Papers › Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry

Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry

3 Mar 2025arXiv:2503.01822archive 2025-07-28

Sai Sumedh R. Hindupur, Ekdeep Singh Lubana, Thomas Fel, Demba Ba

Sparse Autoencoders (SAEs) are widely used to interpret neural networks by identifying meaningful concepts from their representations. However, do SAEs truly uncover all concepts a model relies on, or are they inherently biased toward certain kinds of concepts? We introduce a unified framework that recasts SAEs as solutions to a bilevel optimization problem, revealing a fundamental challenge: each SAE imposes structural assumptions about how concepts are encoded in model representations, which in turn shapes what it can and cannot detect. This means different SAEs are not interchangeable -- switching architectures can expose entirely new concepts or obscure existing ones. To systematically probe this effect, we evaluate SAEs across a spectrum of settings: from controlled toy models that isolate key variables, to semi-synthetic experiments on real model activations and finally to large-scale, naturalistic datasets. Across this progression, we examine two fundamental properties that real-world concepts often exhibit: heterogeneity in intrinsic dimensionality (some concepts are inherently low-dimensional, others are not) and nonlinear separability. We show that SAEs fail to recover concepts when these properties are ignored, and we design a new SAE that explicitly incorporates both, enabling the discovery of previously hidden concepts and reinforcing our theoretical insights. Our findings challenge the idea of a universal SAE and underscores the need for architecture-specific choices in model interpretability. Overall, we argue an SAE does not just reveal concepts -- it determines what can be seen at all.

PaperPDFCode Syntology ranMethod as code

In Syntology View this paper on Syntology, its page in Syntology's graph. That page adds which of its functions are also present in other papers' code Syntology holds, with dates, and its dated run record as JSON and a README badge. It also lists the repositories linked to the paper, the abstract and the calls for agents.

Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

For agents, on Syntology's MCP service (how to connect):

Code

Syntology 37 code samples harvested from 3 repositories linked to this paper: 26 ran (of which 3 constructed an object rather than computing a result; 13 with no instrument failure: 1 honoured, 0 violated, 12 with no contract checked; 13 where Syntology's instrument failed) · 11 unverified. Where Syntology's instrument failed: 5 where our contract draft was wrong; 8 where the contract check could not run; failures of Syntology's instrument, not of the code.

By repository: found in paper text by Syntology: 37 samples from 3 repositories, 26 ran (13 with no instrument failure; 13 where Syntology's instrument failed). The run record, sample by sample. “Ran” means executed on a synthesized input, not that the code is correct or reproduces the paper.

Levels written between 2026-09-22 and 2026-09-25 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed.

Sai-Sumedh/SaeConceptDuality-SpaDE found in paper text by SyntologyMITSyntology: 4 samples, 2 ran (0 instrument-failed) · 2 unverified report
EkdeepSLubana/spadeFormalGrammars found in paper text by SyntologyMITSyntology: 12 samples, 7 ran (0 instrument-failed) · 5 unverified report
KempnerInstitute/Overcomplete found in paper text by SyntologySyntology: 21 samples, 17 ran (13 instrument-failed) · 4 unverified report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

37 samples harvested: 26 ran (of which 3 constructed an object rather than computing a result; 13 with no instrument failure: 1 honoured, 0 violated, 12 with no contract checked; 13 where Syntology's instrument failed) · 11 unverified. “Where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Syntology's record for this page has not changed since , the first build that kept a record date for it; when this build read Syntology's graph is in the build record. That date is not when the samples ran.

1ran · honoured contract
5ran · our draft was wrong
8ran; the contract check could not run
12ran
11unverified

Licence: 4 of the 37 samples are pointer only, meaning Syntology does not serve that copy's text. For 4 of them, this paper's copy is under a permissive licence and cleared, but the stored copy of the same code that Syntology serves is not; the licence label's tooltip gives Syntology's recorded reason. This page shows no code text for any sample; each one links to its file in the repository.

