AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Learning Retrieval Models with Sparse Autoencoders

Thibault Formal, Maxime Louis, Hervé Déjean, Stéphane Clinchant

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (multilingual) · MTEB(Multilingual, v2) retrievalnDCG@10: 61.7 -> 62.3Improved
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (multilingual) · XTREME-UPMRR@10: 56.2 -> 58.6Improved
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (multilingual) · MIRACLnDCG@10: 69.9 -> 71.7Improved
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (MS MARCO-trained) · MTEB CodenDCG@10: 57.3 -> 55.1Hurt
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (MS MARCO-trained) · MTEB ChemTEBnDCG@10: 75.9 -> 70.0Hurt
replace LM-head vocabulary with SAE latent vocabularySPLADE-Llama (MS MARCO-trained) · MTEB EnglishnDCG@10: 52.9 -> 52.9No clear effect
cap document features at Top-K=100 at inferenceSPLARE vs SPLADE · MS MARCO / MTEB Englishperformance drop: 2% vs over 6% dropImproved
add Top-K inference pooling (40,400) to SPLARE (ablation)SPLARE · MTEB(Multilingual, v2) retrievalnDCG@10: 63.8 -> 62.3Hurt
reduce Top-K from (40,400) to (10,100) (ablation)SPLARE · MTEB(Multilingual, v2) retrievalnDCG@10: 62.3 -> 56.0Hurt
use SAE at layer 6 (SPLARE-2B) instead of layer 26 (ablation)SPLARE · MTEB(Multilingual, v2) retrievalnDCG@10: 62.3 -> 59.1Hurt
increase SAE width (ablation)SPLARE (Gemma-2-2B) · MTEB(Eng, v2) retrievalnDCG@10Improved
use intermediate layer (about two-thirds depth) for SPLARE (ablation)SPLARE · MTEB(Eng, v2) retrievalnDCG@10Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy