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.
ORIGINAL PAPER
PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking
WHAT WE KNOW
We introduce Progress-Aligned Recurrence (PAR), which trains the prediction at iteration k to match the state after the first k operations and rescales recurrent states to their first-iteration norms. On swap-S_5, a Qwen2. 5-0.
WATCH NEXT
Review the primary source, validate the main result, and establish whether any listed-company transmission is direct.
EVIDENCE
What the evidence supports so far
Research signals
reasoning
training-method
scaling
What remains unverified
The full methodology, effect size, and limitations still require analyst review.
Company impact remains unverified until a direct economic transmission is established.