Harvested from 3 repositories linked to this paper, official or community; the list below is grouped by repository, each group headed by its name and what it is to the paper. “Ran” means the sample executed on a synthesized input. It does not mean the output is correct, and nothing here reproduces the paper's results. A line beginning “fixture:” under a sample that ran gives the arguments Syntology's fixture called the sample with, each as name:type/rank/element kind (? where none was recorded), and the type of what the fixture's run returned, exactly as the harness recorded them; the arguments and the returned type were recorded by two separate runs of the same fixture. No output value is shown, and a returned type says nothing about whether the output was right. “Level recorded” gives the UTC date when Syntology's graph wrote the sample's verification level, not when the code executed; where the level was written before Syntology's graph kept that date, the row says “level recorded, date not recorded”. For 1 sample that ran, the graph records a signature this page does not show: for a class the fixture built without calling, the harness called the object with inputs it made up itself, so the recorded signature is the harness's probe, not the fixture's arguments. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself. “Our draft was wrong” and “the contract check could not run” are failures of Syntology's instrument, not of the code: the sample ran on Syntology's synthesized input, and then the drafted contract did not describe it, or the check itself raised an error (RAISES), lacked a dependency (DEP_MISSING) or ran out of time (TIMEOUT); hover the label for which. A metamorphic tier, where shown, is evidence of the output's shape, not that the sample implements the method.

Each sample ends with its code_sha256. A sample's code_sha256 is Syntology's identity for that exact code. It is printed in full (16 hexadecimal characters for every sample in this build), and get_code(code_sha256="…") takes it exactly as printed. An agent calls it on Syntology's MCP service (how to connect); click an identity to copy that call.

Repository labels, per sample. official repository: The archive marks this repository official for the paper. named in the paper: The archive records that the paper mentions this repository; it is not marked official. community (archive-listed): In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper. found in paper text by Syntology: Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted. community: Not in the archive's code links for this paper; a community repository Syntology harvested. Samples are grouped by repository: a repository marked official first, then by how directly the repository is tied to the paper, then in the archive's order; never by how many ran. Licence labels name the repository's licence as recorded at harvest. “Pointer only” means Syntology does not serve that copy's text, for one of six reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar); this paper's copy was not located, so its licence is not recorded; or this paper's copy is cleared, but the stored copy of the same code that Syntology serves is not. Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence label for the reason. File links open the file on GitHub at the default branch, which may have changed since the harvest.

A function or class name links to its page when the name occurs in the harvested code of at least 5 papers; the name is a string, not an identity. A census tag appears when Syntology's census ran implementations with that name on one shared input per bucket: it gives how many were compared, from how many papers, and how many distinct outputs they produced, and names this sample's bucket when this sample was one of them.

