titleLate Entry in Looped Language Models
authorsZiqi Jin, Jiaxi Li, Weiwen Xu, Aixin Sun
date2026-11-18
venuearXiv (preprint, not peer reviewed)
idarXiv:2611.24576
affiliationNanyang Technological University
upvotes9 (Hugging Face Daily Papers, 2026-11-21)
linksabs pdf hf code
cite
relatedBeyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation
AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data
MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents
Focused Omni-Modal GUI Agents
PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space
abstractLooped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth. We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache. Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets. To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero extra serial calls. Projecting these finished T-pass wake states inside the existing output GEMM (Wake verification, governed by logit-margin tolerance tau in [0, inf]) unifies greedy-exact T-pass decoding (tau = 0, matching K = T block decoding within bf16 reduction order), margin-gated verification (tau > 0), and non-rollback lossy late entry (tau = inf) inside a single static CUDA graph. On Ouro-2.6B (T = 4), K = 2 Wake reaches 5.2x greedy-exact speedup at tau = 0 (+26% over K = 4 block decoding on GSM8K) and 6.8x at tau = inf (+16.9 points over trained early exit at matched speed). Scaling test-time recurrence across T in {8, 16, 32, 64} on Huginn-3.5B cuts K = T/4 rollbacks 21.5x with depth, driving greedy-exact (tau = 0) speedups to 6.1x (T = 32) and 8.4x (T = 64, +153% over K = 64 block) and tau = inf speedups to 10.6 to 12.5x.

bibtex

@misc{jin2026late,
  title={Late Entry in Looped Language Models},
  author={Ziqi Jin and Jiaxi Li and Weiwen Xu and Aixin Sun},
  year={2026},
  eprint={2611.24576},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2611.24576}
}