When Does Adaptive Halting Win? Selection Rules Matter in Looped Language Models
cs.LG, cs.CL
Submitted: 2026-11-02
Updated: 2026-11-02
Comments: 21 pages, 1 figure. Code: https://github.com/lonewolf15116/lookahead-halting
Code: https://github.com/lonewolf15116/lookahead-halting
License: http://creativecommons.org/licenses/by/4.0/
The gist: Looped language models can stop iterating early, and recent work shows that post-hoc readouts of the loop trajectory often rival the learned exit gate of the looped model Ouro.
Terminology
Abstract
Looped language models can stop iterating early, and recent work shows that post-hoc readouts of the loop trajectory often rival the learned exit gate of the looped model Ouro. Such comparisons reduce each method to one operating point chosen by a selection rule. We ask whether that rule decides the outcome. On a frozen Ouro-1.4B with four loops, we compared a trained stability probe (π 1-x) with Ouro's native gate, adapted to a per-question stopping decision, in pre-registered studies; the confirmatory study was designed after an exploratory finding on development data and evaluated once on a sealed test set of 8,517 questions. In the pre-registered confirmatory comparison, which selected operating points with a bootstrap-screened 1 pp accuracy-loss rule, the probe saved more of the four-loop budget than the native gate (+0.053 gross macro saving over three benchmarks, 95% CI [+0.049, +0.057]), mostly from CommonsenseQA, where the probe's test upper bound on accuracy loss exceeded 1 pp. Under a point-only rule, a pre-specified secondary analysis on the same test set reversed the preference (-0.022 [-0.026, -0.018]), with operating points that exceeded the budget. In pre-specified secondary comparisons, the probe showed no demonstrated advantage over an untrained KL divergence between successive loops (+0.00125 [-0.0041, +0.0065]); we emphasise this result post hoc and claim no equivalence. The probe also broke its budget on one of two out-of-distribution benchmarks (3.26 pp loss on Social IQa). Savings are estimated from stored full-depth traces, not measured inference time, and our screen carries no formal risk-control guarantee. We conclude that halting methods should be compared as signals together with selection procedures, with strong untrained baselines and out-of-distribution checks.
Sources
- Conformal Risk Control
- Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
- PonderNet: Learning to Ponder
- Think-at-Hard: Dynamic Looped Transformers for Improved Reasoning
- Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
- Adaptive Computation Time for Recurrent Neural Networks
- Fast yet Safe: Early-Exiting with Risk Control
- Monitoring latent reasoning in looped language models
- Two-Scale Latent Dynamics for Recurrent-Depth Transformers
- Adaptive Depth in Looped Transformers: Diagnosing Learned Halting Gates and Trajectory Readouts
- Confident Adaptive Language Modeling
- AdaPonderLM: Gated Pondering Language Models with Token-Wise Adaptive Depth
- SpecExit: Accelerating Large Reasoning Model via Speculative Exit
- Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification
- BERT Loses Patience: Fast and Robust Inference with Early Exit
- Scaling Latent Reasoning via Looped Language Models
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis