datkinson/identifying-introspection

Identifying Introspection From the Inside: trained adapters

LoRA adapters from Identifying Introspection From the Inside (Atkinson, Plunkett & Bau, COLM 2026). Code: github.com/diatkinson/identifying-introspection-from-the-inside. Project page: iii.baulab.info.

The adapters teach a model to make A/B decisions on behalf of fictitious characters according to hidden linear preference weights. They are never trained to report those weights; some of them nevertheless report them faithfully. Each folder is one PEFT adapter and contains the personas.json file with its characters and their target weights.

Folder	Base model	What it is
deft-raptor/checkpoint-<step>	Qwen3-32B	Multi-character adapter (100 characters; rank-8 LoRA on all layers), every saved checkpoint up to step 4000 (Figure 1b). Checkpoint 1000 is the paper's unfaithful model and checkpoint 3000 its faithful model (Figures 1, 3, 5).
qwen3-14b-layer-freezing/first-<k>	Qwen3-14B	Multi-character adapters trained only on layers [0, k) (Figure 4).
qwen3-14b-layer-freezing/from-<k>	Qwen3-14B	Trained only on layers [k, 40) (Appendix E).
single-character/<character>/deft-raptor-<1000|3000>	Qwen3-32B + deft-raptor	The 32 pairs of single-character adapters (rank 2) of Figure 5. Each pair is the same character, data and seeds, trained on top of the early (unfaithful) or late (faithful) deft-raptor checkpoint. Use them stacked on that checkpoint (training_config.json). Their final-eval/metrics.json holds decision performance (behavior_goal_corr) and faithfulness (behavior_report_corr), from 32 decision and 16 self-report prompts.
gemma-4-31b/faithful-run/checkpoint-{469,1407}, gemma-4-31b/unfaithful-run/checkpoint-4688	Gemma-4-31B-it	Separately trained high- and low-faithfulness multi-character adapters (Appendix B.6). These are derivatives of Gemma-4 and subject to its license.

Usage

All of the paper's experiments load the base model in 8-bit (bitsandbytes). With the paper's code (pip install -e . in the GitHub repository), any script accepts hf://<this repo>/<folder> wherever it takes an adapter. With PEFT directly:

from peft import PeftModel
from transformers import AutoModelForCausalLM, BitsAndBytesConfig

repo = "datkinson/identifying-introspection"
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B", quantization_config=BitsAndBytesConfig(load_in_8bit=True), device_map="auto"
)
model = PeftModel.from_pretrained(base, repo, subfolder="deft-raptor/checkpoint-3000", adapter_name="backbone")

# A single-character adapter is stacked on its backbone; both are active at once.
model.load_adapter(repo, subfolder="single-character/bird_Willy-Wonka_QkVACLaS/deft-raptor-3000", adapter_name="character")
model.base_model.set_adapter(["backbone", "character"])

Prompts must use the exact templates of the paper (non-reasoning mode, decision and self-report in separate context windows); see introspection.characters in the code repository.

Citation

@inproceedings{atkinson2026identifying,
  title         = {Identifying Introspection From the Inside},
  author        = {Atkinson, David I. and Plunkett, Dillon and Bau, David},
  booktitle     = {Conference on Language Modeling (COLM)},
  year          = {2026},
  eprint        = {2610.07186},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.07186}
}