AI judges can be steered away from biased verdicts by manipulating their internal geometry, new research shows.
A team of researchers studied seven large language models used as automated evaluators across seven categories of scoring bias and nine benchmarks. Rather than probing bias at the input-output level — feeding in tweaked prompts and watching scores shift — they looked inside the models' hidden states. Biased inputs, they found, displace activations along a consistent low-dimensional subspace that becomes more pronounced in deeper layers. Applying "steering vectors" along that subspace could push a clean input toward biased scoring, or pull a biased input back toward a fair baseline.
The practical implication cuts past the usual prompt-engineering fixes. A simple linear projection onto the bias-direction features predicted judge failures on three entirely unseen benchmarks, outperforming text-based detection methods. That means a pipeline could flag likely bad judgments before they propagate — without any changes to the judge model itself.
LLM-as-judge setups have become load-bearing infrastructure in AI evaluation, powering leaderboards and RLHF pipelines alike. Most prior work on their biases — positional preference, verbosity preference, sycophancy — has stayed at the surface, proposing prompt patches that are often benchmark-specific. Mapping bias to activation geometry does not solve the underlying problem, but it suggests that the biases are structured enough to monitor and, in principle, correct at inference time.