AI Paper Atlas

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

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data

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
scale base model from Qwen3-0.6B to Qwen3-14B (ablation)Introspection adapter (SFT) · held-out train-set behaviorsverbalization rate (%): 37.7 -> 77.3Improved
increase number of training behavior categories from one to seven (ablation)Introspection adapter (Qwen3-14B) · held-out train-set behaviorsverbalization rate (%): 20.5 -> 64.4Improved
increase training categories from one to five (ablation)Introspection adapter (Qwen3-14B) · AuditBench OODverbalization rate (%): 1.9 -> 10.0Improved
add DPO refinement stage after SFT (ablation)Introspection adapter (Llama 3.3 70B) · in-distribution and OOD evaluationsverbalization accuracyImproved
include Benign Roleplay dataset in IA training (ablation)Introspection adapter (Llama 70B) · AuditBench misalignment-style modelsverbalization rate (%): 40.5 -> 25.6Hurt
train IA on a model that already contains the behavior (pre-existing behavior) (ablation)Introspection adapter (SFT) · AuditBench Contextual Optimism / Hardcoded Test Casesverbalization rate (%): 22 -> 8 and 26 -> 10Hurt
apply IA to base-model verbalization (add IA vs none) (ablation)Qwen3 models · held-out train-set behaviorsverbalization rate (%)Improved
replace full LoRA with single-layer bias vector (ablation)Introspection adapter (Qwen3-14B) · 8-setting benchmarkintrospection rate vs baseline: 2.38x over baselineImproved
use single-layer rank-1 LoRA instead of full rank-16 all-layer LoRA (ablation)Introspection adapter (Qwen3-14B) · 8-setting benchmarkrelative performance (%): 68.9% of full rank-16 performanceMixed
reduce training samples per behavior 8x (100 to 12) (ablation)Introspection adapter (Qwen-14B) · held-out plus AuditBench-SFT averageverbalization rate: only 6% drop from 8x fewer samplesNo clear effect
exclude Rare-setting models with 0% shown-behavior rate (ablation)Introspection adapter (Qwen14 SFT) · evaluation setsverbalization rateNo clear effect
train on only 40% of behaviors per setting (ablation)Introspection adapter (Llama70B, SFT, 8 settings) · train-set and AuditBenchverbalization rate: ~100% of AuditBench performance retainedNo clear effect

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