Research map / Large language models / Language, safety and interpretability
Learning Retrieval Models with Sparse Autoencoders
Thibault Formal, Maxime Louis, Hervé Déjean, Stéphane Clinchant
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (multilingual) · MTEB(Multilingual, v2) retrieval | nDCG@10: 61.7 -> 62.3 | Improved |
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (multilingual) · XTREME-UP | MRR@10: 56.2 -> 58.6 | Improved |
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (multilingual) · MIRACL | nDCG@10: 69.9 -> 71.7 | Improved |
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (MS MARCO-trained) · MTEB Code | nDCG@10: 57.3 -> 55.1 | Hurt |
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (MS MARCO-trained) · MTEB ChemTEB | nDCG@10: 75.9 -> 70.0 | Hurt |
| replace LM-head vocabulary with SAE latent vocabulary | SPLADE-Llama (MS MARCO-trained) · MTEB English | nDCG@10: 52.9 -> 52.9 | No clear effect |
| cap document features at Top-K=100 at inference | SPLARE vs SPLADE · MS MARCO / MTEB English | performance drop: 2% vs over 6% drop | Improved |
| add Top-K inference pooling (40,400) to SPLARE (ablation) | SPLARE · MTEB(Multilingual, v2) retrieval | nDCG@10: 63.8 -> 62.3 | Hurt |
| reduce Top-K from (40,400) to (10,100) (ablation) | SPLARE · MTEB(Multilingual, v2) retrieval | nDCG@10: 62.3 -> 56.0 | Hurt |
| use SAE at layer 6 (SPLARE-2B) instead of layer 26 (ablation) | SPLARE · MTEB(Multilingual, v2) retrieval | nDCG@10: 62.3 -> 59.1 | Hurt |
| increase SAE width (ablation) | SPLARE (Gemma-2-2B) · MTEB(Eng, v2) retrieval | nDCG@10 | Improved |
| use intermediate layer (about two-thirds depth) for SPLARE (ablation) | SPLARE · MTEB(Eng, v2) retrieval | nDCG@10 | Improved |
Closest papers by meaning
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.89)
- Improving Sparse Decomposition of Language Model Activations with Gated Sparse Autoencoders (NeurIPS 2024 · similarity 0.89)
- Sparse Autoencoders, Again? (ICML 2025 · similarity 0.89)
- Low-Rank Adapting Models for Sparse Autoencoders (ICML 2025 · similarity 0.88)
- Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures (NeurIPS 2026 · similarity 0.88)
- Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
