{"paper":{"title":"Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders","license":"http://creativecommons.org/licenses/by/4.0/","headline":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Arthur Conmy, J\\'anos Kram\\'ar, Neel Nanda, Nicolas Sonnerat, Senthooran Rajamanoharan, Tom Lieberum, Vikrant Varma","submitted_at":"2024-07-19T16:07:19Z","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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","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.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"JumpReLU SAEs deliver higher reconstruction fidelity than Gated or TopK SAEs at fixed sparsity on Gemma 2 9B activations while preserving interpretability.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f331e7f277e79dc657ba9058d62252d6fcd7671a063114430ac775d755258305"},"source":{"id":"2407.14435","kind":"arxiv","version":3},"verdict":{"id":"0db2304a-1b55-49b5-894e-af89d365e568","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T19:06:02.397883Z","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.","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","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.","pith_extraction_headline":"JumpReLU sparse autoencoders deliver higher reconstruction fidelity than Gated or TopK SAEs at matched sparsity on Gemma 2 activations."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"22e9ca61f233f2b0009b65db182ea2ec5ab6674f1046ac505fe93e30d7bfd138"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}