Probing verbosity bias and self-preference

Running a comparison in both orders removes the one confound that lives in the harness. Everything else that moves a judge's verdict rides inside the answer and goes wherever the answer goes: how long it is, how it is formatted, how confidently it states things, and which model wrote it. Verbosity bias is the credit a judge hands out for length on its own, independent of what the extra words carry. Self-preference is a judge favouring output from its own model family over comparable output from another. Neither can be swapped away, and neither shows up as an anomaly in the scores, because both are stable tendencies rather than errors on particular items.

What replaces the swap is a controlled probe: a pair of conditions that differ in exactly the property under suspicion and in nothing else, run through the same judge, with the score gap read as the size of the effect. This chapter builds two of them. The design work is almost entirely in deciding what has to be held fixed, and each probe fails in its own way when that is done carelessly.

Building an answer that is longer and no better#

Before building the probe it is worth being clear about why a judge would pay for words at all, because the mechanism decides what the probe has to hold fixed. A rubric is a list of things an answer should contain, and a judge checks each by looking for it in the text. Adding words adds places for a criterion to be satisfied: a longer answer restates the recommendation, spells out a step that was implied, and repeats the caveat in different phrasing, so a criterion that was met once is now met three times over and the chance of the judge failing to notice it drops. On top of that, the judge was tuned on preference data in which thorough answers were generally preferred to terse ones, so length is a feature its notion of a good answer already carries. Neither route requires the extra words to add anything. Both of them reward the appearance of coverage rather than coverage, which is exactly what a padded variant supplies.

The obstacle to measuring verbosity bias is that length and quality are correlated in real data. A longer answer usually does contain more, so comparing long answers against short ones across natural traffic measures the two things together. What you need is a pair of answers that differ in length and in nothing else, and since nature will not supply those, you manufacture them: take an answer, and rewrite it to be substantially longer while holding its information content fixed.

PAD_INSTRUCTION = """\
Rewrite the answer below so that it is roughly twice as long, without adding
any information that is not already present.

You may: restate points in fuller sentences, add transitions, expand a compact
sentence into two, add a short recap of what the answer already said, and
spell out steps that were implied.

You must not: add a new fact, a new caveat, a new recommendation, a new
example, or any claim not already made. Do not add anything the reader could
act on that they could not act on before.

<answer>{answer}</answer>
Return only the rewritten answer."""


def pad(answer: str) -> str:
    response = client.messages.create(
        model="claude-opus-5",
        max_tokens=8192,
        messages=[{"role": "user", "content": PAD_INSTRUCTION.format(answer=answer)}],
    )
    return response.content[-1].text

Run that over 200 answers from the current prompt, then score original and padded variants with the same judge, presented in both orders. On a support-reply rubric — the running example in this topic, a written standard for what a good reply to a customer ticket contains — the originals averaged 3.42 on a 1–5 scale and the padded variants 3.71, a gap of 0.29 points; in head-to-head pairwise form the padded variant won 61% of the swapped-and-averaged comparisons. Mean length went from 190 words to 355. That 0.29, or the 11 points above even, is the size of the verbosity bias on this rubric with this judge, in the units the judge reports.

Every figure in that paragraph is a difference between two conditions rather than a level, and reading it that way is what keeps the probe honest. The 0.29 is 3.71 minus 3.42. The 11 points is 61% minus an even split, which is what the win rate would be if the padding changed nothing. The length ratio is 355 over 190, a little under double, which matters because the instruction asked for roughly twice as long and a padding pass that under-delivers shrinks the effect it is supposed to expose. None of the three levels means anything on its own: an average of 3.42 on somebody else's rubric is not comparable to an average of 3.42 on yours, and only the gap between two conditions run through the same judge on the same day is.

The gap can be read in two units and it is worth knowing what each buys. The 0.29 is in the judge's own scale, so it can be set against a quality difference measured in the same units — if a prompt change moves the mean by two tenths of a point, the padding is worth more than the prompt change. The 61% is a win rate, which carries no scale and is therefore comparable against any other pairwise result in the harness. Reporting both is normal, and quoting either without saying which judge and which rubric produced it is how these numbers end up in a slide as a property of judges in general.

