Transformers route predictions through tiny parameter subsets regardless of scale
Researchers have identified a striking bottleneck in transformer decision-making: while thousands of components technically influence each prediction, only dozens carry the signal that matters. Across eighteen models, the median prediction relies on just 53 components for 90 percent of its output, with as few as 8 sufficient to regenerate it entirely. Critically, this efficient subset spans only 1 to 3 percent of model parameters and does not scale with model size, suggesting transformers route information through surprisingly narrow pathways. This finding reshapes interpretability work and raises questions about redundancy, pruning potential, and whether current scaling approaches exploit model capacity efficiently.
Modelwire context
ExplainerThe finding that efficient signal pathways don't scale with model size directly contradicts the assumption that larger models simply learn redundant copies of the same computations. This suggests scaling may be adding capacity that isn't being used for the core prediction task.
This connects to the circuit robustness work from earlier today (Training Trajectories paper) which showed that learned mechanisms can be brittle artifacts dependent on training trajectory. Here we see a complementary angle: even if circuits are robust, transformers appear to route most signal through narrow channels regardless of total parameter count. The semantic bottleneck work on brain decoding also maps onto this pattern, showing that high-level representations compress information across domains. Together these suggest bottlenecks are a feature of intelligent systems, not a bug to engineer away.
If researchers can prune away the 97 percent of unused parameters without accuracy loss on held-out test sets (not just training data), that confirms these pathways are genuinely sufficient. If pruning succeeds on one model family but fails on another, watch whether the difference correlates with training procedure rather than architecture, which would validate the circuit brittleness finding.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsTransformers
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Through the Looking Glass: Directly Reading and Writing Transformers”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.