Transcoder Adapters for Reasoning-Model Diffing

Code for training, analyzing, and serving transcoder adapters, as described in Transcoder Adapters for Reasoning-Model Diffing.

Transcoder adapters learn an interpretable approximation of the difference in MLP computation before and after fine-tuning. Each transformer layer's MLP is extended with a sparse parallel branch: output = original_mlp(x) + dec(relu(enc(x))). Only the adapter parameters are trained; the base model is frozen.

Pretrained Checkpoints

Five checkpoints trained on DeepSeek-R1-Distill-Qwen-7B with different L1 sparsity weights:

L1 Weight	Val L0	HuggingFace
0.01	10.3	nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.01-l0-10.3
0.003	4.3	nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.003-l0-4.3
0.001	1.4	nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.001-l0-1.4
0.0003	0.4	nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.0003-l0-0.4
0.0001	0.1	nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.0001-l0-0.1

Load with the custom model class (not compatible with AutoModel):

from models.qwen2_transcoder import Qwen2ForCausalLMWithTranscoder

model = Qwen2ForCausalLMWithTranscoder.from_pretrained(
    "nathu0/transcoder-adapters-R1-Distill-Qwen-7B-l1w0.001-l0-1.4",
    torch_dtype=torch.bfloat16, device_map="auto",
)

See demo/ for full examples including adapter on/off toggling and vLLM serving.

Training

Install dependencies, then run:

pip install -r requirements.txt
python -m training.train --config training/configs/r1_distil_7b.yaml

Training requires two models from the same architecture family: a base model (e.g., Qwen/Qwen2.5-Math-7B) and a target model (e.g., deepseek-ai/DeepSeek-R1-Distill-Qwen-7B). The adapter learns to approximate the MLP difference between them using bridging losses (KL divergence at sampled layer cutoffs + NMSE activation matching).

Configs are YAML files — see training/configs/ for examples. Key settings:

• model_name: base model (MLP weight donor)
• bridging.reference_model_path: target model (distillation target)
• transcoder.n_features: width of the sparse adapter (e.g., 8192)
• transcoder.l1_weight: L1 sparsity penalty — higher values produce sparser adapters with fewer active features per token
• learning_rate: we used 8e-4 for all experiments

The two most important hyperparameters — learning_rate and l1_weight — can be overridden from the command line without editing the config:

python -m training.train --config training/configs/r1_distil_7b.yaml --learning_rate 1e-3 --l1_weight 0.01

Checkpoints are saved as standard HuggingFace format via model.save_pretrained(). Training logs to Weights & Biases by default.

Data

Training uses a stratified subset of OpenThoughts3-1.2M (70% math, 20% code, 10% science), filtered for complete reasoning traces under 10k tokens.

• Pre-built splits: nathu0/transcoder-adapters-openthoughts3-stratified-55k (49,952 train / 4,996 val)
• Reproduction script: misc_scripts/filter_openthoughts_stratified.py

Analysis

For feature activation collection, classification, auto-interp, and attribution graph generation, see analysis/README.md.

Repo Structure

• models/ — Qwen2 + transcoder model class for training/inference (qwen2_transcoder.py) and vLLM serving (qwen2_transcoder_vllm.py)
• training/ — Training script, config system, losses, dataset loader, and experiment configs
• analysis/ — Feature analysis pipeline: activation collection, LLM-judge classification, auto-interp, and RelP attribution graphs
• demo/ — Example scripts for generation with adapters on/off (transformers and vLLM)
• misc_scripts/ — Dataset filtering, hybrid model creation

Citing

If transcoder adapters or this repository is useful in your own research, you can use the following BibTeX entry:

@misc{hu2026transcoderadaptersreasoningmodeldiffing,
  title={Transcoder Adapters for Reasoning-Model Diffing},
  author={Nathan Hu and Jake Ward and Thomas Icard and Christopher Potts},
  year={2026},
  eprint={2602.20904},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.20904},
}