GTFormer dropped self-attention for a graph hierarchy and took a model trained on 512 tokens out to 4.2 million
AspireAI community · arXiv
Self-attention costs memory quadratically and, past the training horizon, its softmax spreads thin across tokens that do not matter. GTFormer replaces it with a bounded-degree graph over the raw token stream: a balanced binary reduction tree, horizontal shift edges that bind adjacent tokens, random expander shortcuts, convex highway gates and a top-down broadcast pass. The architecture is syntax-free, so no parser is needed.
Trained only on sequences of 512 tokens or fewer across eight synthetic long-context tasks, the best variant held 97.8 to 98.1 percent accuracy at lengths from 1,024 up to 4,194,304 tokens, which is 8192 times the training horizon. Memory stayed linear at 22.1 GB at the largest size and inference took 796.62 ms per sequence on a single GPU. Flat Transformer baselines fell from 57.9 to 59.5 percent at 64 tokens to 5.5 to 16.0 percent at 4,096 and then ran out of memory from 16,384 onward.
The honest caveat is that these are procedurally generated synthetic benchmarks with a compact vocabulary of 320 tokens, not real language or code, and the paper is a single-author study that states its own scope as compact embedding dimensions. The paper also reports its top figure as 98.1 plus or minus 0.2 percent in the abstract but as 98.1 plus or minus 3 percent in the results section, and notes the evaluation was stopped at that length because of memory limits. Treat the extrapolation claim as a structural result about routing, not as a drop-in replacement for a frontier transformer.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.