I've stumbled upon a 2018 preprint 'Comparison Training for Computer Chinese Chess' (the final version of the article was published in IEEE Transactions on Games in January 2019) by Wen-Jie Tseng, Jr-Chang Chen, I-Chen Wu, and Ting-Han Wei, who claimed that their modifications of a xiangqi engine Chimo that included eval features that they denoted as LOC2 - features that indicated pairwise interactions of pieces, could be indexed by [piece1color][piece1type][piece1square][piece2color][piece2type][piece2square], for many or even all the pairs of pieces - performed clearly better at very fast TC (0.4 sec/move on Intel Core i5-4690) than the modifications that didn't include such features, despite the former modifications (denoted as EVAL10-13) having >120K eval parameters that were hard to train (and whose training probably hadn't even converged) while the latter (EVAL0-9) had <=730 params.
Seeing how much success A0 and lc0 have had despite having to train millions of weights, I guess that adding and tuning pairwise PSTs with <300K values in total (out of which, at most a few hundred would be used per position) could strengthen an A/B engine, which I certainly wouldn't have imagined before the A0 papers. Is anyone brave enough to spend plenty of effort on this idea? Or is Western chess too different from xiangqi for it to ever work?
The approach can be extended to certain triples of pieces, like a certain piece plus both kings as Pawel initially suggested -
- but that would add even more params, so caution is needed. I've simply brought up an extra data point (the preprint) in support of the general idea of multi-piece PSTs.