A method compares search actions to divide shared credit
In plain words
Search systems can make several rewrites of one query and combine the results, but often receive just one score for the whole group. That makes it hard to tell which rewrite helped in combination with the others.
The paper introduces CASI, a method that reuses one set of generated actions and retrieval results to score different groups of actions. It averages each action’s added value across those groups and uses that credit to guide training.
In the reported five-action setting, CASI-trained systems had higher mean value in offline auction replay than systems trained with two comparison rules.
The research
When a search system combines several rewrites and receives one outcome for the whole set, how can training account for an action’s value in the context of the others? The outcome can guide the set as a whole, but it leaves those context-dependent contributions unresolved.
CASI scores alternative groups of already generated search actions, then averages each action’s contribution across contexts to shape training credit.
For each five-action set, CASI reuses cached retrieval results to score alternative groups under a frozen evaluator. Shapley weighting averages each action’s added value across those contexts, and bounded adjustments supply relative action-level training signals.
On 1,000 held-out requests, CASI-trained policies had mean auction-replay value of 66.70 ± 3.40 across five runs, above Singleton’s 55.00 ± 4.00. This tests complete policy outputs under replay, not whether replay gives production counterfactual value.
The alignment result changed with the evaluator, and the replay reference is not validated as production counterfactual value. CASI assigns credit within a fixed generated set; it does not identify how an earlier action changes which later actions the system generates.
Analysis
Deep
- Original paper authors
- Xinyi Cheng, Tianfu Zhang, Jiwen Li, Xuesen Zhang, Shihao Yu
- Submitted to arXiv
- License
- arXiv non-exclusive