id
int64
1
546k
paper_arxiv_id
stringlengths
10
13
path
stringlengths
1
249
caption
stringlengths
9
170k
label
stringlengths
6
192
name
float64
499
2508.00380
figs/ablation/#1_2.pdf
\caption{Rosenbrock (200D)}
null
null
500
2508.00380
figs/ablation/#1_3.pdf
\caption{Rastrigin (200D)}
null
null
501
2508.00380
figs/ablation/#1_4.pdf
\caption{Levy (200D)}
null
null
502
2508.00380
figs/ablation/#1_5.pdf
\caption{Landing (12D)}
null
null
503
2508.00380
figs/ablation/#1_6.pdf
\caption{Rover (60D)}
null
null
504
2508.00380
figs/ablation/#1_7.pdf
\caption{Walker (198D)}
null
null
505
2508.00380
figs/ablation/#1_8.pdf
\caption{Ant (404D)}
null
null
506
2508.00380
figs/ablation/#1_legend.pdf
null
null
null
507
2508.00331
figures/token_examples/susc_21_wikipedia_en_pos240_endoftext.pdf
\caption{ \textbf{The rainbow serpent: UMAP projection of susceptibility vectors.} (Top) Each point represents a token $y$ in context $x$, positioned according to its 16-dimensional susceptibility vector $\eta_w(xy)$ computed for a 3M parameter language model (one dimension per attention head) at the end of training ($49900$ steps). Token sequences $xy$ are colored by pattern, see \cref{tab:token-patterns}. We mark the anterior-posterior (PC1) and dorsal-ventral (PC2) axes with black lines (Bottom) Susceptibility vector $\eta_w(xy)$ showing each head in order \protect\attentionhead{0}{0}-\protect\attentionhead{0}{7} (gray) \protect\attentionhead{1}{0}-\protect\attentionhead{1}{7} (black) and the token $y$ (black outline) in its context $x$. }
\label{fig:rainbow_serpent}
null
508
2508.00331
figures/umap/umap_2d_overview_12_pclines_nolabels.png
null
null
null
509
2508.00331
figures/umap/umap_2d_overview_12_pclines_nolabels.png
null
null
null
510
2508.00331
figures/umap/umap_2d_overview_12_step900.png
null
null
null
511
2508.00331
figures/umap/umap_2d_overview_12_step9000.png
null
null
null
512
2508.00331
figures/umap/umap_2d_overview_12_step17500.png
null
null
null
513
2508.00331
figures/umap/umap_2d_overview_12_step49900.png
null
null
null
514
2508.00331
figures/appendix/pps_appx_1.pdf
null
null
null
515
2508.00331
figures/umap/umap_2d_overview_12_step900.png
null
null
null
516
2508.00331
figures/umap/umap_2d_overview_12_step9000.png
null
null
null
517
2508.00331
figures/umap/umap_2d_overview_12_step17500.png
null
null
null
518
2508.00331
figures/umap/umap_2d_overview_12_step49900.png
null
null
null
519
2508.00331
figures/appendix/pps_appx_1.pdf
null
null
null
520
2508.00331
figures/pattern_distribution_heatmap_with_std.pdf
\caption{\textbf{Per-token susceptibility PCA} showing mean and standard deviation of loadings of principal components on data patterns across $10$ independent draws of $20000$ tokens from each dataset. The data matrix is the same as that used for UMAP, see \cref{appendix:umap}. The number $30.2$ appearing in ``Top 1\% Negative'' for word start tokens in PC1 means that when we order the tokens in PC1 with negative coefficients by magnitude, 30.2\% of those in the top 1\% by magnitude are word start tokens.}
\label{fig:pattern_dist_heatmap_std}
null
521
2508.00331
figures/spacing/spacing_tokens_2d_step900.png
null
null
null
522
2508.00331
figures/spacing/spacing_tokens_2d_step9000.png
null
null
null
523
2508.00331
figures/spacing/spacing_tokens_2d_step17500.png
null
null
null
524
2508.00331
figures/spacing/spacing_tokens_2d_step49900.png
null
null
null
525
2508.00331
figures/spacing/spacing_tokens_2d_step900.png
null
null
null
526
2508.00331
figures/spacing/spacing_tokens_2d_step9000.png
null
null
null
527
2508.00331
figures/spacing/spacing_tokens_2d_step17500.png
null
null
null
528
2508.00331
figures/spacing/spacing_tokens_2d_step49900.png
