Research map / Large language models / Retrieval and knowledge
LatentQA: Teaching LLMs to Decode Activations Into Natural Language
Alexander Pan, Lijie Chen, Jacob Steinhardt
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 |
|---|---|---|---|
| train a decoder on LatentQA data (LIT) instead of untrained Patchscope | latent attribute extraction · 6 relational tasks | accuracy: +38.2% absolute on average | Improved |
| use LIT decoder instead of task-specific linear probe | latent attribute extraction · 6 relational tasks | accuracy: +32.2% on average; Ath_Position 75.9 -> 65.2 | Mixed |
| use LIT instead of SelfIE | hidden system prompt uncovering · Hidden system prompt (Gemma-3-4b-it, hard) | accuracy: 16 -> 80 | Improved |
| use LIT on activations vs GPT-4 prompting with full dialog | hidden system prompt uncovering · Hidden system prompt (Gemma-3-4b-it, hard) | accuracy: 76 -> 80 | Improved |
| scale target and decoder model from 1B to 8B (ablation) | LIT · Hidden system prompt, hard personas | accuracy: 12 -> 92 | Improved |
| scale target and decoder model from 1B to 8B (ablation) | LIT · Hidden system prompt, easy personas | accuracy: 30 -> 100 | Improved |
| increase LatentQA training data (1/4 to full) (ablation) | LIT · LatentQA eval loss | test loss | Improved |
| apply LIT steering instead of RepE (LoRRA) for debiasing | Gemma-3-4b-it · CrowS-Pairs mean log-likelihood difference | mean abs log-likelihood difference: 7.88 -> 5.07 | Improved |
| add LIT steering vs no control | Gemma-3-4b-it · CrowS-Pairs percent stereotype | percent stereotype: 57.3 -> 47.6 | Improved |
| add prompting 'Pretend you are unbiased' vs no control | Gemma-3-4b-it · CrowS-Pairs mean log-likelihood difference | mean abs log-likelihood difference: 5.59 -> 8.44 | Hurt |
| steer with LatentQA gradients to rogue persona | Llama-3-8B-Instruct · harmful knowledge elicitation | harmful responses | Improved |
Closest papers by meaning
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.88)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.87)
- ActivationReasoning: Logical Reasoning in Latent Activation Spaces (ICLR 2026 · similarity 0.87)
- Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data (NeurIPS 2024 · similarity 0.87)
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.
