Post

mr. TIM
@timkellogg.me

Inverse scaling of reasoning models a research collab demonstrated that there are certain types of tasks where all top reasoning models do WORSE the longer they think things like getting distracted by irrelevant info, spurious correlations, etc. www.arxiv.org/abs/2507.14417

Rémy
@xowap.dev

That's very similar to what I can the "breakdown threshold": Especially with agents, there is a limit number of bullet points/instructions that you can feed to the LLM. Add just one more and everything falls appart (tool calls become unreliable AF, constraints are half-followed, etc)

Rémy
@xowap.dev

The way I've found to get past this is to feed all the instructions (prompt, tools, etc) into a LLM and just ask to re-write all the prompts and documentation without useless or redundant instructions. Drops the prompt sizes massively (60~80% in my case) and makes a HUGE performance boost

0 reposts1 like

Rémy
@xowap.dev

Yup, it's obvious by now that the job of the agent is to manage the big picture, business rules and context curation

mr. TIM
@timkellogg.me

they didn’t mention it, but laziness giving a single agent a chain of 6 tasks vs doing them in parallel an LLM won’t ever be lazy about a single task, so doing it in parallel ensures they all get done