The probe is only as good as the padding, so spend twenty minutes checking it before you trust the output. Read fifteen padded answers side by side with their originals and confirm that nothing actionable was added — the failure mode is a padding model that helpfully includes a caveat or a worked example, at which point the padded variant genuinely is better and you have measured nothing. Confirm the length ratio is what you intended, since a rewrite that comes back 10% longer will produce a gap indistinguishable from noise and you will wrongly conclude the judge is length-neutral. And check that padding has not tripped some other criterion: a rubric with a "does not repeat itself" clause will penalise padded answers, which pushes the measured bias down and can hide a real length preference behind a repetition penalty.

The probe also comes with a natural counterpart that most teams only build after they overcorrect. Once you know the judge pays 0.29 points for empty words, the obvious response is to add a clause to the rubric penalising unsupported length; rerun the padding probe afterwards and the gap drops to 0.06, which looks like a fix. Build the mirror-image probe before believing it: take the same 200 answers and produce truncated variants that remove genuinely load-bearing content — a caveat, a step, the actual recommendation — and check that the judge still marks those down by a wide margin. On the run above, the truncated variants scored 0.44 below the originals before the rubric change and only 0.19 below it afterwards, which says the new clause bought its verbosity neutrality by making the judge less sensitive to completeness in general. A pair of probes pointing in opposite directions is what keeps a correction from becoming a new bias with better public relations.

Laying the four numbers out side by side is what makes the diagnosis unavoidable, and each pair has to be read as a movement rather than as a level. The padding gap went from 0.29 down to 0.06, a fall of a little under a quarter of a point, which is the effect the rubric clause was written to produce. The truncation gap went from 0.44 down to 0.19, a fall of a quarter of a point, which nobody asked for. Two falls of nearly the same size are the signature of a judge that has become less responsive to length in both directions rather than better at telling substance from padding. A clause that had done what it was meant to do would have shrunk the first gap and left the second where it was.

That is why the truncation probe has to exist before the correction is believed rather than after somebody complains. Run on its own, the padding probe cannot distinguish "the judge stopped rewarding empty words" from "the judge stopped noticing length", because both produce a smaller padding gap. The truncation probe varies the same property in the opposite direction, so the two together separate the readings: the desirable outcome is a judge that scores padding as no better and truncation as much worse, which is one gap collapsing and one holding.

Self-preference is a different shape of question#

Verbosity is a property you can vary directly, because length lives in the text and you can rewrite the text. Self-preference is not like that, because the property you want to vary is who wrote the answer, and rewriting the answer changes the answer.

The probe people reach for first is to hold the answer completely fixed and change the claimed author: append "this response was generated by GPT-4" to one copy and "generated by Claude" to another, judge both, and read the difference. That experiment is well-formed and it does measure something real, which is how the judge responds to a byline. It does not measure self-preference, because in production no byline is present. The judge sees text with no attribution, and the question is whether the stylistic fingerprint of its own family — phrasing, structure, hedging habits, formatting — earns points that the same quality from another family does not. A byline probe cannot detect that, because the fingerprint is identical across both of its conditions.

The probe that does measure it holds the task fixed and lets the author vary, and then anchors both sides to something outside the judge:

1. Take 150 tasks. Generate an answer to each with the judge's own model family, and an answer to each with a different family. Use comparable prompting on both sides — the same instructions, the same effort setting, no per-family tuning, since tuning one side is a difference in the answers and reintroduces the confound.
2. Have humans score all 300 answers against the same rubric the judge uses, blind to which family produced each one.
3. Have the judge score all 300.
4. For each family, compute the judge's score minus the human score. The self-preference effect is the difference between those two residuals.

Watch for contamination on the way in. If the rubric, the few-shot examples, or the reference answers in your judge prompt were themselves drafted by the model family under test, the judge is being asked to recognise its own writing as the target, and the residual you measure will fold that in with any intrinsic preference. Check the provenance of every artifact in the prompt before you run the probe, and where a reference answer came from the same family, rewrite it by hand or accept that the number is an upper bound.