Sai-Sumedh/SaeConceptDuality-SpaDE found in paper text by Syntology 4 samples, 2 ran (0 instrument-failed) · 2 unverified
linear_pieces functions/utils.py ran MIT (permissive) · c85ad84383f852e6 · report fixture: signature model:callable/0/?, data:Tensor/2/float, receptive_fields:bool/0/bool, activation_heatmaps:bool/0/bool · returned tuple · level recorded 2026-09-24
read_hyperparameters functions/utils.py ran MIT (permissive) · 974d150ff5d7b382 · report fixture: signature task_id:int/0/int, file_path:str/0/str · returned dict · level recorded 2026-09-24
data_n_loaders functions/get_data.py unverified MIT (permissive) · 0788ffd60162d5e5 · report
get_semicircle functions/get_data.py unverified MIT (permissive) · 538e46bf31beace3 · report
EkdeepSLubana/spadeFormalGrammars found in paper text by Syntology 12 samples, 7 ran (0 instrument-failed) · 5 unverified
define_prior dgp/utils.py ran MIT (permissive) here; the stored copy is pointer only · 4138f619ef3ab02c · report fixture: signature prior_size:int/0/int, alpha:float/0/float, prior_type:str/0/str · returned ndarray · level recorded 2026-09-24
get_module in 37 papers utils/analysis.py ran MIT (permissive) here; the stored copy is pointer only · bfa0b6ce30af730f · report fixture: signature model:callable/0/?, layer:str/0/str · returned Embedding · level recorded 2026-09-24
get_pos_data figures/compute_corrmaps.py ran MIT (permissive) · 56291f3f5cd9fdac · report fixture: signature sequences:Tensor/2/int · returned tuple · level recorded 2026-09-25
get_pos_data figures/compute_metrics.py ran MIT (permissive) · 368f326e3a51cf18 · report fixture: signature sequences:Tensor/2/int · returned tuple · level recorded 2026-09-24
jumprelu sae/saes.py ran fingerprinted MIT (permissive) · 6f7a4cc489051fdf · report fixture: signature x:Tensor/2/float, threshold:Tensor/0/float, bandwidth:Tensor/0/float · returned Tensor · level recorded 2026-09-24
rectangle in 7 papers sae/saes.py ran fingerprinted MIT (permissive) · 7c06d045892ef607 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-24
softplus_inverse in 8 papers
census: 3 implementations from 6 papers, 2 distinct outputs across 2 buckets, one shared input per bucket
sae/saes.py ran MIT (permissive) · c22f704a96321b8f · report fixture: signature input:Tensor/1/float, beta:float/0/float, threshold:float/0/float · returned Tensor · level recorded 2026-09-24
getActivations figures/compute_corrmaps.py unverified MIT (permissive) · 83b36163cec01c2a · report
getActivations figures/compute_metrics.py unverified MIT (permissive) · 8364a1e2bf292345 · report
get_dataloader in 111 papers dgp/dataloader.py unverified MIT (permissive) here; the stored copy is pointer only · 834a92ff49752de4 · report
grammar_evals evals/evals.py unverified MIT (permissive) · 2e4225b09bbdd765 · report
top_k_top_p_filtering in 96 papers
census: 8 implementations from 8 papers, 4 distinct outputs across 2 buckets, one shared input per bucket
model/model.py unverified MIT (permissive) here; the stored copy is pointer only · 4088666d0047518a · report
KempnerInstitute/Overcomplete found in paper text by Syntology 21 samples, 17 ran (13 instrument-failed) · 4 unverified
DictionaryLayer overcomplete/sae/jump_sae.py ran MIT (permissive) · f4b90d4a8d7a8df8 · report fixture: returned DictionaryLayer · level recorded 2026-09-22
JumpReLU overcomplete/sae/jump_sae.py ran MIT (permissive) · 39b7b96e4c0c2ddc · report fixture: returned JumpReLU · level recorded 2026-09-22
TieableEncoder overcomplete/sae/jump_sae.py ran fingerprinted MIT (permissive) · 700cf74572418934 · report fixture: returned TieableEncoder · level recorded 2026-09-22
cauchy_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · fd68d9928b455631 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
cosine_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · 435b7a2c349bf155 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
epanechnikov_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · a567c13de0aceb15 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
gaussian_kernel in 51 papers
census: 9 implementations from 9 papers, 9 distinct outputs across 6 buckets, one shared input per bucket this sample: bucket 3 of 6, arg 1: rank 2, kind float, dtype float32
overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · cb509c068da815e4 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
jump_relu overcomplete/sae/jump_sae.py ran · our draft was wrong MIT (permissive) · 9af77435852d70eb · report fixture: signature x:Tensor/2/float, threshold:Tensor/1/float, kernel_fn:callable/0/?, bandwidth:float/0/float · returned Tensor · level recorded 2026-09-22
normalize_identity overcomplete/sae/jump_sae.py ran · honoured contract fingerprinted MIT (permissive) · ee12fc4a6248f097 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
normalize_l1 overcomplete/sae/jump_sae.py ran · our draft was wrong fingerprinted MIT (permissive) · a7aa0b2ccc793f82 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
normalize_l2 in 7 papers overcomplete/sae/jump_sae.py ran · our draft was wrong fingerprinted MIT (permissive) · efac2c02b2c4b989 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
normalize_max_l1 overcomplete/sae/jump_sae.py ran · our draft was wrong fingerprinted MIT (permissive) · 9dbc9e9b26b1a18e · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
normalize_max_l2 overcomplete/sae/jump_sae.py ran · our draft was wrong fingerprinted MIT (permissive) · b0f9943a802bf49f · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
quartic_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · 2adaee6f999cee83 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
rectangle_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · af2189e3bb817b0b · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
silverman_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · 46f3c01f0b0c7c48 · report fixture: signature x:Tensor/2/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
triangular_kernel overcomplete/sae/jump_sae.py ran; the contract check could not run (RAISES) fingerprinted MIT (permissive) · fa1a887be1a53353 · report fixture: signature x:Tensor/1/float, bandwith:float/0/float · returned Tensor · level recorded 2026-09-22
BaseDictionaryLearning overcomplete/sae/jump_sae.py unverified MIT (permissive) · 3dca837c21e29bc5 · report
EncoderFactory overcomplete/sae/jump_sae.py unverified MIT (permissive) · bbf8fa2e6a8b0d3c · report
JumpSAE overcomplete/sae/jump_sae.py unverified MIT (permissive) · 915ef400e3c55ea7 · report
SAE overcomplete/sae/jump_sae.py unverified MIT (permissive) · eafd4c95fdf03869 · report

