & $76.2\!\pm\!1.9$ & $76.0\!\pm\!1.9$ & $49.9\!\pm\!2.7$ & $59.4\!\pm\!2.7$ & $54.0\!\pm\!2.9$ & $69.9\!\pm\!2.0$ & $62.7\!\pm\!2.7$ & $66.5\!\pm\!2.2$ & $79.2\!\pm\!1.8$ & $64.2\!\pm\!2.2$ & $68.1\!\pm\!2.1$ & $77.9\!\pm\!1.9$ & $\mathbf{80.1\!\pm\!1.7}$ & $79.6\!\pm\!1.7$ & \underline{$79.9\!\pm\!1.7$} \\
9
+ SST-5 & $59.5\!\pm\!1.4$ & $58.8\!\pm\!1.3$ & $59.3\!\pm\!1.3$ & $58.1\!\pm\!1.3$ & $59.6\!\pm\!1.3$ & $55.7\!\pm\!1.2$ & $53.4\!\pm\!1.3$ & $53.9\!\pm\!1.2$ & $52.2\!\pm\!1.3$ & $53.0\!\pm\!1.2$ & $56.2\!\pm\!1.2$ & $55.8\!\pm\!1.2$ & $56.1\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & $55.9\!\pm\!1.3$ & \underline{$60.9\!\pm\!1.3$} & $57.5\!\pm\!1.3$ & $\mathbf{62.1\!\pm\!1.2}$ \\
10
+ Tox & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.5\!\pm\!2.1$ & $67.8\!\pm\!2.1$ & $67.7\!\pm\!2.1$ & $50.3\!\pm\!2.2$ & $62.8\!\pm\!2.3$ & $52.6\!\pm\!2.5$ & $62.8\!\pm\!2.0$ & $63.3\!\pm\!2.2$ & $66.3\!\pm\!2.0$ & $69.5\!\pm\!2.2$ & $63.7\!\pm\!1.9$ & $66.7\!\pm\!1.8$ & $\mathbf{74.2\!\pm\!2.1}$ & $71.0\!\pm\!2.1$ & \underline{$71.9\!\pm\!2.0$} & $71.6\!\pm\!2.1$ \\
11
+ Yelp & $81.6\!\pm\!0.3$ & $81.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $80.0\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $64.4\!\pm\!0.4$ & $81.8\!\pm\!0.3$ & $75.3\!\pm\!0.4$ & $67.0\!\pm\!0.3$ & $65.7\!\pm\!0.4$ & $71.5\!\pm\!0.3$ & $81.6\!\pm\!0.3$ & $71.6\!\pm\!0.3$ & $76.8\!\pm\!0.3$ & $82.2\!\pm\!0.2$ & $\mathbf{83.0\!\pm\!0.3}$ & $81.7\!\pm\!0.3$ & \underline{$82.7\!\pm\!0.3$} \\
12
  \midrule
13
+ \multicolumn{19}{l}{\textsc{deberta}-v3-small} \\
14
+ CoLA & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $77.1\!\pm\!1.8$ & $75.6\!\pm\!1.9$ & $75.9\!\pm\!1.9$ & $67.2\!\pm\!1.9$ & $63.8\!\pm\!2.5$ & $61.9\!\pm\!2.4$ & $70.1\!\pm\!1.9$ & $65.1\!\pm\!2.4$ & $66.8\!\pm\!2.0$ & $79.7\!\pm\!1.7$ & $62.9\!\pm\!2.2$ & $68.0\!\pm\!2.1$ & $79.5\!\pm\!1.6$ & $80.8\!\pm\!1.6$ & \underline{$81.5\!\pm\!1.4$} & $\mathbf{81.5\!\pm\!1.4}$ \\
15
+ SST-5 & $59.0\!\pm\!1.2$ & $59.1\!\pm\!1.2$ & $58.9\!\pm\!1.2$ & $59.5\!\pm\!1.2$ & $60.9\!\pm\!1.2$ & $54.3\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $50.4\!\pm\!1.2$ & $51.2\!\pm\!1.2$ & $57.2\!\pm\!1.3$ & $58.2\!\pm\!1.2$ & $61.8\!\pm\!1.2$ & $57.7\!\pm\!1.1$ & $61.0\!\pm\!1.2$ & $61.7\!\pm\!1.3$ & \underline{$64.9\!\pm\!1.2$} & $62.5\!\pm\!1.1$ & $\mathbf{65.1\!\pm\!1.1}$ \\
16
+ Tox & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $69.2\!\pm\!1.9$ & $66.4\!\pm\!1.9$ & $67.9\!\pm\!1.9$ & $50.0\!\pm\!2.3$ & $66.1\!\pm\!2.2$ & $57.5\!\pm\!2.2$ & $59.8\!\pm\!2.0$ & $61.7\!\pm\!2.2$ & $59.6\!\pm\!2.2$ & $70.4\!\pm\!1.9$ & $62.4\!\pm\!2.0$ & $61.8\!\pm\!2.0$ & $68.9\!\pm\!1.8$ & \underline{$73.8\!\pm\!1.9$} & $73.8\!\pm\!1.9$ & $\mathbf{74.1\!\pm\!1.9}$ \\
17
+ Yelp & $79.2\!\pm\!0.3$ & $78.8\!\pm\!0.3$ & $79.1\!\pm\!0.3$ & $77.0\!\pm\!0.3$ & $78.3\!\pm\!0.3$ & $59.3\!\pm\!0.4$ & $81.1\!\pm\!0.3$ & $71.8\!\pm\!0.4$ & $70.5\!\pm\!0.4$ & $67.3\!\pm\!0.3$ & $69.4\!\pm\!0.4$ & $79.3\!\pm\!0.3$ & $70.8\!\pm\!0.3$ & $75.1\!\pm\!0.3$ & $78.7\!\pm\!0.3$ & \underline{$81.6\!\pm\!0.3$} & $80.4\!\pm\!0.3$ & $\mathbf{81.8\!\pm\!0.3}$ \\
18
  \midrule
19
+ \multicolumn{19}{l}{\textsc{electra}-base} \\
20
+ CoLA & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $76.7\!\pm\!1.8$ & $72.7\!\pm\!1.8$ & $75.5\!\pm\!1.8$ & $61.1\!\pm\!2.4$ & $55.5\!\pm\!2.5$ & $67.3\!\pm\!2.2$ & $71.4\!\pm\!1.7$ & $63.7\!\pm\!2.4$ & $68.5\!\pm\!2.0$ & $79.3\!\pm\!1.8$ & $67.1\!\pm\!2.1$ & $69.9\!\pm\!2.0$ & $79.9\!\pm\!1.7$ & $79.6\!\pm\!1.8$ & $\mathbf{80.9\!\pm\!1.7}$ & \underline{$80.9\!\pm\!1.8$} \\
21
+ SST-5 & $58.7\!\pm\!1.2$ & $55.0\!\pm\!1.2$ & $55.1\!\pm\!1.2$ & $55.1\!\pm\!1.3$ & $58.7\!\pm\!1.2$ & $55.3\!\pm\!1.2$ & $48.4\!\pm\!1.3$ & $54.2\!\pm\!1.3$ & $47.0\!\pm\!1.2$ & $56.8\!\pm\!1.2$ & $58.3\!\pm\!1.2$ & $59.8\!\pm\!1.3$ & $57.5\!\pm\!1.2$ & $60.7\!\pm\!1.2$ & $61.4\!\pm\!1.2$ & \underline{$61.8\!\pm\!1.2$} & $58.5\!\pm\!1.1$ & $\mathbf{62.4\!\pm\!1.1}$ \\
22
+ Tox & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.7\!\pm\!2.1$ & $66.1\!\pm\!2.2$ & $66.5\!\pm\!2.2$ & $65.3\!\pm\!2.1$ & $61.2\!\pm\!2.3$ & $65.2\!\pm\!2.2$ & $57.3\!\pm\!2.1$ & $62.5\!\pm\!2.2$ & $60.4\!\pm\!2.2$ & $69.1\!\pm\!1.9$ & $64.3\!\pm\!2.0$ & $67.1\!\pm\!2.1$ & \underline{$71.1\!\pm\!1.9$} & $70.4\!\pm\!2.1$ & $\mathbf{71.2\!\pm\!2.0}$ & $70.9\!\pm\!2.0$ \\
23
+ Yelp & $77.0\!\pm\!0.3$ & $75.5\!\pm\!0.3$ & $76.6\!\pm\!0.3$ & $72.4\!\pm\!0.4$ & $76.8\!\pm\!0.3$ & $66.0\!\pm\!0.4$ & $81.3\!\pm\!0.3$ & $75.4\!\pm\!0.3$ & $67.0\!\pm\!0.3$ & $66.7\!\pm\!0.4$ & $71.6\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & $70.9\!\pm\!0.3$ & $75.9\!\pm\!0.3$ & $80.9\!\pm\!0.3$ & \underline{$82.5\!\pm\!0.3$} & $81.2\!\pm\!0.3$ & $\mathbf{82.6\!\pm\!0.3}$ \\
