By "evaluator" (aka "eval”), we did indeed mean frameworks for evaluating agent outputs broadly. The article and experiments center on LLM-as-a-judge, where an LLM is the grader, but the argument is ultimately statistical, so it holds regardless of whether the grader is an LLM, a simple supervised model, a set of regex checks, etc.
We were banking on readers being familiar with evals and left out definitions for conciseness, but as Gregaros points out, we could have been more explicit about what we meant.
as long as OpenAI and Anthropic keep subsidizing dirt cheap Codex or Claude Code usage, I'll just keep using them as evaluators. The trick is to have a fresh instance doing the reviewing, not the one that did the work.
> The trick is to have a fresh instance doing the reviewing, not the one that did the work.
In my experience that's not neccessary (some people even claim that you must use models from different vendors), and it's expensive since a fresh instance needs to rebuild all the context that's needed in order to properly and thoroughly review. LLMs have no problem throwing "them 5 minutes ago" under the bus when asked to review something "skeptically" and "with fresh eyes".
Doing it in the same session does save a ton of tokens but I find it's too biased towards its own implementation even if you tell it to use "fresh eyes" or to "act like a code reviewer in a bad mood." Including those strings in your prompt does show some improvement but not nearly as much as making it think from first principles in a fresh instance.
They should define this, but after having read the entire article I think it’s clear they mean “frameworks for evaluating the output of an agent” rather than what first might come to mind as “LLM evals”.
Their thesis is that even when the eval is useless for correctness of a single agentic action in production, it allows you to choose between two agents by cross-comparing in a large aggregated collection of tasks. Effectively: you can tune your agentic parameters.
Nothing new to the idea that taking many samples and averaging can work when a single datapoint doesn’t. Presumably this is part of a conversation in which we’re lacking context.
"LLM evals" is maybe an overused term because it can mean a bunch of things. This article talks about LLM-as-a-judge where an LLM scores another system's outputs.
By "evaluator" (aka "eval”), we did indeed mean frameworks for evaluating agent outputs broadly. The article and experiments center on LLM-as-a-judge, where an LLM is the grader, but the argument is ultimately statistical, so it holds regardless of whether the grader is an LLM, a simple supervised model, a set of regex checks, etc.
We were banking on readers being familiar with evals and left out definitions for conciseness, but as Gregaros points out, we could have been more explicit about what we meant.
reply