null
null
null
529
2508.00331
figures/space_bars.pdf
\caption{\textbf{Average spacing token susceptibilities.} The average per-token susceptibilities for spacing tokens is shown, conditional on the minimal number of preceding spacing tokens. Colors range from green (1) to black (80). Susceptibilities are measured at the end of training (49900 steps).}
\label{fig:spacing-bars}
null
530
2508.00331
figures/dataset_pattern_distribution.pdf
\caption{Percentages of tokens in each dataset which follow a given pattern. Note that not all patterns are mutually exclusive.}
\label{fig:distribution_pattern}
null
531
2508.00331
figures/lm-fig1-top.pdf
\caption{\label{fig:stages} \textbf{Developmental stages of the language model.} Critical points in the local learning coefficient (LLC) curve mark boundaries between distinct \emph{developmental stages} (bottom row; warm hues for increasing LLC, cold for decreasing LLC). }
null
null
532
2508.00331
figures/appendix/pca_fig.png
\caption{A 3D visualization of the first three principal components is shown for the same per-token susceptibility data used to generate the UMAP serpents. Note that the different axes were automatically rescaled.}
\label{fig:3d-pca}
null
533
2508.00331
figures/appendix/thicc.png
\caption{The distribution of tokens projected onto PC2 in susceptibility space, separated into induction pattern tokens and all other tokens.}
\label{fig:thicc}
null
534
2508.00331
figures/spacing/spacing_tokens_by_type_step900.png
null
null
null
535
2508.00331
figures/spacing/spacing_tokens_by_type_step9000.png
null
null
null
536
2508.00331
figures/spacing/spacing_tokens_by_type_step17500.png
null
null
null
537
2508.00331
figures/spacing/spacing_tokens_by_type_step49900.png
null
null
null
538
2508.00331
figures/spacing/spacing_tokens_by_type_step900.png
null
null
null
539
2508.00331
figures/spacing/spacing_tokens_by_type_step9000.png
null
null
null
540
2508.00331
figures/spacing/spacing_tokens_by_type_step17500.png
null
null
null
541
2508.00331
figures/spacing/spacing_tokens_by_type_step49900.png
null
null
null
542
2508.00331
figures/spacing/space_count_histogram.pdf
\caption{Histogram of the number of spacing tokens preceding a spacing token, in the approximately 42k spacing tokens found in the set of 260k sampled token sequences.}
\label{fig:umap_spacing_hist}
null
543
2508.00331
figures/appendix/checkpoint_comparison_2x2_v2.pdf
\caption{The average per-token susceptibilities for spacing tokens is shown across four checkpoints, conditional on the minimum number of preceding spacing tokens. Colors range from green (1) to black (80).}
\label{fig:umap_spacing_bars_2x2}
null
544
2508.00331
figures/umap/seed2/umap_2d_overview_12_step900.png
null
null
null
545
2508.00331
figures/umap/seed2/umap_2d_overview_12_step9000.png
null
null
null
546
2508.00331
figures/umap/seed2/umap_2d_overview_12_step17500.png
null
null
null
547
2508.00331
figures/umap/seed2/umap_2d_overview_12_step49900.png
null
null
null
548
2508.00331
figures/appendix/pps_appx_2.pdf
null
null
null
549
2508.00331
figures/umap/seed2/umap_2d_overview_12_step900.png
null
null
null
550
2508.00331
figures/umap/seed2/umap_2d_overview_12_step9000.png
null
null
null
551
2508.00331
figures/umap/seed2/umap_2d_overview_12_step17500.png
null
null
null
552
2508.00331
figures/umap/seed2/umap_2d_overview_12_step49900.png
null
null
null
553
2508.00331
figures/appendix/pps_appx_2.pdf
null
null
null
554
2508.00331
figures/umap/seed3/umap_2d_overview_12_step900.png
null
null
null
555
2508.00331
figures/umap/seed3/umap_2d_overview_12_step9000.png
null
null
null
556
2508.00331
figures/umap/seed3/umap_2d_overview_12_step17500.png
null