24
  \midrule
25
+ \textbf{MEAN (16 configurations)} & $70.8$ & $70.3$ & $70.5$ & $69.1$ & $70.4$ & $58.2$ & $63.6$ & $62.0$ & $62.3$ & $62.1$ & $65.2$ & $71.8$ & $64.3$ & $67.4$ & $72.3$ & \underline{$73.8$} & $73.1$ & $\mathbf{74.3}$ \\
output/generation/assets/tables/tab_percfg_full_seeds.tex ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table*}[t]\centering\scriptsize
2
+ \setlength{\tabcolsep}{2pt}
3
+ \caption{Misclassification ROC-AUC (\%) for each encoder--dataset configuration over 3 fine-tuning seeds (seeds 0, 1, 2). Every cell is the mean over fine-tuning seeds of the five-detector-seed mean (the single deterministic score for a closed-form method), $\pm$ its standard deviation over fine-tuning seeds, so the bar is how far the value moves when the classifier is retrained; the MEAN row averages the configuration means. Best and second-best values in each row are bold and underlined respectively. Probe suffixes b/m/l denote architecture-specific beginning, middle and final encoder depths. Baselines: SR softmax response, Ent output entropy, DCT DOCTOR, Enr Energy, MaxL MaxLogit, MD Mahalanobis, RMD relative MD, TS TrustScore, CLUE CLUE-2c. Profile is the Mahalanobis profile, Traj the Mahalanobis trajectory, and MT the profile and trajectory together; a cell is blank, and a column is left out of the MEAN row, when a configuration lacks the value at some seed.}
4
+ \label{tab:big_per_cfg_seeds}
5
+
6
+ \resizebox{\textwidth}{!}{%
7
+ \begin{tabular}{l|ccccc|cccc|cccccc|c|ccc}
8
+ \toprule
9
+ & \multicolumn{5}{c|}{output-space (softmax/logits)} & \multicolumn{4}{c|}{representation, ref.-based} & \multicolumn{6}{c|}{supervised probes} & \multicolumn{1}{c|}{final layer} & \multicolumn{3}{c}{evaluated feature sets} \\
10
+ \cmidrule(lr){2-6}\cmidrule(lr){7-10}\cmidrule(lr){11-16}\cmidrule(lr){17-17}\cmidrule(lr){18-20}
11
+ &SR&Ent&DCT&Enr&MaxL&MD&RMD&TS&CLUE&LinP$_\text{b}$&LinP$_\text{m}$&LinP$_\text{l}$&AttP$_\text{b}$&AttP$_\text{m}$&AttP$_\text{l}$&\textbf{Profile}&\textbf{Traj}&\textbf{MT} \\
12
+ \midrule
13
+ \multicolumn{19}{l}{\textsc{roberta}-base} \\
14
+ CoLA & $74.3\!\pm\!0.8$ & $74.3\!\pm\!0.8$ & $74.3\!\pm\!0.8$ & $73.4\!\pm\!1.7$ & $74.1\!\pm\!1.0$ & $43.5\!\pm\!4.0$ & $47.0\!\pm\!1.1$ & $55.6\!\pm\!2.2$ & $72.7\!\pm\!1.1$ & $63.3\!\pm\!0.8$ & $70.6\!\pm\!0.9$ & $75.6\!\pm\!1.3$ & $68.1\!\pm\!0.7$ & $74.4\!\pm\!1.4$ & $77.0\!\pm\!0.5$ & $76.3\!\pm\!1.3$ & $\mathbf{77.6\!\pm\!1.2}$ & \underline{$77.5\!\pm\!1.2$} \\
15
+ SST-5 & $59.1\!\pm\!2.1$ & $58.7\!\pm\!1.3$ & $59.0\!\pm\!1.5$ & $57.6\!\pm\!1.1$ & $59.1\!\pm\!1.2$ & $54.2\!\pm\!2.1$ & $54.1\!\pm\!0.6$ & $53.7\!\pm\!1.1$ & $51.9\!\pm\!0.3$ & $56.0\!\pm\!0.9$ & $58.5\!\pm\!1.0$ & $58.6\!\pm\!1.3$ & $55.5\!\pm\!1.0$ & $57.8\!\pm\!0.8$ & $57.8\!\pm\!1.7$ & \underline{$60.7\!\pm\!0.7$} & $58.9\!\pm\!0.6$ & $\mathbf{61.1\!\pm\!0.1}$ \\
16
+ Tox & $68.9\!\pm\!0.4$ & $68.9\!\pm\!0.4$ & $68.9\!\pm\!0.4$ & $69.6\!\pm\!1.2$ & $69.1\!\pm\!0.2$ & $52.4\!\pm\!14.0$ & $59.4\!\pm\!4.2$ & $68.3\!\pm\!2.5$ & $61.6\!\pm\!6.5$ & $62.3\!\pm\!4.4$ & $69.1\!\pm\!5.2$ & $70.1\!\pm\!2.5$ & $66.1\!\pm\!3.0$ & $67.7\!\pm\!5.9$ & $70.6\!\pm\!2.4$ & $73.0\!\pm\!4.5$ & \underline{$74.5\!\pm\!4.9$} & $\mathbf{74.9\!\pm\!5.4}$ \\
17
+ Yelp & $80.2\!\pm\!0.3$ & $80.1\!\pm\!0.2$ & $80.2\!\pm\!0.2$ & $77.1\!\pm\!0.6$ & $79.0\!\pm\!0.3$ & $67.9\!\pm\!1.0$ & $80.1\!\pm\!0.2$ & $77.9\!\pm\!0.4$ & $67.8\!\pm\!0.4$ & $66.8\!\pm\!0.1$ & $75.2\!\pm\!0.4$ & $79.4\!\pm\!0.2$ & $71.5\!\pm\!0.2$ & $76.1\!\pm\!0.4$ & $79.7\!\pm\!0.3$ & $\mathbf{81.8\!\pm\!0.2}$ & $80.5\!\pm\!0.4$ & \underline{$81.7\!\pm\!0.2$} \\
18
+ \midrule
19
+ \multicolumn{19}{l}{\textsc{deberta}-v3-base} \\
20
+ CoLA & $76.1\!\pm\!2.8$ & $76.1\!\pm\!2.8$ & $76.1\!\pm\!2.8$ & $74.9\!\pm\!5.7$ & $76.3\!\pm\!3.4$ & $55.7\!\pm\!8.1$ & $60.8\!\pm\!4.4$ & $60.5\!\pm\!9.3$ & $70.2\!\pm\!0.8$ & $62.9\!\pm\!0.4$ & $67.1\!\pm\!0.9$ & $78.7\!\pm\!2.1$ & $64.4\!\pm\!0.5$ & $69.5\!\pm\!1.2$ & $78.4\!\pm\!1.0$ & $\mathbf{80.4\!\pm\!0.8}$ & $79.2\!\pm\!0.6$ & \underline{$79.7\!\pm\!0.8$} \\
21
+ SST-5 & $58.5\!\pm\!1.0$ & $57.9\!\pm\!1.4$ & $58.0\!\pm\!1.4$ & $56.8\!\pm\!1.2$ & $58.3\!\pm\!1.4$ & $55.9\!\pm\!0.8$ & $54.0\!\pm\!1.9$ & $53.6\!\pm\!0.6$ & $50.9\!\pm\!1.2$ & $53.2\!\pm\!0.5$ & $55.9\!\pm\!0.8$ & $57.5\!\pm\!1.4$ & $55.9\!\pm\!0.6$ & $56.4\!\pm\!0.7$ & $58.1\!\pm\!1.9$ & \underline{$61.1\!\pm\!0.2$} & $57.8\!\pm\!1.1$ & $\mathbf{62.1\!\pm\!0.5}$ \\
22
+ Tox & $71.5\!\pm\!4.2$ & $71.5\!\pm\!4.2$ & $71.5\!\pm\!4.2$ & $70.1\!\pm\!3.6$ & $70.5\!\pm\!3.8$ & $57.8\!\pm\!6.5$ & $67.1\!\pm\!4.2$ & $61.4\!\pm\!7.9$ & $60.6\!\pm\!2.0$ & $64.1\!\pm\!0.9$ & $66.1\!\pm\!1.0$ & $71.9\!\pm\!2.3$ & $63.8\!\pm\!0.2$ & $66.8\!\pm\!0.1$ & $73.5\!\pm\!2.1$ & $73.8\!\pm\!3.2$ & \underline{$73.8\!\pm\!2.4$} & $\mathbf{74.5\!\pm\!3.3}$ \\
