If an AI grades its own homework, it's generous — does handing grading to a separate, skeptical AI actually fix that?

This looks at why an AI agent grading its own work is too generous, and why giving the grading job to a separate agent works better than trying to make one agent harder on itself. Anthropic's own testing gives the plainest answer. When one agent judged its own output, it praised itself too much, especially on subjective tasks where nothing could be checked automatically. Moving the judging to a dedicated, skeptical evaluator agent was much easier to tune than making the generator self-critical. Because the evaluator kept finding problems, improvement could continue past the point where self-evaluation had stopped finding any Can separating generator and evaluator agents reduce self-praise bias?. The key word is tractable. The split doesn't magically produce better judgment. It turns a hard problem, fighting a model's pull toward its own answers, into an easier one: tuning one component for one job.

The pull itself is real and has a mechanism. Models over-trust answers they generated, partly because an answer the model found highly probable while writing also looks correct when it reads it back. Generation and judgment share the same internal sense of what seems right Why do models trust their own generated answers?. A second finding goes further. When researchers fine-tuned models to get better at recognizing their own summaries, the models' preference for those summaries rose in step. That points to self-recognition as a driver of self-favoritism, though the authors call it initial evidence, not proof Do LLMs favor their own text because they recognize it?. This suggests why separation helps: an evaluator that doesn't experience the text as its own has less reason to go easy on it.

Separation is only the first step. A separate judge can still be shallow or biased. One line of work replaces a single LLM judge with an agent that actively gathers evidence before ruling. It cut the judge's error from 31% to 0.27% on complex tasks. But its memory module passed errors from one step to the next, so even a well-separated evaluator needs its own internal safeguards against mistakes Can agents evaluate AI outputs more reliably than language models?. Another approach builds several evaluator personas from real stakeholder documents and has them debate, so the evaluation comes from several viewpoints instead of one Can personas extracted from documents generalize across evaluation tasks?. Reward models have taken a similar path: having the judge reason through its verdict before scoring makes it stronger Can reward models benefit from reasoning before scoring?.

There's a twist the question doesn't anticipate. A fixed, separate evaluator eventually becomes something to exploit. As the generator improves, static criteria saturate and invite gaming Why do fixed benchmarks fail as agents grow stronger?. One response is to have the evaluator improve alongside the agent it scores. That lets self-improvement work even on creative tasks like writing, where no fixed answer key exists, at a lower token cost Can evaluators improve alongside the agents they score?. A related design idea is to split evaluation itself into independent parts (benchmark, harness, environment), so that reward-hacking shows up in the agent's step-by-step behavior instead of hiding inside one final score How can we make reward-hacking visible in agent evaluation?.

The corpus also points the other way: some work tries to put self-evaluation back inside the model. Post-Completion Learning trains a model to score its own output in the unused space after it finishes answering, at no extra cost at inference time Can models learn to evaluate their own work during training?. So the field hasn't concluded that self-evaluation can't work. The more accurate conclusion is that untrained self-evaluation is lenient by default. A separate evaluator is the cheapest and most reversible fix today, and it fits a broader pattern of agents improving through changes to their scaffolding rather than to their weights Do self-improving agents really split into two distinct loops?.