C.W.K.
Stream
Lesson 03 of 05 · published

The TC⁰ Breakthrough

~12 min · complexity, expressivity, rwkv-7

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

A formal expressivity result, not just a benchmark win

Standard Transformers, by a 2023 result, are bounded by the TC⁰ complexity class — they can only solve problems that can be expressed by constant-depth threshold circuits with polynomial size. There are languages and computational problems strictly outside TC⁰, and a vanilla Transformer provably cannot solve them, no matter how big you make it.

RWKV-7's headline formal result: RWKV-7 provably exceeds TC⁰. Specifically, it can solve the S5 state-tracking problem using only 2 layers (NC¹ complexity — circuits with logarithmic depth, strictly more expressive than TC⁰), and it can recognize all regular languages using 4 layers. These are concrete computational problems that Transformers structurally cannot do at any size.

Where this shows up in practice

The TC⁰ exit isn't a benchmark trick; it predicts a real capability difference. RWKV-7 maintains richer sequential state — counts, parity, regular-language structure — that Transformers struggle with. RWKV-7 2.9B achieves 72.8 average across standard benchmarks, matching Qwen2.5-3B (71.4) and beating Llama 3.2 3B (69.7), despite being trained on only 5.6T tokens compared to 15–18T for those competitors.

That training-token-efficiency story matters. If RWKV is hitting Transformer-comparable benchmarks with one-third the training data, the architecture is doing something the Transformer isn't — likely the richer state dynamics letting the model learn faster from each example. This is one of the clearest "RWKV is not just a Transformer with different math" data points.

The honest caveat

TC⁰ exit is for state-tracking problems. It is not a general expressivity advantage. Transformers excel at parallel associative lookup over the full context — that's a different axis, and on it RWKV (like all recurrent-state architectures) gives ground. The lesson here isn't "RWKV is better than Transformer" — it's "different architectures sit in different complexity classes, and the choice matters more than people credit".

External links

Exercise

Take a small computational task that's TC⁰-hard (e.g., recognizing the language a^n b^n with n up to a few hundred, or counting the number of nested parentheses). Generate synthetic training data, train a small Transformer (≤10M params) and a small RWKV-7 (similar size) on it, and observe the gap. The Transformer will plateau; the RWKV-7 should generalize cleanly. This is the TC⁰ exit, in your own training run.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.