Looped transformers with recall and outer normalization produce reachable, input-dependent fixed points with stable gradients, enabling generalization, while those without recall cannot; a new internal recall variant performs competitively or better.
hub Canonical reference
Universal Transformers
Canonical reference. 89% of citing Pith papers cite this work as background.
64
Pith papers citing it
Background
89% of classified citations
abstract
Recurrent neural networks (RNNs) sequentially process data by updating their state with each new data point, and have long been the de facto choice for sequence modeling tasks. However, their inherently sequential computation makes them slow to train. Feed-forward and convolutional architectures have recently been shown to achieve superior results on some sequence modeling tasks such as machine translation, with the added advantage that they concurrently process all inputs in the sequence, leading to easy parallelization and faster training times. Despite these successes, however, popular feed-forward sequence models like the Transformer fail to generalize in many simple tasks that recurrent models handle with ease, e.g. copying strings or even simple logical inference when the string or formula lengths exceed those observed at training time. We propose the Universal Transformer (UT), a parallel-in-time self-attentive recurrent sequence model which can be cast as a generalization of the Transformer model and which addresses these issues. UTs combine the parallelizability and global receptive field of feed-forward sequence models like the Transformer with the recurrent inductive bias of RNNs. We also add a dynamic per-position halting mechanism and find that it improves accuracy on several tasks. In contrast to the standard Transformer, under certain assumptions, UTs can be shown to be Turing-complete. Our experiments show that UTs outperform standard Transformers on a wide range of algorithmic and language understanding tasks, including the challenging LAMBADA language modeling task where UTs achieve a new state of the art, and machine translation where UTs achieve a 0.9 BLEU improvement over Transformers on the WMT14 En-De dataset.
hub tools
citation-role summary
background 8
method 1
citation-polarity summary
representative citing papers
Long-range dependency in integer multiplication is a mirage from 1D representation; a 2D grid reduces it to local 3x3 operations, letting a 321-parameter neural cellular automaton generalize perfectly to inputs 683 times longer than training while Transformers fail.
Neural networks exhibit grokking on small algorithmic datasets, achieving perfect generalization well after overfitting.
GPT-f, a transformer-based prover for Metamath, generated new short proofs that were accepted into the main library—the first such contribution from a deep-learning system.
Dynamic skipping and looping of LLM layers via a learned prediction network improves math reasoning accuracy while using fewer layers than standard or prior dynamic-depth inference.
LoopMoE is a looped MoE language model that outperforms matched vanilla MoE on 8 of 9 downstream benchmarks at 3B scale and continues to outperform at 9B scale under strictly controlled budgets.
A sleep mechanism with N offline recurrent passes consolidates context into fast weights, improving performance on reasoning tasks where standard transformers fail.
Minimal weight norm of fixed-precision looped neural networks equals Kolmogorov complexity of output string up to log factor, making weight decay match the optimal universal prior up to polynomial factor.
LoopUS converts pretrained LLMs into looped latent refinement models via block decomposition, selective gating, random deep supervision, and confidence-based early exiting to improve reasoning performance.
LoopQ provides a loop-aware PTQ framework for recursive Transformers that mitigates distribution shift, state reuse, and recursive error accumulation, yielding 68.8% higher average accuracy and 87.7% lower perplexity under W4A4 versus static baselines.
LoopCTR trains CTR models with recursive layer reuse and process supervision so that zero-loop inference outperforms baselines on public and industrial datasets.
DepthVAR adaptively allocates per-token computational depth in VAR models using a cyclic rotated scheduler and dynamic layer masking to achieve 2.3-3.1x inference speedup with minimal quality loss.
Looped LLMs converge to distinct cyclic fixed points per layer, repeating feedforward-style inference stages across recurrences.
LA-Sign achieves state-of-the-art skeleton-based sign language recognition on WLASL and MSASL by using recurrent looped transformers with adaptive hyperbolic geometry alignment.
CCDD defines a joint multimodal diffusion on continuous representation space and discrete token space to combine expressivity with explicit token supervision for diffusion language models.
