What is TISD?
On-policy self-distillation trains a student toward a teacher that is the same model conditioned on privileged context (a successful sibling rollout or execution feedback). TISD changes which trajectories are distilled: for each failed rollout it
- finds the position of peak teacher-student divergence and forces the teacher's preferred token there,
- lets the student regenerate the rest of the response, and
- distills the resulting trajectory toward the privileged-context teacher with the SDPO objective.
Results (LiveCodeBench v6, 131 problems)
Table with columns: Model, Avg@4 (%), Pass@4 (%)| Model | Avg@4 (%) | Pass@4 (%) |
|---|
Qwen/Qwen3-8B (base) | 27.5 | 36.1 |
| TISD-Qwen3-8B (this checkpoint, step 80) | 49.6 | 50.4 |
Avg@4 is the mean test-case pass rate over 4 samples per problem; Pass@4 counts a problem as solved if any of the 4 samples passes all hidden tests. Numbers are for this exact checkpoint (training step 80). Paper Table 1 reports the best validation step, averaged over runs where several exist.
Training setup
Table with columns: Item, Value| Item | Value |
|---|
| Data | LiveCodeBench v6 (train split exposes 50% of each problem's tests) |
| Steps | 80 (batch of 32 prompts, 8 rollouts each) |
| Learning rate | 1e-6 |
| Hardware | NVIDIA B200 GPUs, verl FSDP |
| Weights | bf16, merged from the FSDP actor checkpoint |
Full hyperparameters are in the paper appendix and the code repository.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "Taeckyung/TISD-Qwen3-8B"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16", device_map="auto")
The model uses the base model's chat template.
License
Apache 2.0, the same as the base model Qwen/Qwen3-8B. This is a modified (fine-tuned) version of the base model.
Citation
@misc{lee2026tisd, title = {{TISD}: On-Policy Self-Distillation with Trajectory Intervention}, author = {Lee, Taeckyung and Amankos, Rinat and Kim, Jeonghye and Yoon, Hyungjun and Jin, Woogyeol and Lee, Sung-Ju}, year = {2026}, eprint = {2609.30878}, archivePrefix = {arXiv}, url = {https://arxiv.org/abs/2609.30878}}