One Model to Loop Them All: Elastic Latent Recurrence for Language Models
ELR trains one looped checkpoint to run fast at one loop, then spend extra loops for better accuracy.
The paper says prior latent recurrent models paid a “one-loop tax,” underperforming same-size non-recurrent models when latency mattered. ELR avoids that with a gated state update, mixed loop counts during optimization, and conditioning on loop index and budget. On 4B dense and 30B-A3B MoE hybrids, one-loop ELR roughly matches matched baselines, while four loops add sizable ten-task gains. The authors also report cache and routing tricks that reduce the cost of using extra loops. HF Daily Papers' note
The paper says prior latent recurrent models paid a “one-loop tax,” underperforming same-size non-recurrent models when latency mattered. ELR avoids that with a gated state update, mixed loop counts during optimization, and conditioning on loop index and budget. On 4B dense and 30B-A3B MoE hybrids, one-loop ELR roughly matches matched baselines, while four loops add sizable ten-task gains. The authors also report cache and routing tricks that reduce the cost of using extra loops. HF Daily Papers' note
score 5