More test-time recurrence does not automatically mean more useful reasoning: a model can repeat updates without knowing how much progress each iteration should represent, while its internal state drifts in scale. The key insight here is to supervise both the meaning and the magnitude of each recurrent step, turning extra depth into a continuation of the same state-tracking process.
Key Findings
- At iteration k, the model is trained to predict the state after the first k operations. This gives intermediate computation an explicit progress target rather than relying only on the final answer.
- Recurrent states are rescaled to match their first-iteration norms. Removing this norm anchoring reduces the reported extrapolation range from 16× to below 2.2×, showing that stable state scale is central to long rollouts.
- On swap-S₅, Qwen2.5-0.5B trained on at most 32 moves achieves 95.5% mean full-state accuracy at 512 moves. The paper reports 55.7% for RLT and 0.8% for explicit chain-of-thought on the same backbone.
- A learned completion readout can decide when to answer without a supplied stopping depth; prefix self-distillation also offers a route when intermediate labels are unavailable, reaching 85% at 256 moves in the reported setup.
Method and Scope
PAR combines progress-aligned intermediate targets with norm anchoring. Together, they address two distinct failure modes: an iteration that has no defined relationship to input progress, and recurrent updates whose scale drifts as depth increases. The approach is most compelling for sequential tasks with a meaningful state after each operation. The paper reports that graph reachability does not benefit comparably because its parallel frontier expansion lacks a canonical intermediate state for each iteration.
Who It’s For—and the Trade-Off
A strong fit for researchers studying recurrent language models, algorithmic state tracking, and length extrapolation, especially when intermediate states can be defined or inferred from prefixes. It is less directly applicable to problems without a natural step-by-step state, and the headline results are benchmark-specific; they do not establish broad long-context or general reasoning gains.