Sparse autoencoders (SAEs) are widely used to interpret language model activations, but SAE training and latent labelling are typically repeated for every model. Here, we show that a single shared SAE can replace a collection of dedicated per-model SAEs. Our method, SharedSAE, combines a shared dictionary with model-specific encoder-decoder pairs. Unlike the closest prior method, which discards activation magnitudes and requires all models at inference, SharedSAE instead normalizes only selection scores, preserving magnitudes, and uses model dropout for single-model inference. We train SharedSAE on four 1B-scale base language models spanning distinct families and tokenizers. Despite sharing its latents across models, SharedSAE retains 96.6% of dedicated SAEs' mean explained variance; its latent activations exhibit cross-model correlations 1.8 times as high as separate SAEs aligned post-hoc, and its latent descriptions transfer across models. After the dictionary is frozen, new models can be efficiently adapted to it, achieving near-dedicated-SAE reconstruction quality while reusing the shared latent descriptions.
Figures & tables
Figure 1: SharedSAE replaces post-hoc matching with one persistent, adaptable dictionary. Top left: dedicated SAEs admit only partial post-hoc matching of their latents. Top right: SharedSAE trains per-model encoders and decoders around one shared dictionary and later attaches a new model with the dictionary frozen. Bottom left: the criteria a shared SAE must meet, with each method’s score. Bottom right: under solo TopK, SharedSAE ’s latents correlate across models far more strongly than post-hoc matched SAEs, USAE, or SPARC.
Figure 2: SharedSAE. The aligner cuts each model’s token stream into windows that decode to the same text and reads xm at the last token of each window (highlighted). Per-model encoders and decoders surround one shared dictionary: the codes am are normalized and summed, one TopK gives the index set I for all models, and with probability p=0.5 a single random model picks I alone (dice), which trains solo inference. Each decoder reconstructs its own model (self loss) and every other model from that model’s rescaled code.
Figure 3: How many models use a latent: Neff - 4 when all four models select it equally often, 1 for single-model use.
Figure 4: The share of a latent’s code mass one model contributes within a pair, pooled over the six pairs and symmetrized.
Figure 5: Latent selection under the solo TopK, per token: each model’s encoder alone selects its top-128 latents for the same aligned token; the distributions count how many of those latents all four models select at once. Dotted lines mark the means.
Figure 6: Latent selection under the solo TopK, per latent: we obtain the top-50 tokens that activate each latent and report the four-model intersection of these four sets.
Figure 7: Per-latent cross-model value correlation under the joint TopK (top), where all four models vote for the latents to use, and under the solo TopK (bottom), where each model chooses alone. The labels are each system’s median
Qwen3
Llama-3.2
OLMo-2
Gemma-3
Mean
Cross-FVE
Reconstruction (FVE)
Dedicated (ceiling)
0.838
0.807
0.812
0.868
0.831
-
SharedSAE , joint TopK
0.798
0.785
0.787
0.825
0.799
0.749
SharedSAE , solo TopK
0.804
0.794
0.795
0.821
0.803
-
USAE
0.760
0.743
0.750
0.787
0.760
0.640
SPARC, joint TopK
0.769
0.765
0.763
0.631
0.732
0.663
Table 1: Held-out reconstruction and spliced behavior, every system at its own converged budget, under the joint TopK and the solo TopK. For CE-recovered and KL, the reconstruction replaces the residual stream and next-token behavior is scored. SPARC’s codes are unit-norm, so its printed reconstruction uses the oracle per-token rescale into each model’s space; with the deployable one-scalar-per-model median-norm rescale its means drop to 0.643 (joint) and 0.490 (solo). Blanks are due to architectural differences: USAE and the crosscoder have no joint/solo split, the crosscoder has no cross-reconstruction. The dedicated rows are the per-model ceiling, a best-case reference. Bold marks the best shared system per column within each block excluding the dedicated reference.
decoded into model m′
code from m
Qwen3
Llama-3.2
OLMo-2
Gemma-3
Qwen3
0.798
0.738
0.738
0.782
Llama-3.2
0.746
0.785
0.738
0.786
OLMo-2
0.741
0.733
0.787
0.780
Gemma-3
0.739
0.733
0.736
0.825
Dedicated
0.815
0.783
0.795
0.825
Table 2: Self- and cross-reconstruction FVE of SharedSAE under the joint TopK.
