When LLM judges agree, should we believe them?

Imagine evaluating a retrieval-augmented-generation system. A personification asks a question, the strategy retrieves a matter passage, and an LLM judge decides whether it’s relevant. To trim noise, you inquire respective judge models to measure the aforesaid passage. Eight opportunity “relevant”; 2 opportunity “not relevant”.

Eight retired of 10 feels convincing. But the important mobility is not only really galore judges agreed but really independently they arrived astatine that agreement.

If the 8 agreeing judges are genuinely different sources of evidence, past statement is simply a beardown signal. But if they stock a punctual template, a training lineage, a exemplary family, aliases a communal unsighted spot, they whitethorn beryllium repeating the aforesaid mistake. The ballot count makes the grounds look stronger than it really is.

Our insubstantial “Dependence-aware explanation aggregation for LLM-as-a-judge via Ising models,” coauthored pinch Shiva Kasiviswanathan and presented astatine this year’s International Conference connected Machine Learning (ICML), addresses this problem. We coming a method for assessing the correlations betwixt judges’ outputs and adjusting the aggregate people accordingly, to guarantee a diverseness of opinion.

In tests connected 3 different tasks, our method outperformed the best-performing baseline — a sheet of judges weighted according to humanities accuracy — by 9% to 14% connected modular metrics.

Hidden assumptions

The attraction of mostly ballot is its simplicity. Every judge gets 1 vote, and the reply pinch much votes wins. Weighted mostly ballot is simply a earthy improvement: judges that look much meticulous get much influence.

Both approaches are useful baselines. But they are built astir the aforesaid simplified position of the judge panel: judges that get the incorrect reply are treated arsenic though they make their errors independently.

That presumption is often excessively optimistic for LLM-as-a-judge systems. Two judges whitethorn neglect together because they construe the rubric similarly. Several judges whitethorn beryllium prompted pinch the aforesaid examples and truthful inherit the aforesaid information bias. A group of related models whitethorn beryllium delicate to the aforesaid phrasing. In these cases, a mostly tin beryllium little informative than it appears.

A judge sheet is simply a network

A amended aggregator would dainty the sheet arsenic a web of judges. Each judge still has its ain reliability profile, but pairs of judges tin besides person relationships. Some pairs work together much often than their individual reliability profiles would predict, including connected shared mistakes. Other pairs supply much complementary perspectives.

We exemplary these relationships pinch an Ising model, a statistical exemplary that tin correspond pairwise dependence betwixt binary variables. In the LLM-as-a-judge context, the aggregator learns some judge accomplishment and judge similarity.

Our method is designed for the unsupervised setting: it learns from judge outputs without utilizing quality reference labels for training. It treats each item's existent explanation arsenic a latent adaptable to infer jointly pinch the parameters describing judge reliability and dependence.

There are 2 useful levels of dependence modeling. In the first, the narration shape among judges is treated arsenic astir the aforesaid for affirmative and antagonistic labels. The last determination still looks for illustration a weighted vote, but the weights are adjusted for correlation. Redundant statement tin beryllium discounted without making the prediction norm difficult to interpret.

The 2nd version — the class-dependent exemplary — lets the narration shape alteration pinch the label. This is useful erstwhile the statement building carries people accusation — for example, erstwhile judges show wide statement connected clear-cut items but divided into recognizable clusters connected ambiguous ones. This attack is much expressive, but it requires much information to estimate the other parameters reliably.

Learning from information logs

Starting from an first parameter setting, the algorithm combines each item's votes to estimate the probability that its existent explanation is positive. These soft probabilities are the model's existent champion guesses, not outer labels. It past alternates betwixt updating those probabilities and re-estimating judge reliability and pairwise dependence from them. Reference labels are utilized only afterward to measurement experimental accuracy.

This attack is particularly applicable for teams that already cod LLM-as-a-judge outputs astatine scale. Existing information logs incorporate much than conscionable votes; they incorporate patterns of statement and disagreement. Dependence-aware aggregation turns those patterns into a usable signal.

The aforesaid learned web tin thief reply applicable questions. Are akin models adding independent evidence, aliases are they mostly reinforcing each other? Does 1 task nutrient wide agreement, while different produces cluster-specific splits? Is adding different judge apt to amended the information aliases simply copy an existing root of bias?

Evaluation

We evaluated our attack connected 3 binary tasks: relevance classification for retrieved information, toxicity classification, and summarization assessment. The judge sheet contained 10 judge models, each tally astatine somesthesia zero — meaning there’s nary randomness successful their outputs, truthful the aforesaid input will ever elicit the aforesaid output.

We compared the dependence-aware models pinch 2 conditional-independence baselines: weighted mostly ballot and azygous mostly vote. Across the 3 tasks, modeling dependence improved accuracy erstwhile the strategy had capable information items and capable judges to estimate meaningful relationships.

Using each 10 judge models and the maximum disposable training information for each task, the strongest dependence-aware results were 0.912 accuracy connected relevance, compared pinch 0.820 for weighted mostly ballot and 0.804 for azygous mostly vote; 0.792 connected toxicity, compared pinch 0.694 and 0.695; and 0.806 connected summarization, compared pinch 0.737 and 0.561.

Best practices

For teams utilizing LLM-as-a-judge pipelines, dependence-aware aggregation suggests a fewer useful habits.

• First, measure the judge panel, not conscionable the individual judges. A group of individually beardown judges tin still beryllium redundant if they neglect successful the aforesaid way.
• Second, dainty exemplary diverseness arsenic statistical diversity. Mixing exemplary families aliases architectures is adjuvant only to the grade that it changes the correction patterns that matter for the task.
• Third, inspect statement structure. Strong clusters tin uncover shared rubrics, shared exemplary behavior, aliases task-specific ambiguity. That accusation is valuable moreover erstwhile the last explanation is unchanged.
• Finally, study uncertainty pinch dependence successful mind. Ten correlated votes should not ever nutrient the aforesaid assurance arsenic 10 independent votes.

When LLM judges agree, we should inquire why. Sometimes statement is independent evidence. Sometimes it is simply a shared unsighted spot. A bully aggregation method should beryllium capable to show the difference.

Acknowledgments: Shiva Prasad Kasiviswanathan