{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2024:RQHZAYWWV6IWIKPHD7XLFKSSAP","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"b50802e452220cb41cc41233ea1ba2f2115e3db7957ec175d305f65ac0749979","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-07-19T16:07:19Z","title_canon_sha256":"ddba3bc248051a0a77b174db45bc08eacc7bda1175ce1108901b70d6287f634e"},"schema_version":"1.0","source":{"id":"2407.14435","kind":"arxiv","version":3}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2407.14435","created_at":"2026-05-17T23:38:50Z"},{"alias_kind":"arxiv_version","alias_value":"2407.14435v3","created_at":"2026-05-17T23:38:50Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.14435","created_at":"2026-05-17T23:38:50Z"},{"alias_kind":"pith_short_12","alias_value":"RQHZAYWWV6IW","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"RQHZAYWWV6IWIKPH","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"RQHZAYWW","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:11db9647d48502d349d9059ab3f77e9916d7c8cbb1ee63e1d3e3599e33491f87","target":"graph","created_at":"2026-05-17T23:38:50Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"JumpReLU SAEs achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs, and this improvement does not come at the cost of interpretability."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That straight-through estimators applied to the discontinuous JumpReLU produce gradients that reliably optimize the intended sparse reconstruction objective rather than converging to a different local minimum."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"JumpReLU SAEs deliver higher reconstruction fidelity than Gated or TopK SAEs at fixed sparsity on Gemma 2 9B activations while preserving interpretability."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations."}],"snapshot_sha256":"f331e7f277e79dc657ba9058d62252d6fcd7671a063114430ac775d755258305"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"22e9ca61f233f2b0009b65db182ea2ec5ab6674f1046ac505fe93e30d7bfd138"},"paper":{"abstract_excerpt":"Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations. To be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse -- two objectives that are in tension. In this paper, we introduce JumpReLU SAEs, which achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs. We also show that this improveme","authors_text":"Arthur Conmy, J\\'anos Kram\\'ar, Neel Nanda, Nicolas Sonnerat, Senthooran Rajamanoharan, Tom Lieberum, Vikrant Varma","cross_cats":[],"headline":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-07-19T16:07:19Z","title":"Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.14435","kind":"arxiv","version":3},"verdict":{"created_at":"2026-05-15T19:06:02.397883Z","id":"0db2304a-1b55-49b5-894e-af89d365e568","model_set":{"reader":"grok-4.3"},"one_line_summary":"JumpReLU SAEs deliver higher reconstruction fidelity than Gated or TopK SAEs at fixed sparsity on Gemma 2 9B activations while preserving interpretability.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations.","strongest_claim":"JumpReLU SAEs achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs, and this improvement does not come at the cost of interpretability.","weakest_assumption":"That straight-through estimators applied to the discontinuous JumpReLU produce gradients that reliably optimize the intended sparse reconstruction objective rather than converging to a different local minimum."}},"verdict_id":"0db2304a-1b55-49b5-894e-af89d365e568"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:780c6b30f41d3fb69fd39173213743bbc6d30142dffff5559a6f43c1020b6592","target":"record","created_at":"2026-05-17T23:38:50Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"b50802e452220cb41cc41233ea1ba2f2115e3db7957ec175d305f65ac0749979","cross_cats_sorted":[],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-07-19T16:07:19Z","title_canon_sha256":"ddba3bc248051a0a77b174db45bc08eacc7bda1175ce1108901b70d6287f634e"},"schema_version":"1.0","source":{"id":"2407.14435","kind":"arxiv","version":3}},"canonical_sha256":"8c0f9062d6af916429e71feeb2aa5203da4f5109a78c81c141926122bf5f2fee","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"8c0f9062d6af916429e71feeb2aa5203da4f5109a78c81c141926122bf5f2fee","first_computed_at":"2026-05-17T23:38:50.468390Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:50.468390Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"aJabJsw1FbP621eIm3t/OixhR8HgzTJx4Qf5VkPe8yrfKTJGfjkwWjdxiYBxfIp0oPTuzunkxp62HCfl4XzhAQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:50.469156Z","signed_message":"canonical_sha256_bytes"},"source_id":"2407.14435","source_kind":"arxiv","source_version":3}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:780c6b30f41d3fb69fd39173213743bbc6d30142dffff5559a6f43c1020b6592","sha256:11db9647d48502d349d9059ab3f77e9916d7c8cbb1ee63e1d3e3599e33491f87"],"state_sha256":"7ac659ecb506b2449dd66cd3e63167bf9ddaa8ba2816bdc79bdc60be1e65dd2a"}