On: Kimi K2.7-Code: open-source coding model with better token efficiency

[regularfry]

The difference in outcome isn't that big but yes, you need to be more rigorous. For instance I've found that the Kimi K2.5 and K2.6 models will comment out failing tests rather than fix a problem they just caused (mistaking them for "pre-existing failures"), so you need to specifically make commented-out tests break the build. I've not personally had that problem with any of the Anthropic or OpenAI models.

[torginus → regularfry]

I wonder why it's the natural tendency of models to BS or do stuff like this when they don't have the correct answer - it's clear that they can program refusal into them, but for some reason, refusal has to be injected after the fact, and models can't really arrive at the conclusion that they can't answer properly.

[Eridrus → torginus]

I assume it's a lack of care when RLing them.

RL has a tendency to reinforce cheating when the cheats are easier to find than the final solution.

So when making your RL environment, you need to spend a lot of effort on finding ways the model can cheat and penalizing them.