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
ALBERT reduces BERT parameters via embedding factorization and layer sharing, adds inter-sentence coherence pretraining, and reaches SOTA on GLUE, RACE, and SQuAD with fewer parameters than BERT-large.
BERT stores relational knowledge extractable via cloze queries without fine-tuning and matches supervised baselines on open-domain QA tasks.
A looped padded Transformer with parallel gold-CoT cross-entropy supervision matches explicit CoT accuracy at 3B scale and is 2.5–6.9× faster in the thought phase.
Stochastic loop counts during training of looped transformers reduce OOD variance on binary addition, Dyck-1, Unique Set and Copy tasks, with learned RL-Halting further improving the accuracy-stability trade-off.
DiffusionGemma matches Gemma 4 in variable transparency and monitorability after applying an interpretable token bottleneck, despite higher naive serial depth, and shows novel phenomena such as non-chronological reasoning.
LARM enables test-time compute scaling in non-autoregressive ASR via depth-conditioned looping with CTC checkpoints, supervision embeddings, FiLM conditioning, and posterior feedback, yielding lower WER on LibriSpeech with more loops.
Q-K=V projection sharing in transformers matches standard QKV performance with 50% KV cache reduction and combines with GQA/MQA for up to 96.9% reduction across vision and language tasks.
citing papers explorer
-
Stability and Generalization in Looped Transformers
Looped transformers with recall and outer normalization produce reachable, input-dependent fixed points with stable gradients, enabling generalization, while those without recall cannot; a new internal recall variant performs competitively or better.
-
On the Mirage of Long-Range Dependency, with an Application to Integer Multiplication
Long-range dependency in integer multiplication is a mirage from 1D representation; a 2D grid reduces it to local 3x3 operations, letting a 321-parameter neural cellular automaton generalize perfectly to inputs 683 times longer than training while Transformers fail.
-
Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets
Neural networks exhibit grokking on small algorithmic datasets, achieving perfect generalization well after overfitting.
-
Generative Language Modeling for Automated Theorem Proving
GPT-f, a transformer-based prover for Metamath, generated new short proofs that were accepted into the main library—the first such contribution from a deep-learning system.
-
Skip a Layer or Loop It? Learning Program-of-Layers in LLMs
Dynamic skipping and looping of LLM layers via a learned prediction network improves math reasoning accuracy while using fewer layers than standard or prior dynamic-depth inference.
-
LoopMoE: Unifying Iterative Computation with Mixture-of-Experts for Language Modeling
LoopMoE is a looped MoE language model that outperforms matched vanilla MoE on 8 of 9 downstream benchmarks at 3B scale and continues to outperform at 9B scale under strictly controlled budgets.
-
Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference
A sleep mechanism with N offline recurrent passes consolidates context into fast weights, improving performance on reasoning tasks where standard transformers fail.
-
Neural Weight Norm = Kolmogorov Complexity
Minimal weight norm of fixed-precision looped neural networks equals Kolmogorov complexity of output string up to log factor, making weight decay match the optimal universal prior up to polynomial factor.
-
LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models
LoopUS converts pretrained LLMs into looped latent refinement models via block decomposition, selective gating, random deep supervision, and confidence-based early exiting to improve reasoning performance.
-
LoopQ: Quantization for Recursive Transformers
LoopQ provides a loop-aware PTQ framework for recursive Transformers that mitigates distribution shift, state reuse, and recursive error accumulation, yielding 68.8% higher average accuracy and 87.7% lower perplexity under W4A4 versus static baselines.
-
LoopCTR: Unlocking the Loop Scaling Power for Click-Through Rate Prediction
LoopCTR trains CTR models with recursive layer reuse and process supervision so that zero-loop inference outperforms baselines on public and industrial datasets.
-
Depth Adaptive Efficient Visual Autoregressive Modeling
DepthVAR adaptively allocates per-token computational depth in VAR models using a cyclic rotated scheduler and dynamic layer masking to achieve 2.3-3.1x inference speedup with minimal quality loss.
