Give an AI judge the correct answer key — does that actually stop people from gaming it, or just move the cheat somewhere else?

This explores whether giving an AI judge a correct 'answer key' removes the weak spots people can exploit, or only moves them somewhere else. The corpus suggests it does some of both. Reference answers clearly help with one kind of error: judging whether the content is right. When judges were given reference answers along with explicit instructions for using them, their accuracy rose by 6.8%, which was enough for self-improvement training to match a purpose-trained reward model Can reference examples make LLM judges reliable enough for self-improvement?. This matches a broader pattern. Models can't reliably correct their own reasoning without an outside signal Can language models fix their own reasoning mistakes?, so an external anchor is often the thing that makes judging work at all.

The catch is that the most easily exploited judge biases have nothing to do with content. Judges score responses higher when they include fake citations or polished formatting, whatever the quality underneath. These 'authority' and 'beauty' biases work without any access to the model llms-are-susceptible-to-four-exploitable-biases-that-enable-zero-shot-prom Can LLM judges be tricked without accessing their internals?. An answer key tells the judge what is correct. It does not stop the judge from being impressed by how a response looks. Framing cues work the same way: AI judges forgave a rule-breaking piece of writing when told a human wrote it, while human judges became stricter Do authorship labels change how AI judges evaluate rule violations?. None of these surface or context effects are things a reference answer would obviously neutralize.

References can also create new places for errors to hide. Methods that check only the final answer let through reasoning that is wrong but lands on the right result. STaR improves models by keeping only rationales that reach correct answers Can models improve by filtering only on answer correctness?, and VeriFree drops the judge entirely, rewarding reasoning by how likely it makes the reference answer Can reasoning improvement work without answer verification?. Both work well, but both put all their trust in the reference and the final output. In long agent tasks, most failures turned out to be process violations rather than wrong final answers, and checking intermediate steps raised success from 32% to 87% Where do reasoning agents actually fail during long traces?. A reference answer only checks the endpoint, so errors in the steps before it go unchecked.

The corpus offers two complementary responses. The first is to change how the judge evaluates: training judges with reinforcement learning to reason through their verdicts measurably reduces authority, verbosity, position and formatting bias Can reasoning during evaluation reduce judgment bias in LLM judges?. The second is to stop expecting the judge to be unbiased and contain its errors with structure instead. Telling a judge not to be biased doesn't work reliably Can prompting reduce bias in LLM judges reliably?. What helps are mechanical guardrails: run clear-cut checks before debatable ones, measure the judge against human labels, keep test data hidden from whatever is being judged, and plant known cases as alarms Can deterministic checks protect LLM judges from failure?.

The takeaway you might not have expected: a reference answer is less a fix than a decision about where to place your trust. It reduces errors about whether content is correct. It leaves the judge's taste for polish and authority untouched, and it moves the risk onto the quality of the reference itself and onto everything the final-answer check doesn't see. The corpus doesn't directly measure whether references weaken the formatting and authority biases, so that part remains an open question rather than a settled result.