Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.09734 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914551606280192 |
|---|---|
| author | Reddy, Vishnu Vardhan Chatterjee, Sagnik Bhatta, Soumik |
| author_facet | Reddy, Vishnu Vardhan Chatterjee, Sagnik Bhatta, Soumik |
| contents | Most language-model training data shows final artifacts, not the process that produced them. We study a tractable version of this question in tool use: when a model learns a stream of new API domains, does keeping tool-use trajectories help compared with stripping the intermediate API trace? We fine-tune Llama 3.1 8B Instruct with QLoRA on API-Bank using four sequential domain blocks. Condition A strips previous API request/response lines from the prompt and trains the model to predict the next API call. Condition B keeps the trajectory context. In a single-seed pilot, full held-out generation evaluation shows that Condition B reaches 56.9\% final exact full-call accuracy compared with 39.2\% for Condition A. B also improves final API-name accuracy by 7.7 points. However, B uses 25.1\% more training tokens, the run uses one seed, and the task is next-call prediction rather than full dialogue success. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_09734 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Trajectory Supervision for Continual Tool-Use Learning in LLMs Reddy, Vishnu Vardhan Chatterjee, Sagnik Bhatta, Soumik Software Engineering Artificial Intelligence Multiagent Systems Most language-model training data shows final artifacts, not the process that produced them. We study a tractable version of this question in tool use: when a model learns a stream of new API domains, does keeping tool-use trajectories help compared with stripping the intermediate API trace? We fine-tune Llama 3.1 8B Instruct with QLoRA on API-Bank using four sequential domain blocks. Condition A strips previous API request/response lines from the prompt and trains the model to predict the next API call. Condition B keeps the trajectory context. In a single-seed pilot, full held-out generation evaluation shows that Condition B reaches 56.9\% final exact full-call accuracy compared with 39.2\% for Condition A. B also improves final API-name accuracy by 7.7 points. However, B uses 25.1\% more training tokens, the run uses one seed, and the task is next-call prediction rather than full dialogue success. |
| title | Trajectory Supervision for Continual Tool-Use Learning in LLMs |
| topic | Software Engineering Artificial Intelligence Multiagent Systems |
| url | https://arxiv.org/abs/2605.09734 |