Method as code Syntology

What Syntology's graph records as this paper's claimed method, lifted into code: each assembly names its bricks (functions and classes), each with its own role and its own verification level. An assembly is not a check that the code implements the method correctly or reproduces the paper's results. Levels are each brick's own verification level in Syntology's graph (level 0: no recorded run), listed brick by brick and never combined into one level; “ran” means the brick executed on a synthesized input, not that its output is correct, and the instrument figure counts failures of Syntology's instrument, not of the code. Where a model read the code against the paper's claim, each assembly's label verdict is one model's judgement, kept in this page's JSON twin; it is not calibrated, so no card prints it.

1 assembly on this paper: 1 harvested, 0 generated.

harvested Assembly for a new SAE (graph method: a new SAE). 21 bricks; individually on a fixture: 17 ran (of which 3 constructed an object rather than computing a result; 4 with no instrument failure: 1 honoured, 0 violated, 3 with no contract checked; 13 where Syntology's instrument failed) · 4 unverified. The assembly has not been run end-to-end.

  1. JumpSAE entry · level 0 · unverified · 915ef400e3c55ea7
  2. SAE helper · level 0 · unverified · eafd4c95fdf03869
  3. cauchy_kernel helper · level 1 · ran; the contract check could not run (RAISES) · fd68d9928b455631
  4. cosine_kernel helper · level 1 · ran; the contract check could not run (RAISES) · 435b7a2c349bf155
  5. epanechnikov_kernel helper · level 1 · ran; the contract check could not run (RAISES) · a567c13de0aceb15
  6. gaussian_kernel helper · level 1 · ran; the contract check could not run (RAISES) · cb509c068da815e4
  7. jump_relu helper · level 1 · ran · our draft was wrong · 9af77435852d70eb
  8. quartic_kernel helper · level 1 · ran; the contract check could not run (RAISES) · 2adaee6f999cee83
  9. rectangle_kernel helper · level 1 · ran; the contract check could not run (RAISES) · af2189e3bb817b0b
  10. silverman_kernel helper · level 1 · ran; the contract check could not run (RAISES) · 46f3c01f0b0c7c48
  11. triangular_kernel helper · level 1 · ran; the contract check could not run (RAISES) · fa1a887be1a53353
  12. BaseDictionaryLearning helper · level 0 · unverified · 3dca837c21e29bc5
  13. DictionaryLayer helper · level 1 · ran · f4b90d4a8d7a8df8
  14. EncoderFactory helper · level 0 · unverified · bbf8fa2e6a8b0d3c
  15. TieableEncoder helper · level 1 · ran · 700cf74572418934
  16. JumpReLU helper · level 1 · ran · 39b7b96e4c0c2ddc
  17. normalize_identity helper · level 1 · ran · honoured contract · ee12fc4a6248f097
  18. normalize_l1 helper · level 1 · ran · our draft was wrong · a7aa0b2ccc793f82
  19. normalize_l2 helper · level 1 · ran · our draft was wrong · efac2c02b2c4b989
  20. normalize_max_l1 helper · level 1 · ran · our draft was wrong · 9dbc9e9b26b1a18e
  21. normalize_max_l2 helper · level 1 · ran · our draft was wrong · b0f9943a802bf49f

Each brick ends with its code_sha256, Syntology's identity for that exact code; click one to copy the get_code call that returns it. For agents, on Syntology's MCP service: assembly(arxiv_id="2503.01822") returns these assemblies with each brick's role and wiring (how to connect).

Tasks

Bilevel Optimization

Results from the paper archive 2025-07-28

No leaderboard rows for this paper in the archive.

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections