CROSS DOMAIN LATENT SPACE ALIGNMENT FOR ZERO SHOT KNOWLEDGE TRANSFER IN LARGE SCALE NEURO SYMBOLIC FOUNDATION MODELS

The large-scale adoption of neuro-symbolic foundation models poses a critical challenge: zero-shot knowledge transfer is hindered by mismatches in latent distributions, modality heterogeneity, and discrepancies between symbolic and neural representations. In the following paper, a Cross-Domain Latent Space Alignment (CDLSA) framework will be proposed to support strong zero-shot generalisation across structurally different domains without task-specific fine-tuning. Its fundamental issue is the discrepancy between high-dimensional neural representations and structured symbolic representations that, under most circumstances, results in performance drops of 25-35 % when tested on unseen domains. The presented methodology combines contrastive latent projection, probabilistic regularisation of the manifold, and symbolic encoding of constraints into a single optimisation goal. The architecture uses a dual-encoder with shared latent anchors to reduce the Maximum Mean Discrepancy (MMD) and the Kullback-Leibler (KL) divergence between domains while maintaining semantic consistency. Multi-domain experimental assessments of vision-language and logic-reasoning have shown that zero-shot accuracy increases by 21.8%, with a 17.3% error rate in cross-domain generalisation. Latent alignment minimised inter-domain distribution variance by 32.6% and had a higher symbolic consistency score of 0.84 (F1-measure) compared to the initial score of 0.68. Robustness (p<0.01) was tested across five independent runs, and the average effect size (Cohen's d) was 0.79, which is strong enough to be considered practical. Also, the calibration error was reduced by 14.5%, resulting in a more accurate estimate of uncertainty. The findings show that structured latent alignment provides a significant boost to knowledge transfer in neuro-symbolic systems in the zero-shot setting. The CDLSA framework is proposed to create a scalable pipeline from neural representation learning to symbolic reasoning, enabling more reliable and interpretable foundation models in heterogeneous data settings.

Hali tarjima qilinmagan