{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/arxiv-2606-06333","title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","arxiv_id":"2606.06333","date":"2026-06-04","proceeding":null,"authors":["Seyed Arshan Dalili","Mehrdad Mahdavi"],"abstract":"Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional. We show that this assumption mismatches with the multi-dimensional structure of model features, provably inducing feature splitting through two distinct mechanisms. Geometrically, reconstructing a feature of intrinsic dimension $d_i \\ge 2$ to error $\\varepsilon$ with single-direction decoders forces a number of atoms that is exponential in $d_i$. From an end-to-end optimization perspective, this splitting is not merely possible but actively preferred. We prove that there exists a continuous path from the true $d_i$-dimensional basis to a strictly lower risk of the $\\ell_1$-regularized SAE objective, whose descent directions drive any trained dictionary into that exponential regime. A single coherent feature is therefore fragmented across many near-collinear latents, producing spurious multiplicity and obscuring the intrinsic geometry. Motivated by this, we introduce Subspace-Aware Sparse Autoencoders (SASA), which replace single-vector decoders with learned decoder subspaces, enforce block sparsity via Top-$s$ group gating, and adapt each group's effective rank with a nuclear-norm regularizer. We then show that once the block size satisfies $r \\ge d_i$, a single group not only can represent the entire feature slice but is the global minimizer of the SASA objective. This consolidation yields a sample complexity polynomial in $d_i$ rather than exponential -- a decisive advantage given that every training activation costs an LLM forward pass. Empirically, on GPT-2 and Mistral-7B, SASA reduces feature splitting and absorption, improves monosemanticity and interpretability, and matches or exceeds standard SAEs while training on roughly half the token budget.","url_abs":"https://arxiv.org/abs/2606.06333","url_pdf":"https://arxiv.org/pdf/2606.06333","source":{"archive":null,"snapshot":"2025-07-28","note":"not in the Papers with Code archive (frozen at the snapshot)","row_kind":"graph","title_abstract_authors_date":"arXiv metadata, CC0 1.0 (https://info.arxiv.org/help/license)"},"code_links":[],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=2606.06333","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"2606.06333"}},"developers":"https://syntology.ai/developers","read_at":"2026-09-24T18:15:14+00:00","read_at_is":"when the build read Syntology's graph, not when any sample ran","claim":"Per-sample execution status on synthesized fixtures; not a correctness claim about the paper. Samples come from repositories linked to the paper, official or community; repo_kind says which.","repos":[{"mentioned_in_github":null,"is_official":null,"provenance":"deterministic:regex_extraction","mentioned_in_paper":null,"url":"https://github.com/arshandalili/sasa","reach":null}],"summary":{"ran":8,"unverified":3},"by_repo_kind":{"found_in_text":{"samples":11,"ran":8,"repositories":1}},"repo_kind_vocabulary":{"official":"The archive marks this repository official for the paper","named_in_paper":"The archive records that the paper mentions this repository; it is not marked official","listed":"In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper","found_in_text":"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"},"n_pointer_only_for_licence":11,"samples":[{"code_sha256_prefix":"2fac6816fcedb9cf","entry":"load_split","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"eval/absorption.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/eval/absorption.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"2fac6816fcedb9cf"}},{"code_sha256_prefix":"4291a5d025460f82","entry":"pca_dimension","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"analysis/cluster_decoders.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/analysis/cluster_decoders.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"4291a5d025460f82"}},{"code_sha256_prefix":"7e6084a3a5f0c07f","entry":"read_json","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"common.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/common.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"7e6084a3a5f0c07f"}},{"code_sha256_prefix":"52ed43d50830b060","entry":"resolve_device","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"common.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/common.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"52ed43d50830b060"}},{"code_sha256_prefix":"c3168ff0b414d637","entry":"single_token_ids","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"analysis/temporal_subspace.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/analysis/temporal_subspace.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"c3168ff0b414d637"}},{"code_sha256_prefix":"d6a09534d473bdf8","entry":"slugify_label","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"common.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/common.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":true,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"d6a09534d473bdf8"}},{"code_sha256_prefix":"0ed6143cc259a5ff","entry":"spectral_metrics_from_values","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"analysis/rank_analysis.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/analysis/rank_analysis.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"0ed6143cc259a5ff"}},{"code_sha256_prefix":"dda600f8de06c654","entry":"temporal_vocabulary","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"analysis/temporal_subspace.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/analysis/temporal_subspace.py","link_basis":"first_harvest_node","language":"python","status":"ran","verification_level":1,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"dda600f8de06c654"}},{"code_sha256_prefix":"7ed5392a8f60b7cc","entry":"group_coords","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"analysis/temporal_subspace.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/analysis/temporal_subspace.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"7ed5392a8f60b7cc"}},{"code_sha256_prefix":"c6b764437a26dfad","entry":"load_ground_truth_probe","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"eval/absorption.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/eval/absorption.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"c6b764437a26dfad"}},{"code_sha256_prefix":"d8dc02e5116708d3","entry":"resolve_api_key","repo":"arshandalili/sasa","repo_kind":"found_in_text","path":"eval/run_autointerp.py","file_url":"https://github.com/arshandalili/sasa/blob/HEAD/eval/run_autointerp.py","link_basis":"first_harvest_node","language":"python","status":"unverified","verification_level":0,"contract_check":null,"metamorphic_tier":null,"behaviour_fingerprint":false,"licence":"NONE","inline_ok":false,"mcp_get_code":{"code_sha256":"d8dc02e5116708d3"}}]},"arxiv_metadata":{"licence":"arXiv metadata, CC0 1.0 (https://info.arxiv.org/help/license)","fields":["title","abstract","authors","date"],"primary_category":"cs.LG","source":"arxiv_2026.jsonl"},"syntology_extracted_results":{"kind":"leaderboard_placements","source":"Syntology's leaderboard-shaped extractor over the paper's own arXiv-HTML tables: a model pointed at a cell, the number was read from that cell and checked against the board's metric, dataset, split and scale, and an independent check accepted the entry; not reviewed by the paper's authors or the archive's editors","extractor_model":"global.anthropic.claude-sonnet-4-5-20250929-v1:0","verifier_model":null,"prompt_sha":"fa63d4bb9d755694","coverage":{"sentence":"Syntology has checked 6,795 of the 9,581 papers on this site that are newer than the archive; results from the others appear after they are checked.","papers_newer_than_archive":9581,"papers_checked":6795},"entries":[],"not_placed":{"boards":0,"rejected_by_independent_check":0,"refused_by_a_rule":0,"check_did_not_answer":0,"proposed_without_a_cell":0,"declined_by_site":0}}}