Clarity
Resp.
Purity
Faith.
Retained
Within-stream clear labels (clarity ≥0.9 )
SharedSAE shared vote
0.979
0.642
0.463
0.281
10.6%
SharedSAE per-stream
0.976
0.629
0.438
0.217
21.7%
SAE Qwen3
0.975
0.584
0.416
0.333
7.4%
SAE Llama-3.2
0.962
0.558
0.431
0.403
9.4%
SAE OLMo-2
0.954
0.583
0.373
0.320
4.1%
Table 3: FADE evaluation of SharedSAE and dedicated SAE latents. Top: within-stream evaluation of label–stream pairs satisfying clarity ≥0.9 . Retained reports the fraction of evaluated label–stream pairs satisfying the clarity threshold. Middle: cross-stream transfer of source-validated labels. Shared-vote labels are retained if they reach clarity ≥0.9 in at least one stream and are then evaluated on all streams. Clearest per-stream selects, for each latent, the per-stream label with highest source-stream clarity and retains it when this value exceeds 0.9 . Bottom: no clarity filtering is applied. For each latent, the per-stream label with highest source-stream clarity is selected and FADE scores are correlated across the six model pairs.
Figure 8: Most-activating tokens for selected single-concept latents, across the four models SharedSAE was trained on and two models adapted to the frozen dictionary afterwards. min effect is the probe gap in the weakest model, in standard-deviation units. Full contexts are in Appendix O .
Seed TopK
Solo
Held-out model
FVE
/co-tr.
/ded.
FVE
/ded.
Leave-one-out (three-model seed)
Qwen3
0.769
0.997
0.944
0.793
0.972
Llama-3.2
0.753
0.990
0.962
0.769
0.982
OLMo-2
0.751
0.984
0.945
0.776
0.977
Gemma-3
0.787
0.979
0.954
0.812
0.984
Table 4: Adapting a held-out model to a frozen SharedSAE : only the new model’s encoder, decoder, and biases train, so the seed models and their latent descriptions stay untouched.
corr.
top-50
choice
FVE/ded.
Co-trained SharedSAE (4 models)
0.785
31.5
59.0
0.999
Adapted, leave-one-out mean
0.785
31.5
66.0
0.979
Held-out Gemma fold, attach routes
adapted encoder-decoder (ours)
0.785
32.0
60.3
0.984
linear alignment (Atlas-style)
0.722
29.5
57.5
0.692
decoder-only attach
-
-
-
0.795
Table 5: The Section 4 criteria under adaptation, with the adapted model voting solo: median per-latent value correlation and top-50 overlap against the seed models, mean per-token intersection of the solo selections (of 128; four-way above, five-way in the Qwen3.5 and Instruct blocks), and solo FVE as a share of the dedicated SAE.
Figure 9: Validation FVE by epoch at the 4M protocol: dedicated SAEs (left), SharedSAE self-reconstruction (middle), and SharedSAE cross-reconstruction (right, colored by target). Every curve rises monotonically through epoch 8 (dashed line), so the dedicated ceilings and SharedSAE train on to convergence (Section A ). The middle and right panels add the converged SharedSAE run’s logged tail, epochs 41-62, flat at its early stop; a node requeue truncated that run’s earlier log, and the probe stream drawn for epochs 1-8 is a separate run (Appendix C ).
Vote input (10 epochs, no dropout)
Share
Vote shares (Q/L/O/G)
raw votes
0.919
0.07/0.00/0.00/0.93
RMS-cal. inputs
0.925
0.24/0.40/0.30/0.07
vote norm (raw)
0.930
0.25/0.25/0.25/0.25
Vote-source dropout dose (vote norm)
Share
Cross-FVE
Gemma solo
Table 6: Ablations against the settled configuration (vote norm, vote-source dropout p=0.5 , oracle per-token cross rescale, λx=1 ); each block varies one ingredient with the rest held fixed. Share, vote shares (Q/L/O/G), Corr., and Top-50 are defined in Appendix C . Within each block, the best value per column is in bold. Measured at the earlier protocol (800k-position store, k=256 , 10 epochs, held-out shard 3), not rerun at the 4M protocol.
