Hacker News Reader

Comment by jronak | Hacker News Reader

Did you look at the ARC AGI 2? Codex might be overfit for terminal bench

ARC AGI 2 has a training set that model providers can choose to train on, so really wouldn't recommend using it as a general measure of coding ability.

More fundamentally, ARC is for abstract reasoning. Moving blocks around on a grid. While in theory there is some overlap with SWE tasks, what I really care about is competence on the specific task I will ask it to do. That requires a lot of domain knowledge.

As an analogy, Terence Tao may be one of the smartest people alive now, but IQ alone isn’t enough to do a job with no domain-specific training.

A key aspect of ARC AGI is to remain highly resistant to training on test problems which is essential for ARC AGI's purpose of evaluating fluid intelligence and adaptability in solving novel problems. They do release public test sets but hold back private sets. The whole idea is being a test where training on public test sets doesn't materially help.

The only valid ARC AGI results are from tests done by the ARC AGI non-profit using an unreleased private set. I believe lab-conducted ARC AGI tests must be on public sets and taken on a 'scout's honor' basis that the lab self-administered the test correctly, didn't cheat or accidentally have public ARC AGI test data slip into their training data. IIRC, some time ago there was an issue when OpenAI published ARC AGI 1 test results on a new model's release which the ARC AGI non-profit was unable to replicate on a private set some weeks later (to be fair, I don't know if these issues were resolved). Edit to Add: Summary of what happened: https://grok.com/share/c2hhcmQtMw_66c34055-740f-43a3-a63c-4b...

I have no expertise to verify how training-resistant ARC AGI is in practice but I've read a couple of their papers and was impressed by how deeply they're thinking through these challenges. They're clearly trying to be a unique test which evaluates aspects of 'human-like' intelligence other tests don't. It's also not a specific coding test and I don't know how directly ARC AGI scores map to coding ability.

> The only valid ARC AGI results are from tests done by the ARC AGI non-profit using an unreleased private set. I believe lab-conducted ARC AGI tests must be on public sets and taken on a 'scout's honor' basis that the lab self-administered the test correctly

Not very accurate. For each of ARC-AGI-1 and ARC-AGI-2 there is training set and three eval sets: public, semi-private, and private. The ARC foundation runs frontier LLMs on the semi-private set, and the labs give them pre-release API access so they can report release-day evals. They mostly don't allow anyone else to access the semi-private set (except for live Kaggle leaderboards which use it), so you see independent researchers report on the public eval set instead, often very dubious. The private is for Kaggle competitions only, no frontier LLMs evals are possible.

(ARC-AGI-1 results are now largely useless because most of its eval tasks became the ARC-2 training set. However some labs have said they don't train LLMs on the training sets anyway.)

Reply on news.ycombinator.com