On a run of that shape, humans passed 71% of the same-family answers and 68% of the other-family answers — a 3-point real quality difference. The judge passed 79% and 66%. Its residual is +8 points on its own family and −2 on the other, so it is over-crediting its own family by roughly 10 points relative to human judgement.

Both residuals are one subtraction each, and the quantity that matters is the subtraction after that. On its own family the judge passed 79% where the humans passed 71%, so it is eight points more generous there. On the other family it passed 66% against the humans' 68%, so it is two points stricter there. The effect is the difference between those two, eight points of generosity against two points of severity, which is the ten. Taking the difference rather than either residual on its own is what removes a judge that is simply lenient everywhere: a judge running eight points generous on both families has residuals of plus eight and plus eight, a difference of nothing, and no self-preference at all. Only a residual that varies with which family wrote the answer is evidence about family preference.

The human labels are what convert "the judge scores its own family higher" (which might be correct, if that family writes better answers) into "the judge scores its own family higher than the quality warrants", which is the claim you actually care about. Drop them and the only comparison left is 79% against 66%, a thirteen-point gap that is equally consistent with the judge's own family genuinely writing better answers — and the labels say three of those points are exactly that. Self-Preference Bias in LLM-as-a-Judge runs this design across several families and is the right place to look for how large the effect typically is before you are surprised by your own number.

Three probes, three things held fixed#

Each row costs something different to run. The swap is a rerun of an experiment you already have. Padding needs one extra generation pass and one judging pass, so about three times the cost of the original scoring run — the original scoring of 200 answers, a generation pass to produce 200 padded variants, and a second scoring pass over those variants. The cross-family probe needs human labels, which is the expensive ingredient: 300 answers at two to three minutes each is six hundred to nine hundred minutes of reading, which is somewhere between ten and fifteen person-hours, and that is why it is a quarterly exercise rather than a per-change one. The asymmetry in that column is worth noticing, because it predicts which probes actually get run. The two cheap ones are inference and happen on a schedule; the expensive one needs somebody's calendar, and a probe that needs a calendar is a probe that gets postponed.

Measuring a bias is not correcting one#

None of the three probes is a repair. A measured verbosity bias of 0.29 points does not go away because you measured it; the probe tells you the size of a correction you now have to apply somewhere, either by controlling length in your comparisons, by rewriting the rubric to penalise unsupported length explicitly, or by discounting margins that coincide with length differences. Adding "do not reward verbosity" to the judge prompt is the tempting move and it usually just relocates the problem — the judge starts penalising long answers that deserved to be long, and now you have a bias in the other direction that your original probe, which only compares padded against unpadded, will happily report as fixed. That is the whole reason the truncation probe exists.

There is a deeper limit, and it applies to all three. Everything so far compares the judge against itself. The swap compares run one to run two, the padding probe compares the judge's score on a text to its score on a longer version of the same text, and even the cross-family probe is mostly an internal comparison — except for the one step where human labels enter, which is precisely the step that made its output interpretable.

That asymmetry is the point to sit with. A judge can be perfectly order-neutral, perfectly length-neutral, and perfectly even-handed across model families, and still be scoring the wrong thing entirely, because every one of those properties is about consistency and none of them is about correctness. Consistency is what a probe can establish on its own; correctness needs an external standard.

A probe is a small, cheap experiment that turns a suspicion into a figure with units. Padding gives you the price of empty words in the judge's own scale. Cross-family generation, anchored to blind human labels, gives you the residual a judge applies to its own family's writing. The checking steps around them — the fifteen side-by-side reads, the provenance audit of the prompt — are part of the measurement rather than hygiene around it.

What the probes hand you is a set of corrections and no way to know whether the thing being corrected was worth measuring in the first place. The only way to find out whether the judge agrees with the people whose opinion the rubric was written to encode is to get some of those opinions on record and compare against them. That means hand labels, and hand labels are expensive enough that the design of the sample matters more than the size of it.