23
+ Yelp & $81.2\!\pm\!0.3$ & $81.0\!\pm\!0.4$ & $81.2\!\pm\!0.4$ & $79.5\!\pm\!0.8$ & $80.4\!\pm\!0.6$ & $67.3\!\pm\!2.9$ & $82.0\!\pm\!0.2$ & $74.9\!\pm\!0.9$ & $66.3\!\pm\!1.1$ & $65.4\!\pm\!0.3$ & $71.5\!\pm\!0.2$ & $81.6\!\pm\!0.1$ & $71.4\!\pm\!0.2$ & $76.8\!\pm\!0.1$ & $81.9\!\pm\!0.3$ & $\mathbf{83.1\!\pm\!0.1}$ & $81.7\!\pm\!0.2$ & \underline{$82.9\!\pm\!0.1$} \\
24
+ \midrule
25
+ \multicolumn{19}{l}{\textsc{deberta}-v3-small} \\
26
+ CoLA & $76.9\!\pm\!2.3$ & $76.9\!\pm\!2.3$ & $76.9\!\pm\!2.3$ & $76.2\!\pm\!2.3$ & $76.4\!\pm\!2.3$ & $55.7\!\pm\!12.5$ & $62.9\!\pm\!1.6$ & $61.9\!\pm\!1.5$ & $69.7\!\pm\!1.7$ & $63.9\!\pm\!1.1$ & $66.6\!\pm\!1.7$ & $79.2\!\pm\!2.9$ & $62.8\!\pm\!0.0$ & $68.1\!\pm\!0.3$ & $78.1\!\pm\!3.0$ & $78.9\!\pm\!3.5$ & \underline{$80.0\!\pm\!3.0$} & $\mathbf{80.1\!\pm\!3.2}$ \\
27
+ SST-5 & $59.4\!\pm\!0.9$ & $59.5\!\pm\!0.6$ & $59.5\!\pm\!0.7$ & $59.3\!\pm\!0.9$ & $60.6\!\pm\!1.2$ & $55.0\!\pm\!0.7$ & $57.6\!\pm\!1.3$ & $51.6\!\pm\!1.5$ & $51.9\!\pm\!0.8$ & $56.7\!\pm\!0.6$ & $57.4\!\pm\!0.8$ & $61.3\!\pm\!0.8$ & $56.1\!\pm\!1.4$ & $59.7\!\pm\!1.9$ & $61.0\!\pm\!0.7$ & \underline{$63.8\!\pm\!1.0$} & $62.0\!\pm\!1.1$ & $\mathbf{64.3\!\pm\!0.9}$ \\
28
+ Tox & $68.4\!\pm\!0.8$ & $68.4\!\pm\!0.8$ & $68.4\!\pm\!0.8$ & $67.8\!\pm\!1.5$ & $68.2\!\pm\!0.7$ & $54.6\!\pm\!4.1$ & $65.1\!\pm\!1.0$ & $56.8\!\pm\!1.1$ & $58.5\!\pm\!1.3$ & $61.2\!\pm\!0.5$ & $60.8\!\pm\!1.4$ & $69.5\!\pm\!0.9$ & $62.5\!\pm\!0.6$ & $63.3\!\pm\!1.6$ & $68.2\!\pm\!1.2$ & \underline{$72.0\!\pm\!1.9$} & $71.9\!\pm\!1.6$ & $\mathbf{72.4\!\pm\!1.5}$ \\
29
+ Yelp & $79.4\!\pm\!0.2$ & $79.1\!\pm\!0.3$ & $79.3\!\pm\!0.3$ & $76.8\!\pm\!1.2$ & $78.4\!\pm\!0.7$ & $61.0\!\pm\!2.6$ & $81.0\!\pm\!0.1$ & $72.5\!\pm\!0.7$ & $68.0\!\pm\!3.2$ & $67.4\!\pm\!0.1$ & $69.0\!\pm\!0.5$ & $79.5\!\pm\!0.3$ & $71.0\!\pm\!0.1$ & $74.7\!\pm\!0.7$ & $79.3\!\pm\!0.6$ & \underline{$81.7\!\pm\!0.1$} & $80.3\!\pm\!0.2$ & $\mathbf{81.8\!\pm\!0.2}$ \\
30
+ \midrule
31
+ \multicolumn{19}{l}{\textsc{electra}-base} \\
32
+ CoLA & $77.9\!\pm\!1.4$ & $77.9\!\pm\!1.4$ & $77.9\!\pm\!1.4$ & $76.5\!\pm\!3.6$ & $77.5\!\pm\!2.1$ & $65.2\!\pm\!3.5$ & $56.0\!\pm\!2.5$ & $71.0\!\pm\!4.3$ & $72.2\!\pm\!0.8$ & $63.4\!\pm\!0.3$ & $66.8\!\pm\!1.8$ & $80.4\!\pm\!1.1$ & $66.5\!\pm\!0.8$ & $69.7\!\pm\!0.7$ & $80.6\!\pm\!0.8$ & $79.5\!\pm\!0.1$ & \underline{$80.7\!\pm\!0.7$} & $\mathbf{80.9\!\pm\!0.4}$ \\
33
+ SST-5 & $58.3\!\pm\!1.7$ & $54.3\!\pm\!1.0$ & $54.5\!\pm\!0.7$ & $54.4\!\pm\!1.6$ & $57.5\!\pm\!1.1$ & $53.9\!\pm\!1.4$ & $47.3\!\pm\!1.1$ & $54.2\!\pm\!0.3$ & $47.4\!\pm\!0.3$ & $58.0\!\pm\!1.0$ & $58.9\!\pm\!0.6$ & $60.9\!\pm\!1.3$ & $59.2\!\pm\!1.5$ & $61.4\!\pm\!0.6$ & $61.9\!\pm\!0.6$ & \underline{$62.7\!\pm\!1.2$} & $58.6\!\pm\!0.3$ & $\mathbf{62.7\!\pm\!0.8}$ \\
34
+ Tox & $68.8\!\pm\!2.0$ & $68.8\!\pm\!2.0$ & $68.8\!\pm\!2.0$ & $67.0\!\pm\!0.8$ & $68.2\!\pm\!1.6$ & $67.4\!\pm\!2.2$ & $62.5\!\pm\!1.2$ & $66.6\!\pm\!2.1$ & $58.8\!\pm\!1.3$ & $60.6\!\pm\!1.6$ & $58.8\!\pm\!1.4$ & $69.8\!\pm\!0.9$ & $63.8\!\pm\!0.9$ & $64.8\!\pm\!2.7$ & $70.9\!\pm\!0.2$ & $72.1\!\pm\!1.5$ & $\mathbf{73.6\!\pm\!2.2}$ & \underline{$73.3\!\pm\!2.2$} \\
35
+ Yelp & $77.2\!\pm\!0.7$ & $75.4\!\pm\!1.2$ & $76.5\!\pm\!0.9$ & $72.9\!\pm\!2.4$ & $76.5\!\pm\!1.2$ & $67.8\!\pm\!1.6$ & $81.4\!\pm\!0.1$ & $74.1\!\pm\!1.3$ & $67.6\!\pm\!0.5$ & $66.8\!\pm\!0.1$ & $71.7\!\pm\!1.0$ & $80.6\!\pm\!0.3$ & $70.9\!\pm\!0.1$ & $76.0\!\pm\!0.2$ & $80.6\!\pm\!0.6$ & \underline{$82.5\!\pm\!0.2$} & $81.4\!\pm\!0.2$ & $\mathbf{82.7\!\pm\!0.2}$ \\
36
+ \midrule
37
+ \textbf{MEAN (16 configurations)} & $71.0$ & $70.5$ & $70.7$ & $69.4$ & $70.6$ & $58.5$ & $63.7$ & $63.4$ & $62.3$ & $62.0$ & $65.2$ & $72.2$ & $64.3$ & $67.7$ & $72.4$ & \underline{$74.0$} & $73.3$ & $\mathbf{74.5}$ \\
38
+ \bottomrule\end{tabular}%
39
+ }
40
+ \end{table*}
output/generation/assets/tables/tab_seed_cells.tex ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table*}[t]\centering\small
2
+ \caption{Per-configuration spread over 3 fine-tuning seeds (seeds 0, 1, 2). Each entry is mean\,$\pm$\,std over fine-tuning seeds of the five-detector-seed mean ROC-AUC (\%). The difference column is MT minus LastOut within the same fine-tuned classifier, so both readouts see identical predictions; wins count the seeds at which MT is ahead.}
3
+ \label{tab:seed_cells}
4
+ \begin{tabular}{llrrrrrr}
5
+ \toprule
6
+ Encoder & Task & Base acc.\,(\%) & MT & AttnProbe (final) & LastOut & $\Delta$ (MT $-$ LastOut) & wins \\
7
+ \midrule
8
+ RoBERTa & CoLA & 79.2\,$\pm$\,0.5 & 77.50\,$\pm$\,1.25 & 76.96\,$\pm$\,0.55 & 76.28\,$\pm$\,1.44 & $+1.21$\,$\pm$\,0.80 & 3/3 \\
9
