Research map / Large language models / Instruction and fine-tuning
Introspection Adapters: Training LLMs to Report Their Learned Behaviors
Keshav Shenoy, Li Yang, Abhay Sheshadri, Jack Lindsey, Samuel Marks, Rowan Wang
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 |
|---|---|---|---|
| scale base model from Qwen3-0.6B to Qwen3-14B (ablation) | Introspection adapter (SFT) · held-out train-set behaviors | verbalization rate (%): 37.7 -> 77.3 | Improved |
| increase number of training behavior categories from one to seven (ablation) | Introspection adapter (Qwen3-14B) · held-out train-set behaviors | verbalization rate (%): 20.5 -> 64.4 | Improved |
| increase training categories from one to five (ablation) | Introspection adapter (Qwen3-14B) · AuditBench OOD | verbalization rate (%): 1.9 -> 10.0 | Improved |
| add DPO refinement stage after SFT (ablation) | Introspection adapter (Llama 3.3 70B) · in-distribution and OOD evaluations | verbalization accuracy | Improved |
| include Benign Roleplay dataset in IA training (ablation) | Introspection adapter (Llama 70B) · AuditBench misalignment-style models | verbalization rate (%): 40.5 -> 25.6 | Hurt |
| train IA on a model that already contains the behavior (pre-existing behavior) (ablation) | Introspection adapter (SFT) · AuditBench Contextual Optimism / Hardcoded Test Cases | verbalization rate (%): 22 -> 8 and 26 -> 10 | Hurt |
| apply IA to base-model verbalization (add IA vs none) (ablation) | Qwen3 models · held-out train-set behaviors | verbalization rate (%) | Improved |
| replace full LoRA with single-layer bias vector (ablation) | Introspection adapter (Qwen3-14B) · 8-setting benchmark | introspection rate vs baseline: 2.38x over baseline | Improved |
| use single-layer rank-1 LoRA instead of full rank-16 all-layer LoRA (ablation) | Introspection adapter (Qwen3-14B) · 8-setting benchmark | relative performance (%): 68.9% of full rank-16 performance | Mixed |
| reduce training samples per behavior 8x (100 to 12) (ablation) | Introspection adapter (Qwen-14B) · held-out plus AuditBench-SFT average | verbalization rate: only 6% drop from 8x fewer samples | No clear effect |
| exclude Rare-setting models with 0% shown-behavior rate (ablation) | Introspection adapter (Qwen14 SFT) · evaluation sets | verbalization rate | No clear effect |
| train on only 40% of behaviors per setting (ablation) | Introspection adapter (Llama70B, SFT, 8 settings) · train-set and AuditBench | verbalization rate: ~100% of AuditBench performance retained | No clear effect |
Closest papers by meaning
- Looking Inward: Language Models Can Learn About Themselves by Introspection (ICLR 2025 · similarity 0.89)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.87)
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.87)
- Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · similarity 0.86)
- Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.85)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.85)
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.