CE-recovered ↑
KL(clean∥patched)↓
Qwen3
Llama
OLMo-2
Gemma
Qwen3
Llama
OLMo-2
Gemma
Dedicated (8-epoch)
0.966
0.922
0.965
0.943
0.243
0.484
0.309
0.423
SharedSAE (settled)
0.958
0.933
0.970
0.968
0.309
0.438
0.279
0.293
Table 7: Functional splice at the hook layer under the solo TopK: the reconstruction replaces the residual stream and next-token behavior is scored, per stream, against the budget-matched 8-epoch dedicated SAEs (the converged dedicated splice is in Table 1 ). CE-recovered and KL are defined in Appendix C . Best value per column in bold (higher CE-recovered, lower KL).
Qwen3
Llama-3.2
OLMo-2
Gemma-3
Mean
SharedSAE (mean consensus size j=58.9 of 128)
all own 128 latents
0.804
0.794
0.795
0.821
0.803
consensus latents only
0.736
0.721
0.727
0.716
0.725
own strongest j latents
0.758
0.740
0.746
0.739
0.746
latents outside the consensus
-1.139
-0.010
0.032
-8.247
-2.341
USAE, 32 epochs (mean consensus size j=49.5 of 128)
Table 8: Reconstruction (FVE) through the consensus latents. For each token, each model’s encoder alone selects its top-128 latents under the solo TopK; the consensus set is the intersection of the four selections. Each model is then reconstructed from its own code restricted to: all 128 own selections; the consensus set only; its own strongest j latents, a size-matched control; and its selections outside the consensus.
100k window, latents alive ≥ 10 times
Pairwise (of 50)
All-four
System
median (span)
(of 50)
SharedSAE
37 (37-38)
26
Matched dedicated
15 (11-17)
-
USAE (8 epochs)
0 (0-0)
-
300k window, latents with ≥ 50 firings ( n )
Table 9: Top-50 activating-position overlap: how many of a latent’s 50 top-activating positions coincide across models. Top block: median over the six model pairs (span in parentheses) and the all-four count on the core of latents alive at least 50 times in every model. Bottom block: the same statistic under each selection condition on the 300k window, with each system’s latent population in parentheses; USAE codes are per-model native, so only the solo column is defined. Chance overlap is 0 of 50. Dashes: statistic undefined or not computed for the system (Appendix C ). Best value per column in bold.
Table 10: Solo-TopK usage classes at τ=0.1% of tokens, the migration of the 20k-window classes when the window grows to 100k tokens, and per-stream statistics. Med. corr. is the median per-latent cross-model code correlation within the class; populations and definitions in Appendix C .
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Dedicated ceilings trained until validation FVE levels off, under the constant-rate recipe and a 64-epoch cap. Bold: the best validation FVE reached so far - the checkpointed ceiling; faint: the per-epoch series. Dots mark each stream’s peak; the cross marks Llama’s early stop (peak unbeaten for eight consecutive epochs); the dashed line is the 8-epoch budget of every shared system. The second Qwen3 seed (dotted) is indistinguishable from the first throughout. This full-shard window sits 0.5-1.0 points of FVE below the evaluation window of Table 1 (Appendix C ).
Figure 11: The post-hoc-matching reference for the second criterion. Per-latent activation correlations: the best post-hoc match between separately trained dedicated SAEs of different models (blue; converged arms, median 0.45), the same matching between two SAEs of the same model trained from different seeds (gray; median 0.77) - the ceiling the post-hoc route permits with no cross-model disagreement at all - and SharedSAE ’s shared-latent correlation with each model voting alone (orange; median 0.79; the dotted tick marks the joint-TopK median, 0.90). Matching cannot reach its own within-model ceiling across models; the shared dictionary clears that ceiling solo and exceeds it further under the joint vote. SharedSAE ’s near-zero mass is rare latents - 3% of measured latents, 0.1% of solo selections (Section 4.3 ). Matching controls are measured on the 100k window and SharedSAE correlations on the 300k window; the within-model arm is one seed pair on one stream (protocols in Appendix C ).
