scalingfactbullishTTEL establishes strictly dominating Pareto frontiers across sequential reasoning domains by performing token-level error localization and maximally reusing valid prefixesMachine Learning29 Jul 2026http://arxiv.org/abs/2607.21453v1