+ RoBERTa & SST-5 & 52.3\,$\pm$\,0.4 & 61.08\,$\pm$\,0.12 & 57.84\,$\pm$\,1.68 & 59.63\,$\pm$\,0.11 & $+1.45$\,$\pm$\,0.06 & 3/3 \\
10
+ RoBERTa & ToxiGen & 76.3\,$\pm$\,5.5 & 74.85\,$\pm$\,5.36 & 70.57\,$\pm$\,2.35 & 71.63\,$\pm$\,5.80 & $+3.22$\,$\pm$\,0.97 & 3/3 \\
11
+ RoBERTa & Yelp & 81.1\,$\pm$\,0.0 & 81.68\,$\pm$\,0.23 & 79.73\,$\pm$\,0.31 & 81.42\,$\pm$\,0.22 & $+0.26$\,$\pm$\,0.20 & 3/3 \\
12
+ DeBERTa & CoLA & 84.3\,$\pm$\,0.1 & 79.66\,$\pm$\,0.78 & 78.43\,$\pm$\,1.03 & 79.10\,$\pm$\,1.38 & $+0.56$\,$\pm$\,0.92 & 2/3 \\
13
+ DeBERTa & SST-5 & 52.9\,$\pm$\,1.2 & 62.14\,$\pm$\,0.51 & 58.06\,$\pm$\,1.91 & 61.98\,$\pm$\,0.75 & $+0.16$\,$\pm$\,0.26 & 2/3 \\
14
+ DeBERTa & ToxiGen & 78.5\,$\pm$\,0.4 & 74.53\,$\pm$\,3.26 & 73.54\,$\pm$\,2.08 & 74.31\,$\pm$\,4.03 & $+0.21$\,$\pm$\,0.88 & 1/3 \\
15
+ DeBERTa & Yelp & 82.5\,$\pm$\,0.1 & 82.88\,$\pm$\,0.13 & 81.86\,$\pm$\,0.30 & 83.09\,$\pm$\,0.11 & $-0.21$\,$\pm$\,0.02 & 0/3 \\
16
+ DeBERTa-s & CoLA & 82.1\,$\pm$\,0.1 & 80.05\,$\pm$\,3.21 & 78.05\,$\pm$\,3.04 & 78.87\,$\pm$\,4.10 & $+1.18$\,$\pm$\,0.93 & 3/3 \\
17
+ DeBERTa-s & SST-5 & 50.8\,$\pm$\,1.4 & 64.27\,$\pm$\,0.94 & 61.02\,$\pm$\,0.67 & 64.23\,$\pm$\,0.67 & $+0.04$\,$\pm$\,0.28 & 2/3 \\
18
+ DeBERTa-s & ToxiGen & 77.4\,$\pm$\,1.2 & 72.44\,$\pm$\,1.45 & 68.24\,$\pm$\,1.17 & 71.41\,$\pm$\,1.16 & $+1.04$\,$\pm$\,0.75 & 3/3 \\
19
+ DeBERTa-s & Yelp & 79.4\,$\pm$\,0.6 & 81.80\,$\pm$\,0.15 & 79.30\,$\pm$\,0.57 & 81.87\,$\pm$\,0.18 & $-0.07$\,$\pm$\,0.11 & 1/3 \\
20
+ ELECTRA & CoLA & 82.0\,$\pm$\,0.2 & 80.94\,$\pm$\,0.45 & 80.61\,$\pm$\,0.83 & 77.82\,$\pm$\,0.96 & $+3.12$\,$\pm$\,0.59 & 3/3 \\
21
+ ELECTRA & SST-5 & 47.5\,$\pm$\,0.3 & 62.69\,$\pm$\,0.81 & 61.87\,$\pm$\,0.58 & 62.37\,$\pm$\,1.15 & $+0.32$\,$\pm$\,0.64 & 2/3 \\
22
+ ELECTRA & ToxiGen & 78.5\,$\pm$\,0.4 & 73.26\,$\pm$\,2.24 & 70.90\,$\pm$\,0.16 & 69.89\,$\pm$\,2.27 & $+3.37$\,$\pm$\,0.74 & 3/3 \\
23
+ ELECTRA & Yelp & 79.1\,$\pm$\,1.0 & 82.71\,$\pm$\,0.25 & 80.62\,$\pm$\,0.56 & 82.52\,$\pm$\,0.30 & $+0.19$\,$\pm$\,0.07 & 3/3 \\
24
+ \midrule
25
+ \multicolumn{2}{l}{\textbf{Mean}} & & & & $+1.00$ & 37/48 \\
26
+ \bottomrule
27
+ \end{tabular}
28
+ \end{table*}
output/generation/assets/tables/tab_seed_ladder.tex ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table*}[t]\centering\small
2
+ \caption{The ladder under 3 fine-tuning seeds, over the 16 configurations every rung covers at every seed. Each step is measured from the row above it within the same fine-tuned classifier; wins count the configuration$\times$seed pairs in which the step is positive.}
3
+ \label{tab:seed_ladder}
4
+ \begin{tabular}{lrrrrrr}
5
+ \toprule
6
+ Given to the detector & seed 0 & seed 1 & seed 2 & mean\,$\pm$\,std & $\Delta$ & wins \\
7
+ \midrule
8
+ the same scalar, read by the detector & 70.09 & 70.81 & 70.52 & 70.47\,$\pm$\,0.36 & & \\
9
+ final layer, every class & 72.49 & 72.85 & 72.77 & 72.70\,$\pm$\,0.19 & $+2.23$ & 46/48 \\
10
+ every layer, one class & 71.98 & 72.19 & 72.16 & 72.11\,$\pm$\,0.12 & $-0.59$ & 22/48 \\
11
+ every layer, every class (\textbf{Profile}) & 73.76 & 74.03 & 74.10 & 73.96\,$\pm$\,0.18 & $+1.85$ & 44/48 \\
12
+ profile\,+\,trajectory (\textbf{MT}) & 74.25 & 74.46 & 74.87 & 74.53\,$\pm$\,0.32 & $+0.57$ & 35/48 \\
13
+ \bottomrule
14
+ \end{tabular}
15
+ \end{table*}
output/generation/assets/tables/tab_seed_ood.tex ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table*}[t]\centering\small
2
+ \caption{Out-of-distribution detection under 3 fine-tuning seeds: mean AUROC over the 48 scenarios each readout covers at every seed, one column per fine-tuning seed. Every readout is one-class, as in the main OOD table.}
3
+ \label{tab:seed_ood}
4
+ \begin{tabular}{lrrrr}
5
+ \toprule
6
+ Readout & seed 0 & seed 1 & seed 2 & mean\,$\pm$\,std \\
7
+ \midrule
8
+ MD, all layers + Traj (path geometry, shift variant) & 92.72 & 92.10 & 91.88 & \textbf{92.24}\,$\pm$\,0.43 \\
9
+ MD, all layers & 92.20 & 91.95 & 92.00 & 92.05\,$\pm$\,0.13 \\
10
+ MD, all layers + Traj (path geometry) & 92.07 & 91.40 & 91.07 & 91.51\,$\pm$\,0.51 \\
11
+ Profile, no background + Traj (path geometry) & 92.06 & 91.42 & 91.02 & 91.50\,$\pm$\,0.52 \\
12
+ Profile, no background & 91.59 & 91.33 & 91.36 & 91.43\,$\pm$\,0.14 \\
13
+ Traj (path geometry, shift variant) & 91.40 & 90.51 & 90.02 & 90.64\,$\pm$\,0.70 \\
14
+ Traj (path geometry) & 89.76 & 88.67 & 88.30 & 88.91\,$\pm$\,0.76 \\
15
+ Profile, no background + Traj & 85.77 & 82.39 & 83.33 & 83.83\,$\pm$\,1.75 \\
16
+ MD, all layers + Traj & 85.71 & 82.28 & 83.28 & 83.76\,$\pm$\,1.76 \\
17
+ \emph{Document length only} & 82.46 & 82.46 & 82.46 & 82.46\,$\pm$\,0.00 \\
18
+ MT (Profile + Traj) & 83.52 & 80.69 & 81.46 & 81.89\,$\pm$\,1.46 \\
19