Figure 12: Solo per-latent value-correlation distributions inside each three-model leave-one-out seed (step outlines, colored by the held-out model) over the full four-model SharedSAE (orange fill). The four seed medians coincide with each other and with the full dictionary’s.
seed
corr.
top-50
pairwise
N -way
FVE/conv.
without Qwen3-1.7B
0.778
32.0
83.3
71.8 (3-way)
0.945
without Llama-3.2-1B
0.784
32.0
85.9
75.0 (3-way)
0.942
without OLMo-2-1B
0.791
32.0
85.8
74.8 (3-way)
0.935
without Gemma-3-1B
0.792
32.0
80.9
68.5 (3-way)
0.956
full SharedSAE
0.797
31.5
78.0
59.0 (4-way)
0.967
Appendix
Table 11: Internal sharedness of each three-model leave-one-out seed, members voting solo: pooled per-latent value-correlation median and top-50 overlap median (300k window), mean pairwise and N -way per-token selection intersection of 128 (e02 window), and mean solo FVE as a share of the members’ converged dedicated ceilings. The seeds train for 8 epochs, so their shares sit below the converged four-model row by the ceiling gap, not by any property of the roster. The full-model row restates published values.
under the solo TopK
Qwen3
Llama
OLMo
Gemma
SPARC, vote share
0.245
0.276
0.274
0.205
SPARC, solo ∩ joint
0.711
0.625
0.592
0.635
SPARC, latents used
22,933
30,635
31,097
7,719
SharedSAE , latents used
24,328
24,776
24,532
24,964
Four-way intersection of the solo selections, of 128
SPARC 49.2 SharedSAE 58.9
Appendix
Table 12: SPARC’s solo-TopK condition. Left: selection behavior on the scoring window of Table 1 - vote share is each stream’s mean share of the aggregated selection mass, solo ∩ joint the fraction of the joint top- k that a stream’s own top- k reproduces, and latents used the number of distinct latents a stream’s solo vote ever selects (the joint selection uses 7,599 for SPARC and 23,714 for SharedSAE ). Right: the same condition per model pair - how many of the 128 selections the two streams share at a token, and the two per-latent medians of Table 9 and Figure 7 with the population they are measured over (300k window). Our joint selection was verified to reproduce SPARC’s own shared_indices exactly on every batch.
model
ded.
joint
share
solo
solo/ded
sink %
Qwen3
0.807
0.768
0.952
0.767
0.951
0.015
Llama-3.2
0.782
0.738
0.944
0.753
0.963
0.000
OLMo-2
0.712
0.653
0.917
0.681
0.957
0.000
Gemma-3
0.819
0.782
0.954
0.751
0.917
0.000
Appendix
Table 13: Mid-depth ( ∼ 50%) roster: the settled SharedSAE (joint and solo TopK) against dedicated (ded.) ceilings at hooks qwen3-14 llama-8 olmo2-8 gemma-13 (mid ˜50%). share and solo/ded are ratios to the dedicated FVE, and sink % is defined in Appendix C . Cross-FVE mean is 0.652 (max 0.743). Measured at the earlier protocol (800k-position store, k=256 , 10 epochs, held-out shard 3), not rerun at the 4M protocol.
class
@20k
@100k
corr
consensus (all four)
9,987
10,776
0.91
partial (2-3 models)
1,877
1,650
0.85
specific (one model)
1,397
1,144
0.84
joint-only
17
8
0.72
below τ everywhere
19,490
19,190
0.82
Appendix
Table 14: Latent classes by solo-TopK usage (threshold τ=0.1% of tokens): latent counts when classified on a 20k-token window vs. the full 100k held-out shard; “corr” is the class’s median cross-model code correlation at 100k. The migration between the two windows is quantified in Table 10 and Appendix I .
latent
fires/100k
Qwen3-1.7B
Llama-3.2-1B
OLMo-2-1B
Gemma-3-1B
8319
50
free
free
free
free
29780
52
i
i
i
i
11723
75
,’’, ,"
,", ,’’
,’’, ,"
,", ,’’
16628
22
United
United
United
United
30588
20
United
United
United
United
10904
56
bar, Bar
bar, Bar, -bar
bar, Bar
bar, Bar
Appendix
Table 15: Rare shared latents: latents selected on 20-99 of 100k held-out tokens whose top-15 activating positions coincide across all four models. Cells show each model’s top activating tokens, decoded by its own tokenizer. Six of the eight sampled latents are shown; both “United” detectors are kept, an instance of feature splitting. Measured at the earlier protocol (800k-position store, k=256 , 10 epochs, held-out shard 3), not rerun at the 4M protocol.
