PAR aligns looped model steps for longer state tracking
Original title: PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking
On swap-S5, a Qwen2.5-0.5B model trained on at most 32 moves reaches 95.5% full-state accuracy at 512 moves.
The paper introduces Progress-Aligned Recurrence (PAR) for looped language models doing sequential state tracking. Iteration k is trained to match the state after the first k operations, and recurrent states are rescaled to first-iteration norms. On swap-S5, a Qwen2.5-0.5B model trained on at most 32 moves reaches 95.5% mean full-state accuracy at 512 moves, versus 55.7% reported for RLT and 0.8% for explicit chain of thought on the same backbone. Ablations show removing norm anchoring drops extrapolation below 2.2×, and final-answer supervision caps it at 8×. A learned readout can choose when to answer, and prefix self-distillation is used when intermediate labels are unavailable.
Unconfirmed: The 55.7% RLT and 0.8% chain-of-thought figures are reported comparisons, not remeasured here.
- Training length is at most 32 moves; the reported test length is 512 moves.
- Removing norm anchoring cuts the extrapolation factor from 16× to below 2.2×.
- Final-answer supervision instead of progress targets reduces extrapolation to at most 8×.
- Prefix self-distillation uses only final answers when intermediate labels are missing.
- Authors say the gains hold across sequential state-tracking tasks.
From the original article
1 min read at huggingface.co
Looped language models can run more iterations at test time, but additional iterations need not help them solve longer problems. For sequential state tracking, final-answer supervision does not specify how far each iteration should advance, and repeated updates can change the scale of the recurrent state.
By Pan Wang, Yihao Hu, Guangyuan Dong · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.