💬 NLP

LLM Judges Have Dark Current: A Psychometric Datasheet for LLM-as-a-Judge Evaluation

This paper introduces a "Judge Datasheet" protocol that treats LLM-as-a-judge systems as measurement instruments rather than simple scoring devices, proposing a psychometric framework to quantify specific biases like "dark current" and positional preference to ensure reliable evaluation before making downstream claims.

Original authors: Hiroyasu Usami, Keisuke Hara, Ayato Tsuboi, Naohiko Matsuda

Published 2026-06-16✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hiroyasu Usami, Keisuke Hara, Ayato Tsuboi, Naohiko Matsuda

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a team of art critics to judge a painting contest. You want to know who is the best artist, so you ask these critics to compare two paintings and say which one is better.

This paper argues that we have been treating these "AI Critics" (LLM Judges) too simply. We usually just ask them, "Who won?" and report a single number, like "90% accuracy." The authors say this is like buying a thermometer without checking if it's broken, if it reacts to the wind, or if it gives a temperature reading even when there is no heat.

Here is the paper's core message, broken down with simple analogies:

1. The "Dark Current" Problem (The Phantom Signal)

In physics, "dark current" is when an electronic sensor gives a reading even when there is absolutely no light hitting it.

The Paper's Finding: The authors tested AI judges by giving them two identical answers (or even empty answers). A good judge should say, "These are the same, I can't pick a winner."
The Reality: Some judges (like the Llama-3.1-8B model) kept picking a winner anyway, even when the answers were identical. They were "hallucinating" a preference where none existed. This is their "Dark Current."

2. The "Position Bias" (The Seat Preference)

Imagine a judge who always picks the person sitting in a specific chair (e.g., always the left chair), no matter who is actually sitting there.

The Paper's Finding: The authors tested this by swapping the order of the answers. If the judge picks "Answer A" when it's first, but then picks "Answer B" (which is actually the same as A) when it's first, they aren't judging the content; they are just picking a seat.
The Reality: One of the judges (Llama-3.1-8B) was almost entirely driven by this "seat preference." It didn't care about the quality; it just wanted to pick the same position regardless of the content.

3. The "Datasheet" (The ID Card for Judges)

Just as you wouldn't buy a car without a spec sheet telling you its horsepower, fuel efficiency, and safety rating, the authors say we shouldn't use an AI judge without a "Judge Datasheet."

This datasheet measures five specific things:

Dark Current: Does it make up answers when there is no signal?
Stable Sensitivity: Does it give a consistent verdict on two answers of the same quality when their wording or presentation order changes? (It measures consistency on surface-form variation, not quality detection.)
Positional Bias: Does it cheat by picking the same slot/position regardless of the actual content after the order is swapped?
Target Sensitivity: Can it tell the difference between a "good" answer and a "great" answer? (This is the property that detects real/intended quality differences.)
The "Tie" Button: How strict is it about calling a tie?

4. The Three Judges (The Case Study)

The authors tested three different AI models to see what their "Datasheets" looked like:

Judge A (Llama-3.1-8B): This judge appears poorly calibrated for same-quality comparisons in this controlled setting. It has high "Dark Current" (it picks winners even when answers are identical) and is almost entirely driven by "Positional False Preference" (it picks the same slot regardless of content). It is less reliable for comparing similar-quality answers, though it might still be useful for pipeline debugging or for comparisons involving larger quality differences.
Judge B (Qwen2.5-14B): This judge is mixed. It doesn't have "Dark Current" (it stays quiet when there's no signal), and it is very good at spotting big differences in quality. However, when the answers are very similar (specifically same-quality pairs with surface-form variation), its behavior is mixed: sometimes it responds consistently to the surface-form variation, and sometimes it picks based on the order they were shown.
Judge C (Qwen2.5-32B): This is the cleanest judge. It has no "Dark Current," LOW positional false preference (not zero), and it is very good at spotting real quality differences. However, it is a bit "conservative"—it prefers to say "It's a tie" rather than guess when the difference is very small.

5. The "Strict Tie" Experiment

The authors tried a trick: they told the "cleanest" judge (Qwen2.5-32B), "Be stricter! Only pick a winner if you are 100% sure. Otherwise, call it a tie."

The Result: This successfully eliminated false preferences on the tested same-quality pairs (pairs with surface-form variation).
The Catch: It also made the judge miss some real but very small differences. It turned "I think this one is slightly better" into "I'm not sure, it's a tie."
The Lesson: You can change the judge's "strictness" (the criterion) by changing the instructions, but you cannot magically make the judge smarter or more sensitive just by asking nicely. Note: This experiment did not remeasure the "Dark Current" on identical/empty inputs under the strict prompt.

The Bottom Line

The paper does not claim that one of these judges is the "best" for all human tasks, nor does it prove any specific theory about how AI works.

Instead, it claims that before we trust an AI to judge other AIs, we must first measure the judge itself. We need to know if it has "Dark Current," if it's biased by position, and how strict it is. Without this "Datasheet," any score we get from an AI judge is just a number with no context, potentially hiding serious flaws.

Technical Summary: LLM Judges Have Dark Current

Problem Statement

The paper addresses a critical gap in the evaluation infrastructure of Large Language Models (LLMs). While "LLM-as-a-Judge" systems have become the practical default for comparing open-ended model outputs due to the cost and reproducibility issues of human annotation, these judges are currently treated merely as scoring devices. They are typically reported using scalar metrics like accuracy, win-rate, or agreement.