+ Traj (all three families) & 82.23 & 78.99 & 79.98 & 80.40\,$\pm$\,1.66 \\
20
+ MD & 81.50 & 77.59 & 77.58 & 78.89\,$\pm$\,2.26 \\
21
+ Profile & 77.05 & 78.36 & 77.65 & 77.69\,$\pm$\,0.66 \\
22
+ $k$-NN & 80.52 & 76.03 & 76.44 & 77.66\,$\pm$\,2.48 \\
23
+ Traj (correctness deviation) & 75.66 & 76.45 & 75.88 & 76.00\,$\pm$\,0.41 \\
24
+ Energy & 71.45 & 69.74 & 71.73 & 70.97\,$\pm$\,1.08 \\
25
+ MSP & 68.98 & 68.33 & 69.67 & 69.00\,$\pm$\,0.67 \\
26
+ RMD & 67.11 & 65.78 & 66.16 & 66.35\,$\pm$\,0.68 \\
27
+ MT detector, $P(\text{error})$ & 66.56 & 66.05 & 66.21 & 66.27\,$\pm$\,0.26 \\
28
+ Traj (cluster dynamics) & 65.14 & 63.22 & 63.37 & 63.91\,$\pm$\,1.07 \\
29
+ \bottomrule
30
+ \end{tabular}
31
+ \end{table*}
output/generation/assets/tables/tab_seed_summary.tex ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table*}[t]\centering\small
2
+ \caption{Fine-tuning seed replicates. Every classifier of the 16-configuration grid was fine-tuned 3 times (seeds 0, 1, 2); each column is the grid mean ROC-AUC (\%) at that seed, over five detector seeds. The last three columns give the mean and standard deviation over fine-tuning seeds, the mean difference to LastOut over every configuration$\times$seed pair, and the number of pairs in which the row is ahead.}
3
+ \label{tab:seed_summary}
4
+ \begin{tabular}{lrrrrrr}
5
+ \toprule
6
+ Method & seed 0 & seed 1 & seed 2 & mean\,$\pm$\,std & $\Delta$ & wins \\
7
+ \midrule
8
+ \multicolumn{7}{l}{\emph{Output- and distance-space baselines}} \\
9
+ SR & 70.76 & 71.52 & 70.70 & 70.99\,$\pm$\,0.46 & $-2.53$ & 6/48 \\
10
+ Entropy & 70.30 & 71.05 & 70.27 & 70.54\,$\pm$\,0.44 & $-2.98$ & 4/48 \\
11
+ MaxLogit & 70.45 & 71.11 & 70.34 & 70.63\,$\pm$\,0.42 & $-2.89$ & 6/48 \\
12
+ Energy & 69.06 & 69.95 & 69.10 & 69.37\,$\pm$\,0.50 & $-4.16$ & 4/48 \\
13
+ DOCTOR & 70.47 & 71.18 & 70.42 & 70.69\,$\pm$\,0.43 & $-2.84$ & 5/48 \\
14
+ TrustScore & 62.03 & 65.09 & 63.11 & 63.41\,$\pm$\,1.55 & $-10.12$ & 0/48 \\
15
+ CLUE-2c & 62.28 & 61.90 & 62.57 & 62.25\,$\pm$\,0.33 & $-11.28$ & 0/48 \\
16
+ MD (final layer) & 58.18 & 60.14 & 57.04 & 58.45\,$\pm$\,1.56 & $-15.07$ & 0/48 \\
17
+ RMD (final layer) & 63.56 & 64.64 & 62.75 & 63.65\,$\pm$\,0.95 & $-9.88$ & 0/48 \\
18
+ RMD, all layers & 71.98 & 72.19 & 72.16 & 72.11\,$\pm$\,0.12 & $-1.42$ & 11/48 \\
19
+ \multicolumn{7}{l}{\emph{Supervised probes}} \\
20
+ LinProbe (final layer) & 71.78 & 72.30 & 72.40 & 72.16\,$\pm$\,0.33 & $-1.37$ & 10/48 \\
21
+ AttnProbe (begin) & 64.27 & 64.24 & 64.53 & 64.35\,$\pm$\,0.16 & $-9.18$ & 0/48 \\
22
+ AttnProbe (mid) & 67.39 & 67.36 & 68.36 & 67.70\,$\pm$\,0.57 & $-5.83$ & 0/48 \\
23
+ AttnProbe (final layer) & 72.33 & 72.51 & 72.22 & 72.35\,$\pm$\,0.15 & $-1.18$ & 10/48 \\
24
+ \multicolumn{7}{l}{\emph{Final layer, read by the same detector}} \\
25
+ RMD scalar, read by the detector & 70.09 & 70.81 & 70.52 & 70.47\,$\pm$\,0.36 & $-3.05$ & 0/48 \\
26
+ RMD per class, final layer & 72.49 & 72.85 & 72.77 & 72.70\,$\pm$\,0.19 & $-0.82$ & 7/48 \\
27
+ Raw \texttt{[CLS]}, final layer & 71.78 & 72.24 & 72.24 & 72.09\,$\pm$\,0.26 & $-1.44$ & 14/48 \\
28
+ Logits\,+\,distances, final layer & 73.06 & 73.61 & 73.90 & 73.53\,$\pm$\,0.43 & & \\
29
+ \multicolumn{7}{l}{\emph{Representation pools (ours in bold)}} \\
30
+ \textbf{Mahalanobis trajectory} & 73.08 & 73.16 & 73.60 & 73.28\,$\pm$\,0.28 & $-0.25$ & 20/48 \\
31
+ \textbf{Mahalanobis profile} & 73.76 & 74.03 & 74.10 & 73.96\,$\pm$\,0.18 & $+0.44$ & 30/48 \\
32
+ \textbf{MT} (profile\,+\,trajectory) & 74.25 & 74.46 & 74.87 & \textbf{74.53}\,$\pm$\,0.32 & $+1.00$ & 37/48 \\
33
+ \bottomrule
34
+ \end{tabular}
35
+ \end{table*}
output/generation/assets/tables/tab_trajladder.tex CHANGED
@@ -1,14 +1,14 @@
1
  \begin{table}[t]\centering\small
2
- \caption{The trajectory's columns in three groups, each added to the profile at the same position. \emph{A} is derivable from the profile; \emph{B} introduces correctness-conditioned reference centroids but reads one layer at a time; \emph{C} is the group whose every column requires two or more layers, and is the one the reported representation uses: MT is the profile plus C, and contains no A or B column. The groups are disjoint, so these are alternatives at one position rather than a chain. B, a 52-column single-layer alternative, recovers about two fifths of the increment of C.}
3
  \label{tab:trajladder}
4
  \begin{tabular}{@{}>{\raggedright\arraybackslash}p{0.33\columnwidth}rrrr@{}}
5
  \toprule
6
  Added to the profile & ROC-AUC & $\Delta$ & wins & $p$ \\
7
  \midrule
8
- nothing (profile alone) & 73.76 & & & \\
9
- A --- derivable from the profile & 74.06 & $+0.30$ & $11/16$ & $0.0507$ \\
10
- B --- one layer, new reference geometry & 73.97 & $+0.21$ & $11/16$ & $0.0443$ \\
