The paper
A new arXiv preprint (2609.27220) from ArXiv CS.CL describes LOCKR, a hidden-state trajectory-guided planner for diffusion language models. The authors identify a reasoning failure they call stable-but-wrong lock-in: an answer stabilizes early around an incorrect value while substantial denoising remains. They report that surface-level decoding signals — confidence, entropy, margin, and answer stability — are not sufficient to reliably separate correct from erroneous lock-in. LOCKR frames selective reasoning repair as a lightweight test-time planning problem: it decides when to spend extra computation, expands targeted repair branches, and picks a continuation using trajectory-aware verification. [1]
Reported results
Across two diffusion language models and three mathematical reasoning benchmarks, the authors report that hidden-state trajectories consistently beat surface signals and single hidden snapshots at both detecting wrong lock-in and selecting repairs. On natural evaluation distributions, the paper claims absolute accuracy gains of 2.21 to 5.37 percentage points across all five evaluated settings, with repair rates between 22% and 41%. These are the authors' own results on a preprint, not independently verified; the benchmarks are mathematical reasoning only. [1]
Sources
- LOCKR: A Hidden-State Trajectory-Guided Planner for Detecting and Repairing Stable-but-Wrong Lock-In in Diffusion Language Models
ArXiv CS.CL (Computation and Language) · Reporting ·