TurnOPD: Turn-Aware Distillation for Long-Horizon Agents¶
Local core-mechanism evidence. Paper outcomes and local results are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | TurnOPD (arXiv 2607.05804) |
| Affiliations | Academic author team |
| First public release | 2026-07-07 |
| Authors' code | No official author repository found as of 2026-07-29. |
| Adapter / method | turn-opd |
| Local code | src/auto_research/agent_research/ |
Original paper¶
Background and contribution¶
Use probe statistics to choose rollout depth adaptively, gradually moving token KL budget into turn-normalized supervision.
flowchart LR
P[Turn probe] --> B[Adaptive depth] --> R[Truncated trajectory] --> N[Turn-normalized KL] --> U[Student update]Key figure from the paper¶
Original Figure 1: TurnOPD. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
L^\star=\operatorname{Budget}(\hat I_1,\ldots,\hat I_T),\quad \mathcal L=\sum_{t\le L^\star}\frac{1}{|y_t|}\mathrm{KL}(\pi_\theta^t\Vert\pi_T^t).
\]
Paper results¶
At equal wall time, the paper improves validation accuracy and the accuracy–time frontier on ALFWorld, WebShop, and Multi-Hop Search. No production A/B.
Local reproduction¶
ScaleMCP mini-suite executes depth probing, truncation, and normalization. Joint success 1.0000, average cost 1.3333, 40 turns saved. Fixed-seed results.
Reproduction limits¶
Teacher behavior comes from deterministic task oracles, not a task-specific LLM. Mechanism only.