11
- C --- two or more layers (\textbf{MT}) & 74.25 & $+0.49$ & $13/16$ & $0.0076$ \\
12
  \bottomrule
13
  \end{tabular}
14
  \end{table}
 
1
  \begin{table}[t]\centering\small
2
+ \caption{The trajectory's columns in three groups, each added to the profile at the same position. \emph{A} is derivable from the profile; \emph{B} introduces correctness-conditioned reference centroids but reads one layer at a time; \emph{C} is the group whose every column requires two or more layers, and is the one the reported representation uses: MT is the profile plus C, and contains no A or B column. The groups are disjoint, so these are alternatives at one position rather than a chain. B, a 52-column single-layer alternative, recovers about three fifths of the increment of C.}
3
  \label{tab:trajladder}
4
  \begin{tabular}{@{}>{\raggedright\arraybackslash}p{0.33\columnwidth}rrrr@{}}
5
  \toprule
6
  Added to the profile & ROC-AUC & $\Delta$ & wins & $p$ \\
7
  \midrule
8
+ nothing (profile alone) & 73.96 & & & \\
9
+ A --- derivable from the profile & 74.21 & $+0.25$ & $10/16$ & $0.1439$ \\
10
+ B --- one layer, new reference geometry & 74.31 & $+0.35$ & $12/16$ & $0.0063$ \\
11
+ C --- two or more layers (\textbf{MT}) & 74.53 & $+0.57$ & $13/16$ & $0.0052$ \\
12
  \bottomrule
13
  \end{tabular}
14
  \end{table}
output/generation/clustered_inference.json CHANGED
@@ -7,16 +7,16 @@
7
  "n_tasks": 4,
8
  "cells_per_task": "4,4,4,4",
9
  "seed_mismatch_cells": 0,
10
- "delta_cfg": 1.6325,
11
- "delta_task": 1.6325,
12
- "wins": 14,
13
- "p_config": 0.00058,
14
  "p_task": 0.125,
15
  "p_task_floor": 0.125,
16
- "ci_lo": 0.6733,
17
- "ci_hi": 2.862,
18
- "rank_biserial": 0.8971,
19
- "icc_task": 0.3941
20
  },
21
  {
22
  "comparison": "profile over collapsed",
@@ -26,16 +26,16 @@
26
  "n_tasks": 4,
27
  "cells_per_task": "4,4,4,4",
28
  "seed_mismatch_cells": 0,
29
- "delta_cfg": 1.7847,
30
- "delta_task": 1.7847,
31
  "wins": 15,
32
- "p_config": 0.002136,
33
  "p_task": 0.125,
34
  "p_task_floor": 0.125,
35
- "ci_lo": 0.4662,
36
- "ci_hi": 3.458,
37
- "rank_biserial": 0.8235,
38
- "icc_task": 0.5209
39
  },
40
  {
41
  "comparison": "trajectory over profile",
@@ -45,16 +45,16 @@
45
  "n_tasks": 4,
46
  "cells_per_task": "4,4,4,4",
47
  "seed_mismatch_cells": 0,
48
- "delta_cfg": 0.494,
49
- "delta_task": 0.494,
50
  "wins": 13,
51
- "p_config": 0.007629,
52
  "p_task": 0.25,
53
  "p_task_floor": 0.125,
54
- "ci_lo": 0.0999,
55
- "ci_hi": 0.7452,
56
- "rank_biserial": 0.7353,
57
- "icc_task": 0.2808
58
  },
59
  {
60
  "comparison": "profile over raw CLS",
@@ -64,16 +64,16 @@
64
  "n_tasks": 4,
65
  "cells_per_task": "4,4,4,4",
66
  "seed_mismatch_cells": 0,
67
- "delta_cfg": 1.9783,
68
- "delta_task": 1.9783,
69
- "wins": 14,
70
- "p_config": 0.000427,
71
  "p_task": 0.125,
72
  "p_task_floor": 0.125,
73
- "ci_lo": 0.792,
74
- "ci_hi": 3.6517,
75
- "rank_biserial": 0.9118,
76
- "icc_task": 0.7064
77
  },
78
  {
79
  "comparison": "MT over raw CLS",
@@ -83,35 +83,16 @@
83
  "n_tasks": 4,
84
  "cells_per_task": "4,4,4,4",
85
  "seed_mismatch_cells": 0,
86
- "delta_cfg": 2.4723,
87
- "delta_task": 2.4723,
88
- "wins": 15,
89
- "p_config": 6.1e-05,
90
  "p_task": 0.125,
91
  "p_task_floor": 0.125,
92
- "ci_lo": 1.3124,
93
- "ci_hi": 4.1624,
94
- "rank_biserial": 0.9853,
95
- "icc_task": 0.7228
96
- },
97
- {
98
- "comparison": "profile over attention",
99
- "baseline": "single_v1",
100
- "contender": "rmdcls_only",
101
- "n_cfgs": 16,
102
- "n_tasks": 4,
103
- "cells_per_task": "4,4,4,4",
104
- "seed_mismatch_cells": 0,
105
- "delta_cfg": 1.9034,
106
- "delta_task": 1.9034,
107
- "wins": 13,
108
- "p_config": 0.009186,
109
- "p_task": 0.25,
110
- "p_task_floor": 0.125,
111
- "ci_lo": 0.3057,
112
- "ci_hi": 4.1408,
113
- "rank_biserial": 0.7206,
114
- "icc_task": 0.7035
115
  },
116
  {
117
  "comparison": "profile over strongest probe",
@@ -121,16 +102,16 @@
121
  "n_tasks": 4,
122
  "cells_per_task": "4,4,4,4",
123
  "seed_mismatch_cells": 16,
124
- "delta_cfg": 1.4317,
125
- "delta_task": 1.4317,
126
- "wins": 12,
127
- "p_config": 0.015503,
128
  "p_task": 0.125,
129
  "p_task_floor": 0.125,
130
- "ci_lo": 0.7176,
131
- "ci_hi": 2.1459,
132
- "rank_biserial": 0.6765,
133
- "icc_task": 0.0
134
  },
135
  {
136
  "comparison": "MT over strongest probe",
@@ -140,53 +121,34 @@
140
  "n_tasks": 4,
141
  "cells_per_task": "4,4,4,4",
142
  "seed_mismatch_cells": 16,
143
- "delta_cfg": 1.9257,
144
- "delta_task": 1.9257,
145
- "wins": 14,
146
- "p_config": 0.002136,
147
  "p_task": 0.125,
148
  "p_task_floor": 0.125,
149
- "ci_lo": 1.4186,
150
- "ci_hi": 2.6682,
151
- "rank_biserial": 0.8235,
152
- "icc_task": 0.0
153
  },
154
  {
155
