AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

LatentQA: Teaching LLMs to Decode Activations Into Natural Language

Alexander Pan, Lijie Chen, Jacob Steinhardt

ICLR 2026 · Large language models › Retrieval and knowledge · approach: Explanatory and multiple-choice QA benchmarks · 35 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
train a decoder on LatentQA data (LIT) instead of untrained Patchscopelatent attribute extraction · 6 relational tasksaccuracy: +38.2% absolute on averageImproved
use LIT decoder instead of task-specific linear probelatent attribute extraction · 6 relational tasksaccuracy: +32.2% on average; Ath_Position 75.9 -> 65.2Mixed
use LIT instead of SelfIEhidden system prompt uncovering · Hidden system prompt (Gemma-3-4b-it, hard)accuracy: 16 -> 80Improved
use LIT on activations vs GPT-4 prompting with full dialoghidden system prompt uncovering · Hidden system prompt (Gemma-3-4b-it, hard)accuracy: 76 -> 80Improved
scale target and decoder model from 1B to 8B (ablation)LIT · Hidden system prompt, hard personasaccuracy: 12 -> 92Improved
scale target and decoder model from 1B to 8B (ablation)LIT · Hidden system prompt, easy personasaccuracy: 30 -> 100Improved
increase LatentQA training data (1/4 to full) (ablation)LIT · LatentQA eval losstest lossImproved
apply LIT steering instead of RepE (LoRRA) for debiasingGemma-3-4b-it · CrowS-Pairs mean log-likelihood differencemean abs log-likelihood difference: 7.88 -> 5.07Improved
add LIT steering vs no controlGemma-3-4b-it · CrowS-Pairs percent stereotypepercent stereotype: 57.3 -> 47.6Improved
add prompting 'Pretend you are unbiased' vs no controlGemma-3-4b-it · CrowS-Pairs mean log-likelihood differencemean abs log-likelihood difference: 5.59 -> 8.44Hurt
steer with LatentQA gradients to rogue personaLlama-3-8B-Instruct · harmful knowledge elicitationharmful responsesImproved

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