🤖 AI Summary
This work addresses the opacity of neural network internal states during program execution, which hinders verification of correctness and trustworthiness. The authors propose a symbolic neural CPU architecture that enables interpretable program execution through trajectory supervision, explicitly modeling instruction selection, register reads and writes, memory access, and write-back semantics. This architecture is the first to support verifiable full execution trajectories, preserving symbolic path consistency under quantized simulation and incorporating a fixed-point replay mechanism to eliminate numerical drift. By integrating loop control, differentiable ALU routing, target-masked write-back, behavioral cloning, and Actor-Critic reinforcement learning—while aligning with RV32I instruction set semantics—the system precisely reproduces reference executions at 16-bit precision and maintains symbolic path consistency over thousands of instructions even at 8-bit quantization, demonstrating controllable and interpretable low-precision neural execution.
📝 Abstract
Neural networks can learn algorithmic input-output mappings, but trusting a learned executor requires more than a correct final answer because the state transitions that produce it are usually hidden. To make those transitions visible, we introduce a trace-supervised symbolic neural CPU, a factorized learned execution architecture that combines recurrent control, an explicit operation router over a fixed differentiable arithmetic-logic unit bank, destination-masked register writeback, complete trajectory supervision and matched fixed-point replay. The model exposes the selected operation, source and destination registers, register trajectory, memory signals and writeback semantics at every step. On the principal 16-wide benchmark, the non-quantized executor reproduces reference execution exactly, while the eight-bit quantization-simulated executor preserves the symbolic operation path through programs of 1,000 instructions. When the same execution is evaluated against a matched fixed-point replay, the residual numerical drift disappears, showing that it comes from a mismatch between continuous and low-precision reference semantics rather than from execution failure. We compare recurrent, Transformer, temporal-convolution, temporal graph-inspired and state-space controllers, and the ablations show that operation-gate supervision is necessary for an inspectable execution path. Hidden-opcode memory-pressure tasks expose the remaining limits in delayed state use and temporal binding. We also extend the interface with ValueMemory, hybrid adaptive leaky integrate-and-fire controllers, candidate-constrained symbolic control trained through behaviour cloning and actor-critic reinforcement learning, and an RV32I base-integer semantic bridge. Together, these results establish a trace-verifiable framework for interpretable, low-precision and controllable neural execution.