null
null
557
2508.00331
figures/umap/seed3/umap_2d_overview_12_step49900.png
null
null
null
558
2508.00331
figures/appendix/pps_appx_3.pdf
null
null
null
559
2508.00331
figures/umap/seed3/umap_2d_overview_12_step900.png
null
null
null
560
2508.00331
figures/umap/seed3/umap_2d_overview_12_step9000.png
null
null
null
561
2508.00331
figures/umap/seed3/umap_2d_overview_12_step17500.png
null
null
null
562
2508.00331
figures/umap/seed3/umap_2d_overview_12_step49900.png
null
null
null
563
2508.00331
figures/appendix/pps_appx_3.pdf
null
null
null
564
2508.00331
figures/umap/seed4/umap_2d_overview_12_step900.png
null
null
null
565
2508.00331
figures/umap/seed4/umap_2d_overview_12_step9000.png
null
null
null
566
2508.00331
figures/umap/seed4/umap_2d_overview_12_step17500.png
null
null
null
567
2508.00331
figures/umap/seed4/umap_2d_overview_12_step49900.png
null
null
null
568
2508.00331
figures/appendix/pps_appx_4.pdf
null
null
null
569
2508.00331
figures/umap/seed4/umap_2d_overview_12_step900.png
null
null
null
570
2508.00331
figures/umap/seed4/umap_2d_overview_12_step9000.png
null
null
null
571
2508.00331
figures/umap/seed4/umap_2d_overview_12_step17500.png
null
null
null
572
2508.00331
figures/umap/seed4/umap_2d_overview_12_step49900.png
null
null
null
573
2508.00331
figures/appendix/pps_appx_4.pdf
null
null
null
574
2508.00331
figures/pca/pc_loadings_seed1.pdf
\caption{\textbf{Head loadings of per-token susceptibility PCA for seed $1$.} In \citet[Appendix G]{wang2024differentiationspecializationattentionheads} it was found that in this seed the previous-token heads are \protect\attentionhead{0}{1} and \protect\attentionhead{0}{4}, the current-token head is \protect\attentionhead{0}{5} and the induction heads are \protect\attentionhead{1}{6} and \protect\attentionhead{1}{7}.}
\label{fig:pattern_dist_heatmap_seed1}
null
575
2508.00331
figures/pca/pc_loadings_seed2.pdf
\caption{\textbf{Head loadings of per-token susceptibility PCA for seed $2$.} In \citet[Appendix G]{wang2024differentiationspecializationattentionheads} it was found that in this seed the previous-token head is \protect\attentionhead{0}{1}, the current-token head is \protect\attentionhead{0}{6} and the induction head is \protect\attentionhead{1}{0}.}
\label{fig:pattern_dist_heatmap_seed2}
null
576
2508.00331
figures/pca/pc_loadings_seed3.pdf
\caption{\textbf{Head loadings of per-token susceptibility PCA for seed $3$.} In \citet[Appendix G]{wang2024differentiationspecializationattentionheads} it was found that in this seed the previous-token head is \protect\attentionhead{0}{4}, the current-token head is \protect\attentionhead{0}{2} and the induction head is \protect\attentionhead{1}{6}.}
\label{fig:pattern_dist_heatmap_seed3}
null
577
2508.00331
figures/pca/pc_loadings_seed4.pdf
\caption{\textbf{Head loadings of per-token susceptibility PCA for seed $4$.} In \citet[Appendix G]{wang2024differentiationspecializationattentionheads} it was found that in this seed the previous-token heads are \protect\attentionhead{0}{0} and \protect\attentionhead{0}{3}, the current-token head is \protect\attentionhead{0}{6} and the induction heads are \protect\attentionhead{1}{1} and \protect\attentionhead{1}{3}.}
\label{fig:pattern_dist_heatmap_seed4}
null
578
2508.00331
figures/umap/umap_2d_overview_01_pc.png
null
null
null
579
2508.00331
figures/umap/umap_2d_overview_02_pc.png
null
null
null
580
2508.00331
figures/umap/umap_2d_overview_01_pc.png
null
null
null
581
2508.00331
figures/umap/umap_2d_overview_02_pc.png
null
null
null
582
2508.00331
figures/token_examples/susc_22_wikipedia_en_pos409_It.pdf