Token level: ten strongest positions decoded in
each model’s own tokenizer (medians over 6 pairs)
SharedSAE
Matched ded.
Modal decoded-token agreement
0.63
0.45
Decoded token-set Jaccard
0.50
0.19
Concept level: AP retention of each model’s
five most selective detectors
Appendix
Table 16: Semantic readout under SharedSAE and under the matched dedicated control, at the token level and at the concept level. The AP-in-own-model column gives each detector’s base selectivity; retention is the ratio of a detector’s average precision read out in the receiving model to that base value. Protocol in Appendix C . Bold: higher retention per row; ties at the printed precision are both bolded.
Latent
Per-stream Qwen3
Per-stream OLMo-2
Per-stream Gemma-3
Per-stream Llama-3.2
Shared Vote
16291
word ‘in’ or ‘of’
word ‘of’ or ‘to’ or punctuation marks
word ‘of’ or ‘in’
word ‘of’ or ‘is’ or ‘in’
prepositions and articles
2317
four-year term or similar duration
four-year term or similar duration
year term election context
four-year term or similar duration
four-year term or similar duration
6418
using or a before handheld gps
handheld gps usage
word-initial ‘a’ or ‘using’
word-initial ‘a’ or ‘the’
word-initial ‘a’ or ‘using’
2727
word-initial uppercase fragment
word-initial uppercase fragment
word-initial uppercase fragment
word-initial uppercase fragment
word-initial uppercase fragment
14390
possessive form
possessive form
word-initial uppercase fragment
possessive ‘s or whose
possessive form
22504
should
should
should fragment
should
should
Appendix
Table 17: Examples of shared SharedSAE latents and their per-stream semantic descriptions.
Figure 13: Qualitative steering examples for shared SharedSAE latents activated across all four models. The same shared latent is steered independently in each model, illustrating that semantically aligned latents can induce related behavioral shifts while leaving room for model-specific effects.
Dedicated baseline
t
p
SAE Gemma-3
-1.154
0.259
SAE Llama-3.2
-1.002
0.327
SAE OLMo-2
-0.210
0.840
SAE Qwen3
-0.420
0.679
Appendix
Table 18: Welch two-sample tests comparing FADE faithfulness between SharedSAE shared-vote labels and each dedicated SAE baseline on retained label–stream pairs. None of the comparisons is significant at p<0.05 .
Figure 14: Same latent, same effect: for each of the 280 mined candidate latents, the overlap of the top-30 tokens the latent promotes across the four models (median over the six pairs), measured causally by injection at the hook layer (filled) and by the linear logit lens (outline). A mismatched-latent null sits at 0.00. String-level overlap is a lower bound: the same concept often surfaces in different languages and subword forms across tokenizers (Appendix N ).
Figure 15: Cross-model steering dose–response, mean over the 12 steered latents. Top: concept-rate lift over each dose regime’s own unsteered baseline; the dotted line is the matched-norm random-latent control. Bottom: judged coherence; arrowheads mark unsteered baselines. Doses left of the dashed line use repetition-penalized sampling, right of it plain sampling. The three width-2048 streams move nearly in lockstep; Gemma-3’s sampler bounds its judged coherence at every arm.
#1537 - positive appraisal (English and French probes)
Qwen3
(r1, in 4/4 top-10) …ists. |The spine and ribs ofgnathodon saturator| It’s a very nice display. And while the amount … †
(r5, in 4/4 top-10) …hem in a history that is both analytical and narrative…A superb book that will be a standard m … †
(r10, in 4/4 top-10) …inifa should get an Israel Business award! What a Kiddushem! Nice to hear about the positive asp … †
Llama-3.2
(r1, in 4/4 top-10) …ists. |The spine and ribs ofgnathodon saturator| It’s a very nice display. And while the amount … †
Appendix
Table 19: Latent 1537: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 38–43 distinct passages per model.
