Decoding Transformer's Take on Magnitude

How do transformer models comprehend magnitude? It's a question that has puzzled researchers, with theories ranging from logarithmic spacing to linear encoding. Yet recent work suggests something different entirely. Let me break this down.

Logarithmic Geometry Prevails

Applying psychophysics tools, researchers assessed three instruction-tuned models, Llama, Mistral, and Qwen, each featuring 7-9 billion parameters. Across domains like numbers, time, and space, they consistently found log-compressive geometry. Representational similarity analysis (RSA) correlations with a Weber-law dissimilarity matrix showed high alignment, ranging from 0.68 to 0.96 across 96 model-domain-layer cells. Here's what the benchmarks actually show: linear geometry never took the lead.

Behavior vs. Representation

Interestingly, the numbers tell a different story behavior. One model exhibited a human-like Weber fraction (WF = 0.20), yet both models performed no better than chance in temporal and spatial discrimination tasks. Strip away the marketing and you see that possessing a logarithmic geometry doesn't equate to behavioral competence.

The Layer Conundrum

Another intriguing finding comes from causal interventions. Early layers seemed important in magnitude processing, with 4.1 times the specificity, yet later layers, where geometry was strongest, weren't as engaged (showing only 1.2 times the specificity). So, is it all about the architecture? The architecture matters more than the parameter count, it seems.

Corpus analysis also sheds light: efficient coding with an alpha of 0.77 was confirmed. This indicates that the statistics of training data alone can lead to log-compressive magnitude geometry. But should we be surprised that geometry alone doesn't guarantee success?

Why Should We Care?

These findings challenge assumptions about how transformers grasp magnitude. If geometry isn't enough for behavioral competence, what's missing? Does this mean we need new ways of training or model design? Frankly, it's a reminder that there's more to AI than just crunching numbers. As we push the frontiers of what these models can do, understanding their capabilities, and limitations, becomes increasingly important.