Variable-Width Transformers: X-shaped architecture outperforms uniform-width baselines with 22% fewer FLOPs
Researchers propose the ><former (X-shaped transformer), a decoder-only architecture that uses wider early and late layers with narrower middle layers, implemented via a parameter-free residual resizing mechanism. Evaluated on models from 200M to 2B dense parameters and 3B MoE, the architecture consistently outperforms parameter-matched uniform-width baselines on language modeling loss. The design yields a 22% reduction in FLOPs and 15% reduction in KV cache memory under fitted scaling curves, suggesting nonuniform width allocation is a viable path to more compute-efficient language models.
Related guides (2)
Related events (8)
Dynamic short convolutions yield 1.33–1.60× compute advantage over standard Transformers
A new arXiv preprint introduces dynamic short convolutions as an architectural primitive for Transformers, using input-dependent filters to combine locality bias with increased expressivity. Experiments across 150M–2B parameter language models show consistent perplexity improvements over standard Transformers and static convolution variants, with scaling-law fits indicating a 1.33× compute advantage when applied to key/query/value vectors and 1.60× when added after every linear layer. The technique also improves linear RNNs (Mamba-2, Gated DeltaNet) and mixture-of-experts architectures, with custom Triton kernels making training practical.
Hyperfitting Explained: Terminal Geometric Expansion in Final Transformer Layers Drives Diversity Gains
This paper investigates the 'hyperfitting' phenomenon—where fine-tuning LLMs to near-zero loss on small datasets improves open-ended generation and reduces repetition—and demonstrates it is mechanistically distinct from temperature scaling. Entropy-matched control experiments falsify both the temperature-equivalence and static vocabulary reweighting hypotheses, instead localizing the effect to a 'Terminal Expansion' in the final transformer block where feature-space dimensionality expands by ~80.8 dimensions, enabling promotion of deep-tail tokens via context-dependent rank reordering. The authors introduce Late-Stage LoRA, a targeted fine-tuning strategy updating only the final 5 layers, achieving robust generation with minimal parameter updates.
Looped World Models introduce iterative latent depth as a new scaling axis for world simulation
A new arXiv preprint introduces Looped World Models (LoopWM), a parameter-shared transformer architecture that iteratively refines latent environment states to achieve up to 100x parameter efficiency over conventional world models. The approach uses adaptive computation to scale depth dynamically per prediction step, addressing the tension between long-horizon simulation fidelity and deployment cost. The authors position iterative latent depth as a new scaling axis orthogonal to model size and training data.
OrpQuant: Geometric Orthogonal Residual Projection for Multiplier-Free Power-of-Two Transformer Quantization
This paper introduces Orthogonal Residual Projection (ORP), an algorithm-hardware co-design framework for ultra-low-bit quantization of LLMs and Vision Transformers targeting edge deployment. ORP addresses the structural limitations of Power-of-Two (PoT) quantization by formulating quantization as a dual-basis geometric projection that synthesizes higher-resolution residual lattices using only shift-and-add operations, eliminating multipliers. At 3-bit (W3/A16), ORP achieves 6.10 perplexity on LLaMA-2-7B, competitive with MAC-intensive baselines like AWQ, while reducing full-model calibration time to ~15 minutes. RTL synthesis at 28nm confirms hardware efficiency by mitigating timing bottlenecks from dense multiplier trees.
Looped Diffusion Language Models (LoopMDM): Depth Scaling via Layer Looping
LoopMDM introduces selective looping of early-middle transformer layers in masked diffusion language models, achieving a depth-scaling effect without adding parameters. The approach matches same-size MDM performance with up to 3.3× fewer training FLOPs and outperforms deeper non-looped MDMs on reasoning benchmarks, including up to 8.5 points improvement on GSM8K. Inference-time compute scaling is enabled by varying loop counts, with adaptive loop scheduling providing additional efficiency gains. Attention analysis suggests looping works by promoting interactions among masked token positions.
Predictor-gated bank-wise sparsity recipe for dense-to-sparse LLM upcycling from Qwen2.5-8B
A new arXiv preprint introduces a continual training recipe to convert dense LLMs into channel-sparse models without post-hoc pruning. Starting from a Qwen2.5-8B checkpoint, the method uses a low-rank predictor to gate FFN channel routing, achieving 4x sparsity in FFN intermediate activations via a bank-wise top-k rule at 32K context. The routing module is trained on the main language modeling path, making the resulting sparsity hardware-oriented rather than approximate. The authors also identify and patch a layer-local long-context failure mode on the RULER-CWE benchmark.
Training-Free Looped Transformers: Inference-Time Recurrence via ODE-Motivated Layer Reapplication
The paper introduces a method to retrofit recurrence onto frozen pretrained transformer checkpoints at inference time by looping a contiguous mid-stack block of layers without any fine-tuning or architectural changes. Naive block reapplication degrades performance, so the authors motivate their approach by treating pre-norm transformer blocks as forward Euler ODE steps and replacing one large update with smaller damped sub-steps. Evaluated across seven model families including dense, sparse MoE, and MLA+MoE architectures, the method yields consistent benchmark improvements (e.g., +2.64 pp on MMLU-Pro for Qwen3-4B-Instruct) at no training cost.
Ω-QVLA: Training-Free W4A4 Quantization for Full Vision-Language-Action Models Including Diffusion Action Heads
Omega-QVLA is a post-training quantization framework that compresses both the LLM backbone and the diffusion-based action head of VLA models to uniform W4A4 precision without mixed-precision schemes or fine-tuning. It combines composite SVD-Hadamard rotation for weight energy equalization with per-step DiT activation scaling to handle dynamic-range drift across denoising steps. On the LIBERO benchmark, it achieves 98.0% and 87.8% task success on Pi 0.5 and GR00T N1.5 respectively—matching or exceeding FP16 baselines—while reducing static memory footprint by 71.3%. Real-world manipulation experiments confirm the approach generalizes beyond simulation.