#7393 - general negative affect (moral-wrongdoing and distress probes)
Qwen3
(r1, in 4/4 top-10) But the power it would need to guarantee that no one is poor would be so great it could crush the natural rights and liberty of individuals.
(r5, in 3/4 top-10) Smederevo citizens, frightened and hurt by a harsh occupation, and on April 6, by the bombing of the city, too, withdrew to their homes and only came out on the streets when needed: to finish their jobs and to procure food.
(r10, in 3/4 top-10) Those disciples of Christ who have experienced their status to any degree have become keenly aware of the intimate relationship they have with Marya relationship that at the moment is perhaps closer than that with Christ. For it is Mary who brings us to Christ.
Llama-3.2
(r1, in 4/4 top-10) He concludes-echoing historian Rick Atkinson’s excellent recent account of the campaign, The Day of Battle-that despite its terrible cost, the fight in Italy played a decisive role in defeating Germany.
Appendix
Table 20: Latent 7393: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 30–32 distinct passages per model.
#27599 - epistemic hedging: conjecture and possibility
Qwen3
(r1, in 4/4 top-10) All three were calibrated using the same unit of measurement, the cubit; the Egyptians broke the cubit into smaller units, which allowed them to keep remarkably accurate records, perhaps more accurate than would have been warranted for the purposes of merely agriculture and taxation.
(r5, in 4/4 top-10) But, Namibian elephants are of great interest to tourists, and this may be the key to their salvation in this country that has been described as the land that God created in anger.
(r10, in 2/4 top-10) But, Namibian elephants are of great interest to tourists, and this may be the key to their salvation in this country that has been described as the land that God created in anger.
Llama-3.2
(r1, in 4/4 top-10) The American Blacks played a crucial role in the very beginning of the Sydney steel plant, but never stayed around to reap any benefits.
Appendix
Table 21: Latent 27599: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 35–40 distinct passages per model.
#23065 - conditional mood in hypothetical clauses
Qwen3
(r1, in 2/4 top-10) Simulations are supporting the education system to better equip students with practical knowledge, so when it comes to tackling real-life situations in their future careers they can perform better. Taking into consideration the benefits offered by simulations in the education process, it would be a shame to miss out on them.
(r5, in 3/4 top-10) The feature was named by Blanchard and Davis,66 who called it the Bone Springs arch. It would seem from their paper that they considered the feature to be anticlinal, and to have a similar, opposing flank to the north.
(r10, in 3/4 top-10) Thoth - Link With Sirius Sirius Mythology - Zep Tepi Star - Star Tetrahedron Thoth The Scribe used sacred geometry - geometry of the stars to write our program placing coded messages in the heavens for you to find. Creators - Guardians - ’God/Dog’ - Dog Star - would live beyond the stars in a higher realm of consciousness.
Llama-3.2
(r1, in 3/4 top-10) Didn’t scientific research recently concluded that when in the dark (can’t be seen!), people are more likely to steal? It would seem that we all agree on the moral high grounds but consistently fail when faced such tests.
Appendix
Table 22: Latent 23065: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 40–43 distinct passages per model.
#20341 - programming operations, in code and in prose about code
Qwen3
(r1, in 4/4 top-10) …ned a new file via my IDE and turned it into a list via read lines . Why? Because I want to †
(r5, in 4/4 top-10) …e opened a new file via my IDE and turned it into a list via read lines. Why? Because I want †
(r10, in 3/4 top-10) When I use [’Some name__________1.5 6.5 6.5\n’, ’Another name________6.3 1.2 1.5\n’] 1.5 3.5 4.5 2.5 3.5 4.5 5.5 3.5 4.5 fileFolder = open(’TEXTFILE’, ’r’) readFile = fileFolder.readlines () for line in readFile: line = line.split("_") grades = line[-1] grades = map(float,line[-1]) could not convert string to float: .
Llama-3.2
(r1, in 4/4 top-10) map is operating on each character of the string, because when you iterate over a string you get individual characters.