- "comparison": "profile over final layer in full",
156
- "baseline": "lastall_only",
157
- "contender": "rmdcls_only",
158
- "n_cfgs": 16,
159
- "n_tasks": 4,
160
- "cells_per_task": "4,4,4,4",
161
- "seed_mismatch_cells": 0,
162
- "delta_cfg": 0.2834,
163
- "delta_task": 0.2834,
164
- "wins": 13,
165
- "p_config": 0.033539,
166
- "p_task": 0.125,
167
- "p_task_floor": 0.125,
168
- "ci_lo": 0.2241,
169
- "ci_hi": 0.392,
170
- "rank_biserial": 0.6029,
171
- "icc_task": 0.0
172
- },
173
- {
174
- "comparison": "MT over final layer in full",
175
- "baseline": "lastall_only",
176
  "contender": "rc_trajC",
177
  "n_cfgs": 16,
178
  "n_tasks": 4,
179
  "cells_per_task": "4,4,4,4",
180
- "seed_mismatch_cells": 0,
181
- "delta_cfg": 0.7773,
182
- "delta_task": 0.7773,
183
- "wins": 14,
184
- "p_config": 0.00058,
185
  "p_task": 0.125,
186
  "p_task_floor": 0.125,
187
- "ci_lo": 0.3547,
188
- "ci_hi": 1.0621,
189
- "rank_biserial": 0.8971,
190
- "icc_task": 0.1854
191
  }
192
  ]
 
7
  "n_tasks": 4,
8
  "cells_per_task": "4,4,4,4",
9
  "seed_mismatch_cells": 0,
10
+ "delta_cfg": 1.6793,
11
+ "delta_task": 1.6793,
12
+ "wins": 16,
13
+ "p_config": 3.1e-05,
14
  "p_task": 0.125,
15
  "p_task_floor": 0.125,
16
+ "ci_lo": 0.598,
17
+ "ci_hi": 3.2376,
18
+ "rank_biserial": 1.0,
19
+ "icc_task": 0.7412
20
  },
21
  {
22
  "comparison": "profile over collapsed",
 
26
  "n_tasks": 4,
27
  "cells_per_task": "4,4,4,4",
28
  "seed_mismatch_cells": 0,
29
+ "delta_cfg": 1.8525,
30
+ "delta_task": 1.8525,
31
  "wins": 15,
32
+ "p_config": 0.000427,
33
  "p_task": 0.125,
34
  "p_task_floor": 0.125,
35
+ "ci_lo": 0.6271,
36
+ "ci_hi": 3.7348,
37
+ "rank_biserial": 0.9118,
38
+ "icc_task": 0.7736
39
  },
40
  {
41
  "comparison": "trajectory over profile",
 
45
  "n_tasks": 4,
46
  "cells_per_task": "4,4,4,4",
47
  "seed_mismatch_cells": 0,
48
+ "delta_cfg": 0.5685,
49
+ "delta_task": 0.5685,
50
  "wins": 13,
51
+ "p_config": 0.005157,
52
  "p_task": 0.25,
53
  "p_task_floor": 0.125,
54
+ "ci_lo": 0.1857,
55
+ "ci_hi": 0.914,
56
+ "rank_biserial": 0.7647,
57
+ "icc_task": 0.2253
58
  },
59
  {
60
  "comparison": "profile over raw CLS",
 
64
  "n_tasks": 4,
65
  "cells_per_task": "4,4,4,4",
66
  "seed_mismatch_cells": 0,
67
+ "delta_cfg": 1.8736,
68
+ "delta_task": 1.8736,
69
+ "wins": 15,
70
+ "p_config": 0.000153,
71
  "p_task": 0.125,
72
  "p_task_floor": 0.125,
73
+ "ci_lo": 0.8858,
74
+ "ci_hi": 3.3727,
75
+ "rank_biserial": 0.9559,
76
+ "icc_task": 0.7239
77
  },
78
  {
79
  "comparison": "MT over raw CLS",
 
83
  "n_tasks": 4,
84
  "cells_per_task": "4,4,4,4",
85
  "seed_mismatch_cells": 0,
86
+ "delta_cfg": 2.4421,
87
+ "delta_task": 2.4421,
88
+ "wins": 16,
89
+ "p_config": 3.1e-05,
90
  "p_task": 0.125,
91
  "p_task_floor": 0.125,
92
+ "ci_lo": 1.6163,
93
+ "ci_hi": 3.9073,
94
+ "rank_biserial": 1.0,
95
+ "icc_task": 0.7141
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
96
  },
97
  {
98
  "comparison": "profile over strongest probe",
 
102
  "n_tasks": 4,
103
  "cells_per_task": "4,4,4,4",
104
  "seed_mismatch_cells": 16,
105
+ "delta_cfg": 1.6116,
106
+ "delta_task": 1.6116,
107
+ "wins": 14,
108
+ "p_config": 0.00058,
109
  "p_task": 0.125,
110
  "p_task_floor": 0.125,
111
+ "ci_lo": 0.6647,
112
+ "ci_hi": 2.2691,
113
+ "rank_biserial": 0.8971,
114
+ "icc_task": 0.2713
115
  },
116
  {
117
  "comparison": "MT over strongest probe",
 
121
  "n_tasks": 4,
122
  "cells_per_task": "4,4,4,4",
123
  "seed_mismatch_cells": 16,
124
+ "delta_cfg": 2.1801,
125
+ "delta_task": 2.1801,
126
+ "wins": 16,
127
+ "p_config": 3.1e-05,
128
  "p_task": 0.125,
129
  "p_task_floor": 0.125,
130
+ "ci_lo": 1.4576,
131
+ "ci_hi": 2.9025,
132
+ "rank_biserial": 1.0,
133
+ "icc_task": 0.2611
134
  },
135
  {
136
+ "comparison": "MT over strongest published trajectory score",
137
+ "baseline": "CoE-R",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
138
  "contender": "rc_trajC",
139
  "n_cfgs": 16,
140
  "n_tasks": 4,
141
  "cells_per_task": "4,4,4,4",
142
+ "seed_mismatch_cells": 16,
143
+ "delta_cfg": 14.3383,
144
+ "delta_task": 14.3383,
145
+ "wins": 16,
146
+ "p_config": 3.1e-05,
147
  "p_task": 0.125,
148
  "p_task_floor": 0.125,
149
+ "ci_lo": 11.2221,
150
+ "ci_hi": 17.2289,
151
+ "rank_biserial": 1.0,
152
+ "icc_task": 0.1876
153
  }
154
  ]
output/generation/last_build.json CHANGED
@@ -1,31 +1,36 @@
1
  {
2
- "generated_utc": "2026-09-06T14:04:11Z",
3
- "git_commit": "856213e59c2408b898b7bad8560e565d0afd4f5a",
4
  "assets": [
 
 
 
 
 
5
  {
6
  "path": "figs/fig_mechanism.pdf",
7
- "bytes": 20835,
8
- "sha256_16": "43f0bde766c28985"
9
  },
10
  {
11
  "path": "figs/fig_ood.pdf",
12
- "bytes": 32813,
13
- "sha256_16": "c7d21e4ef790fc22"
14
  },
15
  {
16
  "path": "figs/fig_percfg_bars.pdf",
