GLINN: General Language Interface for Neural Networks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Continuous internal activations in deep neural networks often encode complex strategic decisions that resist direct human interpretability. In this paper, we introduce GLINN (General Language Interface for Neural Networks), a unified bidirectional framework that translates continuous thought vectors (z ∈ ℝ⁷⁶⁸) of an expert dual-head neural network (SE- ResNet-20) into natural language explanations, and conversely reconstructs, compiles, and steers latent vectors from text alone. We identify and solve a critical failure mode in unweighted high-dimensional autoencoding: the 0.13% dimensional dilution trap, where a 1D evaluation manifold receives negligible gradient signal under standard cosine distance, causing massive positive centroid drift (> +25 logits) and freezing final evaluation outputs in the saturated flat tail of tanh. By attaching the chess engine's pre-tanh value head output layer directly into the GLINN co-training loop and optimizing solely on evaluation difference via non-saturating Smooth L1 backpropagation and Group Relative Policy Optimization (GRPO), we achieve: (1) a 68.6% reduction in logit calibration error (24.85 → 7.80) and a 54.8% reduction in position evaluation MAE (0.875 → 0.396); (2) decisive sign agreement on critical won/lost tactical positions (flipping reconstructed logits from +16.12 to -10.35 on decisive Black-winning boards, matching the engine's -1.000 evaluation); (3) complete preservation of global latent fidelity (99.7% cosine retention); and (4) direct causal natural language steering, where counterfactually editing natural language explanations causally shifts the engine's internal evaluation by up to 9.23 logits and induces full win/loss decision reversals.