Large language models are plagued by a critical flaw: once a reasoning error occurs mid-generation, subsequent tokens often compound the mistake, leading to unrecoverable outputs. This paper introduces a novel solution, Latent Phase-Shift Rollback (LPSR), designed to address this fundamental limitation without requiring fine-tuning or additional forward passes.
Halting Compounding Errors with Latent Phase-Shift Rollback
LPSR operates by monitoring the residual stream at a critical layer during each generation step. It employs a dual gate, combining cosine similarity and entropy, to detect abrupt directional reversals—akin to phase shifts—in the model's internal state. Upon detection, LPSR rolls back the KV-cache and injects a pre-computed steering vector, effectively correcting the erroneous trajectory. This mechanism bypasses the need for gradient computation or further training, offering an inference-time fix. On the MATH-500 benchmark, an 8B model equipped with LPSR achieved a remarkable 44.0% accuracy, a substantial +15.2 percentage point improvement over standard autoregressive generation (28.8%). Crucially, LPSR significantly outperforms prompted self-correction, which scores only 19.8%, by a margin of +24.2 percentage points.