17
- "bytes": 39719,
18
- "sha256_16": "5bdb3356f3c607f3"
19
  },
20
  {
21
  "path": "figs/fig_percfg_grid.pdf",
22
- "bytes": 28010,
23
- "sha256_16": "ea76576f52ba35cd"
24
  },
25
  {
26
  "path": "figs/fig_percfg_gridv.pdf",
27
- "bytes": 27284,
28
- "sha256_16": "51e14b203320b682"
29
  },
30
  {
31
  "path": "figs/percfg_cells.pdf",
@@ -35,42 +40,42 @@
35
  {
36
  "path": "tables/tab_classaxis.tex",
37
  "bytes": 575,
38
- "sha256_16": "5ba03dd6fdd82ef8"
39
  },
40
  {
41
  "path": "tables/tab_clustered.tex",
42
- "bytes": 2412,
43
- "sha256_16": "dd7ad41192353a28"
44
  },
45
  {
46
  "path": "tables/tab_controls.tex",
47
- "bytes": 1195,
48
- "sha256_16": "96eae8a4d11d9071"
49
  },
50
  {
51
  "path": "tables/tab_datasets.tex",
52
- "bytes": 828,
53
- "sha256_16": "6cee7b0a6a7d2727"
54
  },
55
  {
56
  "path": "tables/tab_ladder.tex",
57
- "bytes": 1592,
58
- "sha256_16": "eff0e3cf0a21781f"
59
  },
60
  {
61
  "path": "tables/tab_main.tex",
62
- "bytes": 2721,
63
- "sha256_16": "f4d9c5a29f31b9a8"
64
  },
65
  {
66
  "path": "tables/tab_mechanism.tex",
67
  "bytes": 809,
68
- "sha256_16": "dc59de15f61e913c"
69
  },
70
  {
71
  "path": "tables/tab_ood.tex",
72
- "bytes": 3057,
73
- "sha256_16": "38af5f9e98b6f98c"
74
  },
75
  {
76
  "path": "tables/tab_oodtiers.tex",
@@ -79,23 +84,48 @@
79
  },
80
  {
81
  "path": "tables/tab_percfg.tex",
82
- "bytes": 2081,
83
- "sha256_16": "d837655ab0053b8a"
84
  },
85
  {
86
  "path": "tables/tab_percfg_full.tex",
87
- "bytes": 8747,
88
- "sha256_16": "1df42a5ceb7a6f4f"
89
  },
90
  {
91
  "path": "tables/tab_percfg_full_rows.tex",
92
- "bytes": 7009,
93
- "sha256_16": "0c646a35e52d1d55"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94
  },
95
  {
96
  "path": "tables/tab_trajladder.tex",
97
- "bytes": 1115,
98
- "sha256_16": "5f89db9b201b4314"
99
  }
100
  ]
101
  }
 
1
  {
2
+ "generated_utc": "2026-09-08T16:45:49Z",
3
+ "git_commit": "6d0a8c27584b411290c31120507c3a0341c16943",
4
  "assets": [
5
+ {
6
+ "path": "figs/fig_channel_auc.pdf",
7
+ "bytes": 44185,
8
+ "sha256_16": "a056696bcc80ff6c"
9
+ },
10
  {
11
  "path": "figs/fig_mechanism.pdf",
12
+ "bytes": 20823,
13
+ "sha256_16": "9f1973cb462db61f"
14
  },
15
  {
16
  "path": "figs/fig_ood.pdf",
17
+ "bytes": 66773,
18
+ "sha256_16": "a8c082be3b51d8ec"
19
  },
20
  {
21
  "path": "figs/fig_percfg_bars.pdf",
22
+ "bytes": 38712,
23
+ "sha256_16": "4f7c055ae34227cf"
24
  },
25
  {
26
  "path": "figs/fig_percfg_grid.pdf",
27
+ "bytes": 27715,
28
+ "sha256_16": "9b9b79886abf4c12"
29
  },
30
  {
31
  "path": "figs/fig_percfg_gridv.pdf",
32
+ "bytes": 53375,
33
+ "sha256_16": "75d63196b507c0e2"
34
  },
35
  {
36
  "path": "figs/percfg_cells.pdf",
 
40
  {
41
  "path": "tables/tab_classaxis.tex",
42
  "bytes": 575,
43
+ "sha256_16": "b51e686e5fc32ef9"
44
  },
45
  {
46
  "path": "tables/tab_clustered.tex",
47
+ "bytes": 1624,
48
+ "sha256_16": "eed5f4c361bef29a"
49
  },
50
  {
51
  "path": "tables/tab_controls.tex",
52
+ "bytes": 1253,
53
+ "sha256_16": "0a45099596082871"
54
  },
55
  {
56
  "path": "tables/tab_datasets.tex",
57
+ "bytes": 1004,
58
+ "sha256_16": "c24a256d950e9c8a"
59
  },
60
  {
61
  "path": "tables/tab_ladder.tex",
62
+ "bytes": 1613,
63
+ "sha256_16": "4441c91d10001d4a"
64
  },
65
  {
66
  "path": "tables/tab_main.tex",
67
+ "bytes": 2485,
68
+ "sha256_16": "2b1b33db615d7271"
69
  },
70
  {
71
  "path": "tables/tab_mechanism.tex",
72
  "bytes": 809,
73
+ "sha256_16": "7e12be613f0d4583"
74
  },
75
  {
76
  "path": "tables/tab_ood.tex",
77
+ "bytes": 3332,
78
+ "sha256_16": "56e2bfce62885727"
79
  },
80
  {
81
  "path": "tables/tab_oodtiers.tex",
 
84
  },
85
  {
86
  "path": "tables/tab_percfg.tex",
87
+ "bytes": 1876,
88
+ "sha256_16": "4912b835ca60d787"
89
  },
90
  {
91
  "path": "tables/tab_percfg_full.tex",
92
+ "bytes": 7937,
93
+ "sha256_16": "519be4932a7bbdcc"
94
  },
95
  {
96
  "path": "tables/tab_percfg_full_rows.tex",
97
+ "bytes": 6383,
98
+ "sha256_16": "c7eb49d15158c907"
99
+ },
100
+ {
101
+ "path": "tables/tab_percfg_full_seeds.tex",
102
+ "bytes": 8097,
103
+ "sha256_16": "5d16d21d89330c06"
104
+ },
105
+ {
106
+ "path": "tables/tab_seed_cells.tex",
107
+ "bytes": 2748,
108
+ "sha256_16": "45ee66b0ed76b88b"
109
+ },
110
+ {
111
+ "path": "tables/tab_seed_ladder.tex",
112
+ "bytes": 983,
113
+ "sha256_16": "bb6a59c391610cc7"
114
+ },
115
+ {
116
+ "path": "tables/tab_seed_ood.tex",
117
+ "bytes": 1906,
118
+ "sha256_16": "2750c5c09cfe2beb"
119
+ },
120
+ {
121
+ "path": "tables/tab_seed_summary.tex",
122
+ "bytes": 2663,
123
+ "sha256_16": "40621a032a731c9e"
124
  },
125
  {
126
  "path": "tables/tab_trajladder.tex",
127
+ "bytes": 1117,
128
+ "sha256_16": "7050c2abd10ffc81"
129
  }
130
  ]
131
  }