-
A Mechanistic Analysis of Looped Reasoning Language Models
Looped LLMs converge to distinct cyclic fixed points per layer, repeating feedforward-style inference stages across recurrences.
-
LA-Sign: Looped Transformers with Geometry-aware Alignment for Skeleton-based Sign Language Recognition
LA-Sign achieves state-of-the-art skeleton-based sign language recognition on WLASL and MSASL by using recurrent looped transformers with adaptive hyperbolic geometry alignment.
-
Coevolutionary Continuous Discrete Diffusion: Make Your Diffusion Language Model a Latent Reasoner
CCDD defines a joint multimodal diffusion on continuous representation space and discrete token space to combine expressivity with explicit token supervision for diffusion language models.
-
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
-
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
-
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
ALBERT reduces BERT parameters via embedding factorization and layer sharing, adds inter-sentence coherence pretraining, and reaches SOTA on GLUE, RACE, and SQuAD with fewer parameters than BERT-large.
-
Language Models as Knowledge Bases?
BERT stores relational knowledge extractable via cloze queries without fine-tuning and matches supervised baselines on open-domain QA tasks.
-
Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers
A looped padded Transformer with parallel gold-CoT cross-entropy supervision matches explicit CoT accuracy at 3B scale and is 2.5–6.9× faster in the thought phase.
-
Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping
Stochastic loop counts during training of looped transformers reduce OOD variance on binary addition, Dyck-1, Unique Set and Copy tasks, with learned RL-Halting further improving the accuracy-stability trade-off.
-
How Transparent is DiffusionGemma?
DiffusionGemma matches Gemma 4 in variable transparency and monitorability after applying an interpretable token bottleneck, despite higher naive serial depth, and shows novel phenomena such as non-chronological reasoning.
-
Test-Time Compute Scaling for ASR with Depth-Conditioned Looped Transformers
LARM enables test-time compute scaling in non-autoregressive ASR via depth-conditioned looping with CTC checkpoints, supervision embeddings, FiLM conditioning, and posterior feedback, yielding lower WER on LibriSpeech with more loops.
-
Do Transformers Need Three Projections? Systematic Study of QKV Variants
Q-K=V projection sharing in transformers matches standard QKV performance with 50% KV cache reduction and combines with GQA/MQA for up to 96.9% reduction across vision and language tasks.
-
Fixed-Point Masked Generative Modeling
FP-MGMs with consistency loss and three-state reuse (CoFRe) reduce parameters by up to 38.8% and improve low-budget perplexity and FID versus standard masked generative models on text and images.
-
Contribution Weights: A Geometrical Analysis of Self-Attention Transformers
Contribution Weights combine attention, value magnitude, and directional alignment to measure token influence more faithfully than attention alone, and show attention sinks actively suppress information via a convex sink-rate to output-norm relationship.
-
Looped Diffusion Language Models
LoopMDM loops early-middle layers in masked diffusion models to match same-size MDM performance with up to 3.3x fewer training FLOPs and outperform on reasoning tasks by up to 8.5 points on GSM8K.
-
Generative Recursive Reasoning
GRAM is a latent-variable generative model that performs recursive reasoning via stochastic trajectories, trained with amortized variational inference to support multi-hypothesis reasoning and unconditional generation.
-
Elastic Attention Cores for Scalable Vision Transformers
VECA learns effective visual representations using core-periphery attention where patches interact exclusively via a resolution-invariant set of learned core embeddings, achieving linear O(N) complexity while maintaining competitive performance.
-
ZAYA1-8B Technical Report
ZAYA1-8B is a reasoning MoE model with 700M active parameters that matches larger models on math and coding benchmarks and reaches 91.9% on AIME'25 via Markovian RSA test-time compute.
-
State Stream Transformer (SST) V2: Parallel Training of Nonlinear Recurrence for Latent Space Reasoning
SST V2 introduces parallel-trainable nonlinear recurrence in latent space to let transformers reason continuously across positions, delivering +15 points on GPQA-Diamond and halving remaining GSM8K errors over matched baselines.