The authors argue that once a judge is used to validate another system, it functions as a measurement instrument. Like any physical instrument, it possesses intrinsic properties that must be characterized before its readings can be trusted. Specifically, judges may exhibit:

Dark Current: False preferences generated in the absence of any evaluative signal (e.g., on empty or identical inputs).
Positional Bias: Preferences driven by presentation order rather than content.
Cross-Sensitivity: Sensitivity to nuisance variations (surface form) rather than the target construct.
Criterion Drift: Variations in the operating point (tie vs. preference) induced by prompting.

The paper posits that reporting only scalar outcomes hides these distinct failure modes, potentially leading to invalid downstream scientific claims.

Methodology: The Judge Datasheet Protocol

To address this, the authors introduce a Judge Datasheet protocol, a metrological framework designed to measure the judge itself before it is used to measure other models. The protocol is grounded in psychophysics and Signal Detection Theory (SDT), utilizing a constructively controlled stimulus ladder.

Core Components

True-Vacuum Inputs (A0): The protocol tests the judge on inputs with no evaluative signal, including empty strings, whitespace, and identical non-empty pairs. This measures Dark Current (DC), defined as the rate of false preference when no signal exists.
Controlled Quality Ladder (A1): A prefix-chain checklist stimulus is constructed where higher-quality responses are strict supersets of lower-quality ones. This creates a Pareto-dominant ladder of quality differences ( $\Delta Q$ ) to measure Target Sensitivity.
Direction-Stability Decomposition ( $\Delta 0$ ): For same-quality pairs, the protocol evaluates both presentation orders (A, B) and (B, A). It decomposes raw false preference into:
- Stable Cross-Sensitivity (SCS): Content-stable preferences that persist regardless of order (indicating sensitivity to surface form).
- Positional False Preference (PFP): Preferences driven by the presentation slot (e.g., always choosing slot 1), which reverse the canonical content direction upon order reversal.
- One-Sided Commit (OSC): Cases where a preference is made in only one order.
- No-Preference: Valid ties or abstentions.
Criterion Shift Probing: The protocol tests how changing the prompt (e.g., enforcing a "strict tie" criterion) shifts the operating point without altering the underlying stimulus resolution.

Metrics

The protocol reports specific quantities including Dark Current, Raw $\Delta 0$ False Preference, Stable Cross-Sensitivity, Positional False Preference, Target Sensitivity ( $P_{correct}$ ), Detection Threshold ( $\Delta^*_{75}$ ), and Miss-by-Tie rates.

Key Contributions

The paper makes five specific contributions:

Judge Datasheet Protocol: A standardized measurement framework combining true-vacuum tests, controlled quality ladders, and criterion-shift probing.
Direction-Stability Decomposition: A method to separate raw same-quality false preference into stable cross-sensitivity, positional bias, and one-sided commits, revealing that high false preference does not necessarily imply stable content sensitivity.
Controlled Stimulus Ladder: A prefix-chain checklist design with Pareto dominance and specific controls for $\Delta 0$ same-subset and different-subset variants.
Three-Judge Case Study: An empirical evaluation of Llama-3.1-8B, Qwen2.5-14B, and Qwen2.5-32B, demonstrating distinct metrological profiles.
Criterion-Shift Evidence: Demonstration that prompting can move the tie/preference criterion (reducing false positives) but does not increase the resolution of the measurement (cannot detect signals below the instrument's inherent noise floor).

Results

The case study reveals that the three judges occupy fundamentally different metrological profiles, which would be obscured by scalar reporting:

Llama-3.1-8B: Exhibits high Dark Current (0.667) and a raw $\Delta 0$ false preference of 1.0. However, its Stable Cross-Sensitivity is negligible (0.033), while Positional False Preference is extremely high (0.967). It is characterized as "Presentation-conflicted," making it unsuitable for same-quality comparisons but potentially useful for pipeline debugging.
Qwen2.5-14B: Shows zero Dark Current and high target sensitivity. However, its raw $\Delta 0$ false preference (0.992) is a mix of Stable Cross-Sensitivity (0.45) and Positional False Preference (0.53). It is labeled "Mixed stable-positional."
Qwen2.5-32B: Demonstrates the cleanest profile with zero Dark Current, zero Stable Cross-Sensitivity, and low Positional False Preference (0.083). It is described as "Vacuum-clean" with low artifacts.
Criterion Shift: Applying a strict tie prompt to Qwen-32B eliminated its $\Delta 0$ false preference (reducing it to 0.0) but absorbed marginal $\Delta 1$ target signals into ties. Crucially, $\Delta 5$ sensitivity was preserved, and no "wrong-choice" errors were introduced. This confirms that prompting moves the criterion, not the resolution.

Claims and Significance

The authors maintain a modest scope regarding their claims:

No Downstream Validation: The paper does not confirm the "downstream mechanism hypothesis" (orientation in LLM evaluation) that motivated the work. It does not validate the orientation mechanism itself.
No Universal Trends: The authors do not claim a universal size-family trend or that Qwen-32B is a general-purpose evaluator. The results are specific to this controlled stimulus family.
Metrology First: The primary contribution is the Judge Datasheet protocol itself. The paper argues that before any downstream scientific claims are made using an LLM judge, the measuring device must be characterized for dark current, bias, and sensitivity.
Scalar Insufficiency: The results demonstrate that scalar win-rates or agreement scores are insufficient because they collapse distinct failure modes (e.g., positional bias vs. stable cross-sensitivity) into a single number.

In summary, the paper establishes a metrological baseline for LLM evaluation, asserting that "LLM judges are measurement instruments" that require multi-axis measurement to ensure the validity of the evidence they produce.