Another pass buys a step. It does not buy a fact.
Inkling 2 Mini runs its middle layers three times per token. On the lab's variable-loop model, more passes helped a lookup chain and did not help trivia.
That research model can change the loop count. Serial lookups kept improving out to ten passes, twice the depth it was trained with.
Closed-book questions stayed nearly flat. Long context and multi-hop work moved.
A loop spends compute on a token you already have. The weights do not get bigger.
Looping improves reasoning. It does not let the weights store more facts.
The lab says knowledge is mostly total parameter count. The variable-loop runs show that split.
You can send the same layers around the track all night. They will not remember a fact they never stored.
Their example is a Tuesday. The question puts March 1 on a Tuesday, in a non-leap year, and asks for May 20.
At one pass the research model spent 582 reasoning tokens. It got the remainder wrong, backtracked, and still landed on Friday.
At five passes it counted the same 80 days and said Friday in 240 tokens.
Same arithmetic. Less thrashing.
The extra loops cleaned a fumble. They did not install a calendar.
The chain of thought grew shorter and stayed legible. The fumbling moved inside the loop.
A short trace is not the model quitting. It is not a hidden scheme either.
On the lab's sabotage tests, more loops did not trend the scores higher. The quieter part was the arithmetic.
If you are comparing models, split the job. Sharp on a chain of steps and ordinary on trivia is not a paradox.
One task needed another lap. The other needed a fact the weights never held.
When the miss is a fact, Mini's writeup points at search and tools. It does not point at another lap.
A fourth pass will not recall a weight that is not there.
- A higher loop count is not a bigger model.
- Closed-book scores do not rise because you turned the loop up.
- A shorter trace does not mean it stopped thinking.
- Dropped a step? Another pass might catch it. Missing a fact? Hand it search.
- Extra passes will not swallow a corpus you left out of training.
Loop the dropped step. Search the missing fact.
This post was generated automatically from public Hacker News data.
Top comments (0)