Xin Zhang

Xin Zhang


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

20 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
XZXin Zhang
Xin Zhang
LGLiyuan Gao
Liyuan Gao
HZHanyao Zeng
Hanyao Zeng
VSVictor Sheng
Victor Sheng
Background: Neuro-symbolic AI combines neural networks with explicit knowledge such as rules, programs or knowledge graphs. Papers describe how the two parts are combined but seldom state which part produces the answer once the system is trained. This decides what the system can guarantee and how it can fail. Objectives: We classify neuro-symbolic systems by the computation that produces the answer at test time and ask what each class implies for design and evaluation. Methods: We assembled 288 papers with title queries, citation snowballing and queries for well-known methods. We read every paper and assigned it to one of five classes with a decision procedure that uses only what papers report. In three classes the answer comes from a neural network, a symbolic procedure or a probabilistic model alone. In the two hybrid classes, neural and symbolic computation both run on every input (tightly coupled) or a rule chooses between separate neural and symbolic answers (loosely coupled). Each paper was also labeled by task, input, kind of output and use of large language models. To see whether evaluations show where errors arise, how often no answer is returned and what an answer costs, we also audited the experiments of 39 randomly sampled tightly coupled papers and all loosely coupled ones. Results: A symbolic or probabilistic procedure helps compute the answer at test time in 85% of the papers whose outputs must be formally valid, against 72% of those that output labels or scores. Of the 39 sampled tightly coupled papers, only 1 separates network errors from symbolic errors by giving the symbolic part correct inputs, 8% report how often no answer is returned and 36% report what an answer costs. Just 12 of the 288 papers let one side answer when the other is skipped or fails, so in almost all hybrid systems a failure on either side becomes a wrong or missing answer. Conclusions: Which computation produces the answer can be read from what papers report, and it says more about what a system guarantees than the neuro-symbolic label does. A symbolic procedure belongs at test time when every output must be valid, and knowledge can enter through training when outputs are labels or scores. Where a symbolic procedure runs at test time, its guarantee covers only the property it checks and only the inputs the network hands it. Papers on hybrid systems should therefore test the symbolic part on correct inputs, and report how often an answer is returned and what it costs, so that systems with and without a symbolic check can be compared. We release every label, so readers can check them and classify new systems with the same decision procedure.