Appendix
Table 23: Latent 20341: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 8–9 distinct passages per model; this latent concentrates in the window’s few code-bearing documents (largest passage up to 44% of its top 50).
#30007 - sentence-initial imperatives: calls to action
Qwen3
(r1, in 4/4 top-10) Glyptodonts became extinct at the end of the last ice age along with a large number of other megafaunal species, including pampatheres, the giant ground sloths, and the Macrauchenia. Finding an ancient shell like this is surely not an ordinary discovery.
(r5, in 1/4 top-10) The second class is known as “blockchain technology”, and consists of both cryptocurrencies like Bitcoin and Ethereum, and alternative distributed ledger designs like Hyperledger and ErisDB.
(r10, in 3/4 top-10) If this ratio is used to design the board the board size would be increased in area by greater than 9 times.
Llama-3.2
(r1, in 2/4 top-10) …d her bed, pulling down the covers, and patting the pillow). Give your child some advance notice … †
Appendix
Table 24: Latent 30007: examples at activation ranks 1/5/10 per model under its own solo vote; the activating word is bold; ( r , in k /4 top-10) gives the example’s rank and how many models share the position. Top-50 activations span 38–40 distinct passages per model.
Large language model (LLM) interpretability studies how internal representations and computations produce outputs, and how they can be explained or controlled.
Sparse autoencoders (SAEs) decompose language model activations into sparse features, yet these models traditionally encode each token independently, failing to expose information that persists across a sequence. We first show that temporal persistence can naturally emerge in standard SAE features: after a feature activates, the hidden state remains aligned with its direction, and past activations help reconstruct later hidden states. How long this lasts varies widely across features. We therefore introduce Persistent Sparse Autoencoders (Persistent SAEs), an extension of standard SAEs that learns a persistence coefficient for each feature, allowing the model to learn feature-specific timescales from reconstruction alone. Our experiments show that Persistent SAEs retain competitive reconstruction quality while learning a spectrum of timescales: short-timescale (fast) features stay locally interpretable, whereas long-timescale (slow) features accumulate information that identifies the current context. Moreover, we show in a prompt-injection monitoring case study that slow features preserve injection-related signals and remain causally effective over long contexts. These results suggest that Persistent SAEs offer new opportunities for interpreting and monitoring language models via persistent sparse features.
Sparse autoencoders (SAEs) have been used widely to decompose and interpret neural network activations, especially those of transformer language models. One key issue with SAEs is their inability to directly model multidimensional features. Instead, SAEs may tile such features by a set of independent directions that must be grouped together after the SAE training phase, impeding discoverability and interpretation of learned feature representations. We begin to address this issue by introducing the Sparse MIXture of Autoencoders (SMIXAE) architecture. Empirically, we provide evidence that SMIXAE models have success both in directly learning previously identified manifold structures, as well as finding novel structures, within the open source Gemma 2 2B and 9B models. Finally, we discuss several limitations and point towards areas for future work.
Collin Francel
Department of Arts and Sciences, University of Alabama, Tuscaloosa, AL, United States.
Sparse autoencoders (SAEs) seek to decompose the dense activations of large language models (LLMs) into sparse, potentially interpretable features using a learned overcomplete dictionary. In this work, we ask whether it is possible to enhance the semantic selectivity of the learned dictionary atoms by linearly shaping the dense activations prior to learning an SAE. We explore rank-reducing transformations based on the recently introduced J-space corresponding to the averaged Jacobian from the current layer to the LLM output, projecting activations onto the ``live'' subspace spanned by eigenmodes of the Jacobian with larger singular values (which we expect to be more semantically relevant). We also introduce a new approach for training SAEs in which (a)~an affine shaping transformation is learned along with the dictionary atoms, (b)~the atom selection mechanism is based on the TopK projections along the one-dimensional subspace spanned by the atoms, and (c)~reconstruction is carried out using the same atoms. We call this approach a Matched Filter SAE (MF-SAE) based on analogies with communication theory. The promise of shaping is demonstrated via evaluations against dense activations and state-of-the-art SAEs on benchmarks including SAEBench and SAE Probes.