\caption{ \textbf{Token streaks.} Susceptibility vector $\eta_w(xy)$ showing each head in order \protect\attentionhead{0}{0}-\protect\attentionhead{0}{7} (gray) \protect\attentionhead{1}{0}-\protect\attentionhead{1}{7} (black) and the token $y$ (black outline) in its context $x$. The examples are taken from the streaks (A)-(E) identified in \cref{fig:umap_other_angles}. Note the similarity of the susceptibility vectors of the tokens \tokenbox{the}, \tokenbox{to}, \tokenbox{and}. }
\label{fig:token_streaks}
null
583
2508.00458
figure/receiver.eps
\caption{Schematic of a RA receiver. The vapor cell contains ground-state atoms that are excited to Rydberg states via coupling and probe lasers. Incoming RF signals interact with these Rydberg atoms, modifying the transmission of the probe laser that is subsequently detected by the photodetector.}
\label{receiver}
null
584
2508.00458
figure/CS1_no_b.eps
\caption{\label{fig:constellation}Visualization of LOAM constellation points with $M =4$ and the phase of $c$ is selected from the discrete set $\{0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4}\}$. (a) LO-free case showing constellation points co-linearly arranged along a ray from the origin. (b) Weak reference signal case displaying non-symmetrically distributed constellation points with the first point $x_{1}$ anchored at $c$. (c) Strong reference signal case showing symmetrically distributed constellation points centered around the origin.}
null
null
585
2508.00458
figure/CS1_no_b.eps
null
null
null
586
2508.00458
figure/CS1_small_b.eps
null
null
null
587
2508.00458
figure/CS1_large_b.eps
null
null
null
588
2508.00458
figure/A2_Constellation_Comparison_LOfree_M4.eps
\caption{\label{fig03} SER performance comparison between the proposed LOAM scheme and conventional modulation schemes (PAM, QAM, PSK) under different reference signal conditions and constellation sizes. Top row (a-c): $4$-point constellations ($M=4$). Bottom row (d-f): $64$-point constellations ($M=64$).}
null
null
589
2508.00458
figure/A2_Constellation_Comparison_LOfree_M4.eps
null
null
null
590
2508.00458
figure/A2_Constellation_Comparison_LOsmall_M4.eps
null
null
null
591
2508.00458
figure/A2_Constellation_Comparison_LOlarge_M4.eps
null
null
null
592
2508.00458
figure/A2_Constellation_Comparison_LOfree_M64.eps
null
null
null
593
2508.00458
figure/A2_Constellation_Comparison_LOsmall_M64.eps
null
null
null
594
2508.00458
figure/A2_Constellation_Comparison_LOlarge_M64.eps
null
null
null
595
2508.00519
./HBondImp_ev.eps
\caption{Band structure of a tight-binding model with a Hermitian bond impurity for a system of $L=32$ as a function of $t'$. As $t'=1$ two states, one from the bottom, one from the top of the band, separate from the overall band. We calculated the bulk-boundary correspondence indicator ($D_L(\epsilon)$) and found it to be divergent for these states, but not for any of the others.}
\label{fig:HBondImp_ev}
null
608
2508.00646
img/results_q3_tool_acceptance_stacked.pdf
\caption{Digital Tool Acceptance (n=5)}
\label{fig:q3_tool_acceptance}
null
825
2508.00355
figures/RL_evaluate_modify.pdf
null
null
null
596
2508.00519
./HBondImp_D.eps
\caption{Bulk-boundary correspondence indicator, $D_L(\epsilon)$ for a Hermitian system with a Hermitian bond impurity. Panel (a) shows $D_L(\epsilon)$ as a function of $t'$ for all states for three system sizes. As $t'$ increases above $t'=1$ for a certain set of states $D_L(\epsilon)$ diverges. The larger the system size, the steeper the increase. Panel (b) shows $D_L(\epsilon)$ as a function of the energy eigenvalue, for two values of $t'$, $t'=0.5,1.6$. The arrows indicate the extremal states which separated from the band and are localized for $t'=1.6$.}
\label{fig:HBondImp_D}
null