CAST: Game solvers as turn-level teachers¶

Deterministic mini-suite mechanism, not large-model game training.

Paper information¶

Field	Details
Paper	CAST (arXiv 2607.25308)
Institution	USTC / Nanjing University / Wuhan University.
First public release	2026-07-28
Author code	Wloner0809/CAST.
Local adapter or method key	`cast`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

Use successive differences in solver state value as a solver advantage that supplies turn-level credit alongside sparse outcome reward.

flowchart LR
  S[Adjacent environment states] --> V[Query solver value]
  V --> D[Value difference]
  D --> A[Add to outcome advantage]

Original paper figure¶

The original figure shows solver-supported credit. From the original paper; copyright remains with the authors.

Core equation¶

\[ A_t^{\mathrm{solver}}=V_{\mathrm{solver}}(s_{t+1})-V_{\mathrm{solver}}(s_t),\quad A_t=A^{\mathrm{outcome}}+\lambda A_t^{\mathrm{solver}}. \]

Reported results¶

The paper reports beating baselines in Sokoban, Minesweeper and Rush Hour on seen and unseen difficulty, with highest average zero-shot results on ALFWorld and WebShop. No production A/B.

Local reproduction¶

Deterministic shortest paths supply state values in PlanBench mini. Joint success 1.0000, average cost 1.5000, solver queries / turn updates 2040 / 360.

auto-research agent-eval --method cast --episodes 120 --seed 42

Fixed metrics. Accurate solver verifies credit assignment, not LLM capability or large-scale RLVR on the three games.