generate text by iteratively denoising, and sometimes their answers stabilize early around an incorrect value, a failure mode the authors call stable-but-wrong lock-in. Surface signals like confidence, entropy, margin, and answer stability cannot reliably distinguish correct from erroneous lock-in. The authors propose LOCKR, a test-time planner that uses the hidden-state trajectory to decide when to allocate extra computation, expand targeted repair branches, and select the most promising continuation via trajectory-aware .
Across two diffusion language models and three mathematical reasoning benchmarks, hidden-state trajectories consistently outperform both surface-level signals and single hidden snapshots for detecting wrong lock-in and selecting repairs. On natural evaluation distributions, LOCKR achieves absolute accuracy gains of 2.21 to 5.37 percentage points across all five settings, with repair rates ranging from 22% to 41%. This establishes hidden diffusion trajectories as actionable signals for selective test-time reasoning repair.