The news. On July 13, 2026, researchers released Inside the Unfair Judge, a mechanistic-interpretability study of why LLM-as-judge scoring goes wrong. Across 7 judges, 7 bias types, and 9 benchmarks, biased inputs push the judge's hidden state along low-dimensional, type-specific directions that sharpen with depth. Projecting an answer onto those directions predicts judge failures on 3 entirely unseen benchmarks, beating text-based features. And steering the activations along a bias direction moves the score both ways, with an effect an order of magnitude larger than a random direction of the same length. Read the paper →
Picture a judging panel with one crooked referee. Every match they score tilts the same way, because they keep a thumb pressed on one side of the scale. Hand them a different contestant or reword the rules, and the referee still leans the same way. An LLM grading another model's answers can carry that same kind of lean — and rather than only rewording the prompt, this paper goes looking for the lean inside the model. The surprising part is where the thumb lives. Alongside the words of the question, it shows up as a direction inside the model's own vector space of hidden states — the numbers the judge computes before it ever writes a score.
For years, people studied judge bias from the outside: reword the prompt, flip the option order, and watch the score wobble. This paper looks inside too, and finds that the bias is not scattered noise but a low-dimensional pattern — each kind of bias occupies its own compact subspace, a handful of directions you can point at in the judge's activations. Feed the judge a neutral answer and a biased one, compare their hidden states, and the difference lines up along that bias direction. The signal even gets sharper the deeper you look, as the hidden state flows down the network's stacked layers. Because meaning in these models is already a matter of direction in a high-dimensional space, a bias can be a compact bundle of directions rather than a thousand tangled quirks.
Finding a direction is one thing; proving it causes the bias is another. So the authors do something a prompt study can't: they grab the judge's activations and push them along the bias direction, like leaning on the referee's thumb from the inside. Push forward, and a clean answer suddenly earns the biased score. Push the same amount in reverse, and a biased score slides back toward its baseline — the score the judge gives when the bias isn't in play. The direction behaves like a steering wheel for the verdict — turnable both ways.
Here is where it earns its keep. Picture one answer the judge is scoring, and change just two things: the direction you nudge its hidden state, and keep the step length fixed. Nudge one step along the bias direction and the score jumps toward the biased verdict; take a step of the exact same length in a random direction and the score barely twitches — the paper reports the bias-direction effect is roughly an order of magnitude (~10×) larger than a matched-length random nudge (the paper reports this ratio, not a single per-step score). That gap is the whole point: if a random direction did just as much, the bias would be noise you could never isolate; because one specific direction does ~10× more, the bias is a real, findable, steerable thing. And reading that direction pays off downstream — projecting an answer onto it predicts the judge's mistakes on 3 unseen benchmarks better than the text does, turning "our grader is a little biased" into a number you can watch for drift and steer back toward baseline.
| Approach | What it inspects | Predicts failures on unseen benchmarks? | Acting on the bias |
|---|---|---|---|
| Prompt / input-level probing (prior work) | the wording you feed in | Text-based features — the paper reports these predict less well (paper) | Prompt-level mitigations (rephrase, reorder) |
| Activation geometry (this paper) | the judge's hidden state | Yes — beats text features, on 3 unseen benchmarks (paper) | Reverse-steer to restore baseline scoring |
Goes deeper in: LLM Internals → Embeddings → Tokens in Space
Related explainers
- kNNGuard turns LLM hidden activations into a training-free guardrail — another intervention that reads the hidden state, rather than the prompt, to catch bad behavior with no retraining.
- LLM-as-a-Verifier scales agent feedback with logit-score expectations — the flip side: making an LLM's own scores into a reliable feedback signal, when the judge behaves.
- Agent leaderboards mislead under distribution shift — why an eval you trust can still fail to transfer; a biased judge is one reason a leaderboard doesn't hold up.