LoopUS: Turning Pretrained LLMs into Looped Latent Reasoning Models Without Retraining From Scratch

Forum topic · 小凯 · 2026-06-12

Summary

LoopUS (Looped Depth Up-Scaling), a post-training framework from Pusan National University, converts pretrained LLMs into looped latent refinement models without architectural surgery or full retraining. The method partitions transformer layers by cosine distance of adjacent hidden states: encoder layers, a middle looped reasoning block, and a decoder layer. A Mamba-style selective gate damps hidden-state drift during iteration via convex interpolation, while random deep supervision (sampling K=5 of B=20 loop steps for backprop) cuts training token requirements to 3B—17-20x fewer than prior looped approaches (McLeish et al., 52B; Bae et al., 60B). A confidence head adaptively stops iteration (average 3.39 steps on Qwen3-4B, max 8), and a monotonicity loss ensures each loop step is corrective. On Qwen3-1.7B, LoopUS reduces WikiText perplexity by 19.5% and LAMBADA perplexity by 39.1%, with consistent accuracy gains up to Phi-4 14B (+1.7 avg). Per-iteration KV caches also yield up to 2.49x generation speedup on Qwen3-8B.

LoopUS (Looped Depth Up-Scaling) is a post-training framework from Pusan National University and Changwon National University that recasts pretrained LLMs into looped latent refinement models—preserving pretrained weights and only reusing middle layers repeatedly at inference time.

Key points

Layer partitioning by cosine distance: Analyzing Qwen3-1.7B shows early layers rapidly transform token embeddings, middle layers evolve slowly in a stable plateau, and the final layer jumps toward the output vocabulary. Middle layers are therefore looped; first/last layers act as encoder/decoder.
Selective gating (Mamba-style damping): To prevent hidden-state drift during looping, a learned input-dependent decay factor α ∈ (0,1) performs convex interpolation between the current state and the block's output: h^(b+1) = α ⊙ M(h^(b)) + (1-α) ⊙ h^(b).
Random deep supervision: Forward unrolls B=20 loop steps, but gradients are computed on a uniformly sampled K=5 steps per batch (others run under no_grad). This stabilizes training and cuts training to only 3B tokens (FineWeb-Edu + CC-MAIN-2025-26), versus 52B (McLeish et al. 2025) and 60B (Bae et al. 2025) for prior looped methods—17-20x fewer.
Confidence head: A lightweight head predicts whether the hidden state is good enough; iteration stops when q ≥ 0.6 (max 8 steps). Average stopping depth on Qwen3-4B is 3.39 steps—compute is allocated on demand.
Monotonicity loss: L_mono^(b) = SiLU(L_LM^(b) − L_LM^(b−1)) penalizes any loop step that worsens next-token loss, ensuring each iteration is corrective.

Results

Qwen3-1.7B: WikiText ppl 21.0 → 16.9 (−19.5%); LAMBADA ppl −39.1%; average accuracy 53.7 → 55.3.
Qwen3-4B: average accuracy 60.3 → 62.1; WikiText ppl −15.2%.
Qwen3-8B: average accuracy 63.2 → 65.4.
Phi-4 14B: average accuracy 67.0 → 68.6.
TinyLlama 1.1B comparison: LoopUS reaches 49.4 avg accuracy with 3B tokens, matching or beating methods trained on 52B/60B tokens.
Hidden-state dynamics: token-level probability for "32 * 64 =" evolves from ~2×10⁻⁵% to 81.9% after one iteration and 89.8% after four; trajectories converge to a fixed point rather than diverging.
KV-cache speedup: with per-iteration KV caches (encoder/decoder run once), generation of 1024 tokens speeds up 1.64x (1.7B), 2.31x (4B), and 2.49x (8B).

Training setup

AdamW, lr 5×10⁻⁵, cosine schedule with 300 warmup steps, bf16 mixed precision, seq len 1024, FlashAttention-2, B=20 loop depth with K=5 supervised steps, max 8 inference loops. Hardware: L40S / RTX PRO 6000 / H200.

Ablations

Removing any component—selective gating, encoder-decoder partitioning, random deep supervision, decay gating (vs sigmoid), SiLU (vs ReLU/SELU/SoftPlus), or replacing deep supervision with TBPTT—degrades final loss or stability. All components are necessary.

Limitations

Validated only on autoregressive transformers; loop depth is fixed at B=20 for training (deeper loops unexplored); the 0.6 confidence threshold is heuristic. Future directions include diffusion LLMs, natively loop-aware pretraining, and much deeper loops.

Conclusion

LoopUS argues that pretrained transformer depth can be reinterpreted as a budget for adaptive test-time compute: partition the layers, gate the loop, subsample gradients, and stop adaptively—all in latent space, without outputting intermediate steps or altering the architecture.

References: Park, Lee, Kim, & Bae, "LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models", arXiv:2605.11011. Project page: https://thrillcrazyer.github.io/LoopUS