-
Do Not Imitate, Reinforce: Iterative Classification via Belief Refinement
RIC replaces single-pass label imitation with RL-driven iterative belief refinement, recovering cross-entropy optima while enabling adaptive halting via a value function.
-
Universal Transformers Need Memory: Depth-State Trade-offs in Adaptive Recursive Reasoning
Memory tokens are required for non-trivial performance in adaptive Universal Transformers on Sudoku-Extreme, with 8-32 tokens yielding stable 57% exact-match accuracy while trading off against ponder depth.
-
One Step Forward and K Steps Back: Better Reasoning with Denoising Recursion Models
Denoising Recursion Models train multi-step noise reversal in looped transformers and outperform the prior Tiny Recursion Model on ARC-AGI.
-
Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task
Transformers show limited adaptive depth use on relational reasoning, with clearer evidence after finetuning on the task.
-
ELT: Elastic Looped Transformers for Visual Generation
Weight-shared looped transformers trained with intra-loop self-distillation match MaskGIT-class FID/FVD at roughly 4x fewer parameters and support any-time inference across loop counts.
-
LPC-SM: Local Predictive Coding and Sparse Memory for Long-Context Language Modeling
LPC-SM is a hybrid architecture separating local attention, persistent memory, predictive correction, and control with ONT for memory writes, showing loss reductions on 158M-parameter models up to 4096-token contexts.
-
From Words to Amino Acids: Does the Curse of Depth Persist?
Protein language models exhibit consistent depth inefficiency where most task-relevant computation occurs in a subset of layers, mirroring patterns in large language models.
-
On the Spatiotemporal Dynamics of Generalization in Neural Networks
Deriving a neural cellular automaton from locality, symmetry, and stability postulates produces 100% accurate addition generalization from 16-digit to 1-million-digit inputs.
-
Dr.LLM: Dynamic Layer Routing in LLMs
Dr. LLM retrofits frozen LLMs with MCTS-supervised per-layer routers for skip/execute/repeat decisions, delivering up to +3.4% accuracy and 5-layer savings on reasoning tasks with strong out-of-domain generalization.
-
Mixture-of-Depths: Dynamically allocating compute in transformer-based language models
Mixture-of-Depths enables transformers to dynamically allocate compute by routing only the top-k tokens through each layer's full computations, matching baseline performance with a fraction of the FLOPs per forward pass and up to 50% faster sampling.
-
Scaling Data-Constrained Language Models
Repeating training data up to 4 epochs yields negligible loss increase versus unique data for fixed compute, and a new scaling law accounts for the decaying value of repeated tokens and excess parameters.
-
Solving math word problems with process- and outcome-based feedback
On GSM8K, outcome-based supervision achieves similar final-answer error rates to process-based with less labeling, but process-based or learned reward models are needed to reach 3.4% reasoning error among correct solutions.
-
Language Models (Mostly) Know What They Know
Language models show good calibration when asked to estimate the probability that their own answers are correct, with performance improving as models get larger.
-
Emergent Abilities of Large Language Models
Emergent abilities are capabilities present in large language models but absent in smaller ones and cannot be predicted by extrapolating smaller model performance.
-
A General Language Assistant as a Laboratory for Alignment
Ranked preference modeling outperforms imitation learning for language model alignment and scales more favorably with model size.
-
Scaling Laws for Transfer
Effective data transferred from pre-training to fine-tuning is described by a power law in model parameter count and fine-tuning dataset size, acting like a multiplier on the fine-tuning data.
-
R-Transformer: Recurrent Neural Network Enhanced Transformer
R-Transformer integrates RNNs with multi-head attention to model local and global sequence dependencies without position embeddings and reports large-margin gains over prior methods on diverse tasks.
-
Widening the Representation Bottleneck in Neural Machine Translation with Lexical Shortcuts
Gated lexical shortcut connections added to the transformer yield 0.9 BLEU average gains on five WMT directions while lowering the lexical content stored in hidden states.
-
Self Multi-Head Attention for Speaker Recognition
Self multi-head attention applied after CNN encoding of spectrograms outperforms temporal and statistical pooling for speaker verification on VoxCeleb1 with 18% relative EER reduction.