A small edit can look like a new skill. Measure the failure you removed before you call the score learning.

That is the part worth stealing from a Hacker News thread about models that read and write neural network weights directly. Researchers at MIT CSAIL asked GPT-6 Astra to make Qwen3.5-0.8B better at math by editing weights. No gradients. No training loop. Astra traced a reasoning loop to one direction written by the MLPs in layers 15 and 19, then subtracted that direction from 64 output columns in each of those layers. The patch is 131,072 weights. On a model they describe as having 0.8 billion weights, that is 0.016%.

The scoreboard is easy to screenshot and easy to strip of context. MATH-500 went from 7.9% to 48.1%. GSM8K went from 13.8% to 67.1%. IFEval, which measures instruction following, went from 21.0% to 62.3%. Qwen's own model card already warns that this small model is prone to thinking loops. On the unedited model, 85 of the first 96 practice answers ran into the 32,768-token limit. After the edit that stuck, the share of MATH-500 answers hitting that limit fell from 93% to 33%.

The attempts that missed are the lesson. With 10 forward passes, MATH-500 only reached 12.3%, and 87.2% of answers still died at the cap. With 25 passes, 14.1% and 86.4%. The score did not climb while the model learned math. It sat still until the loop broke, and then it jumped. The loop is not special to math, which is why a non-math test moved too. A model that never lands the last line fails every exam that grades the last line.

The authors then ran the control that separates a blocker from a skill. Gemma 3 270M rarely loops. They report that only 8.5% of its answers reach its token limit, against 93% for that Qwen. Same style of direct weight edit. MATH-500 started at 6.8%, and no edited run moved it by more than about a point. The best figure in their table is 7.7%. On their account, Astra found ordinary wrong arithmetic and invalid algebra, with no single component to switch off. Fixing a loop lets the model finish answers it could already work out. Removing an obstacle that hid an ability does not install an ability the model lacked.

The same shape shows up in systems that have no weights at all.

Suppose an API's success rate climbs after you delete one retry bug: on timeout, call yourself until the gateway hangs up. Callers start getting a response. The handler did not get wiser. It stopped spending the whole deadline on a polite panic. Requests that were wrong for a real reason are still wrong. Success rate cannot tell those two stories apart. A math benchmark cannot either, once you crop the column that says the answer never finished.

Or take a function that returns early on a weird input, so the suite is a pile of timeouts. You delete the early return. The tests go green. The changelog says the score jumped. What jumped was termination. The missing branch is still missing. It is only visible now. That is worth shipping, and it is still not a new algorithm.

Do not publish the headline number and leave the blocker in a footnote you drop from the slide.

MATH-500 and "hit the token limit" are one measurement. Ten and twenty-five passes barely moved either one. Fifty passes moved both. If the score rises while the loop rate stays put, you might have taught something. If they move together, you unblocked a finish. Deleting while (true) is not a math curriculum. It is a favor to whoever pays for the electricity.

Before you trust a delta tomorrow, write the suspected blocker in one sentence, in the language of the system you have.

"It repeats until the sampler stops." "It retries until the gateway kills the request." "The test times out because we return before the hard case." Log that rate beside the score, on the same examples, before and after. Then find a control where the blocker is already rare. If the control sits still, you removed an obstacle. Put the remaining errors on a real plan: data, training, or code that implements the case you do not have yet.

Ship the unblocker. A model that can do the step, and then repeats until the token limit cuts it off, should have that habit removed. The patch was small because the habit was narrow. Do not send the next reader looking for a math lesson inside 131,072 weights that were only holding the door shut. The answers it could already work out were waiting behind that door.

Source: https://rikiyat.github.io/blog/weight-space-programming/

Hacker News: https://news.ycombinator.com/item?id=50341174

This post was generated automatically from public Hacker News data.