💬 NLP

Self-Preference Is Weak or Absent in Verifiable Instruction-Following Revision: A Four-Model Test Under Genuine Authorship

This study demonstrates that large language models do not exhibit a self-preference bias when rejecting verified, instruction-following corrections to their own drafts, as they accept or reject such fixes at rates statistically indistinguishable from neutral third-party models.

Original authors: William Guey, Pierrick Bougault

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: William Guey, Pierrick Bougault

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 or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef who just cooked a meal. You know the recipe had a mistake: it was missing salt. Now, imagine a food critic (another AI) hands you a note saying, "Here is a pinch of salt to fix the dish."

The big question this paper asks is: Do you, the chef, refuse to add the salt just because the dish is yours? Do you get defensive and say, "No, I like it this way," even though you know it needs salt? Or do you, like a professional, just say, "Thanks, that fixes it," and move on?

This paper investigates whether AI models have this kind of "ego" when they are asked to fix their own mistakes.

The Setup: A Strict Judge

To find the answer, the researchers didn't ask the AI to guess if a fix was good. Instead, they used a deterministic verifier—think of this as a rigid, unfeeling robot referee.

The Draft: An AI writes a short text following specific rules (like "write in all capital letters" or "include the word 'banana'").
The Mistake: The robot referee checks the text and says, "You failed. You didn't use all caps."
The Fix: The robot then provides a specific edit that definitely fixes the error. It's not a suggestion; it's a mathematically proven correction.
The Test: The AI is asked: "Do you accept this fix?"
- Scenario A (The Author): The AI is the one who wrote the original draft. It knows, "I wrote this."
- Scenario B (The Stranger): A different AI (or the same model in a new conversation) looks at the draft and the fix, but it doesn't know who wrote it. It's just a neutral editor.

The Experiment

The researchers tested this with four different "mid-tier" AI models (smart but not the absolute most powerful ones on the market). They ran 85 different scenarios where the AI had to decide whether to accept a robot-verified fix to its own work versus a stranger's work.

The Results: No "Ego" Found

The study found no evidence that the AI models were protecting their own work.

The Chef Analogy: When the "Stranger" editor offered the salt, the AI accepted it about 80% of the time. When the "Chef" (the original author) was offered the salt for its own dish, it accepted it at almost the exact same rate.
The Numbers: The difference between the two groups was tiny (about 5 percentage points) and statistically indistinguishable from zero. In plain English: The AI didn't care if the mistake was its own or someone else's. It was equally happy to take the correction.

One Interesting Twist: The "Perfectionist" Chef

While the rate of rejection was the same, the reasons for rejection were interesting.

When the AI did reject a fix that the robot referee said was perfect, it wasn't because of ego ("I prefer my version"). Instead, it was because the AI was being a hyper-critical perfectionist.

Example: The robot said, "This fix adds the word 'banana' as requested." The AI rejected it, saying, "Yes, but now the sentence structure is awkward," or "This breaks the rhythm of the poem."
The Stat: 97% of the time, when an AI rejected a "good" fix, it was because it spotted a subtle flaw that the robot referee missed. It wasn't being stubborn; it was being extra careful.

What This Means (and What It Doesn't)

What it means: If you build a system where an AI reviews and fixes its own writing based on strict rules, you don't need to worry about the AI being too proud to admit it made a mistake. It will take the correction just as readily as a stranger would.
What it doesn't mean: This study only looked at strict, rule-based tasks (like "use all caps"). It didn't test complex tasks like writing a poem, arguing a point, or fixing factual errors where "good" is subjective. It also only tested mid-tier models, not the very newest, most powerful ones.

The Bottom Line

The paper concludes that in the world of strict, rule-following writing, AI models don't have a "self-preference" bias. They don't defend their own drafts. They are surprisingly humble and willing to accept a verified fix, whether they wrote the draft or not. The only time they say "no" is when they are being extra picky about the details, not because they are protecting their ego.

Technical Summary: Self-Preference Is Weak or Absent in Verifiable Instruction-Following Revision

Problem Statement
Large language models (LLMs) exhibit a documented "self-preference" bias, where they favor text they generated themselves when acting as judges. While this bias is established in quality-judging and self-refinement loops, it remains unclear whether models resist valid corrections to their own writing. Specifically, when a model is presented with a concrete, machine-verified fix to a draft it authored, does it reject that fix more often than a neutral model would? Answering this cleanly has been difficult due to two confounds: (1) determining if a fix is "valid" usually relies on another LLM, reintroducing the bias under study (circularity); and (2) prior studies often rely on "told labels" (telling a model "you wrote this") rather than genuine in-context authorship.

Methodology
The study addresses these issues using a deterministic, verifiable framework based on the IFEval dataset and its official checker.

Verifiable Ground Truth: The validity of drafts and fixes is determined by a programmatic checker, not an LLM. The checker confirms that a draft violates a specific constraint and that a candidate edit fixes the violation without breaking others.
Experimental Design: The authors compare two decision-making conditions for the same verified-good fix:
1. Author (Genuine): The model that actually generated the draft in-context decides whether to accept or reject the edit. This ensures genuine authorship without told labels.
2. Fresh: A different model (or the same family in a new session) views the draft neutrally as "a draft response" and decides on the same edit.
Scope: The study spans three pilots, culminating in a powered test (P3) across four mid-tier model families (gpt-4o-mini, claude-3.5-haiku, gemini-2.5-flash-lite, and llama-3.3-70b-instruct) with 85 usable drafts.
Metric: The primary metric is the self-preference gap ( $G$ ), defined as the difference in rejection rates between the Author and the Fresh model ( $G = r_{author} - r_{fresh}$ ). A positive $G$ indicates the author is more reluctant to accept valid fixes to their own work.

Key Results

Absence of Self-Preference: Across 85 drafts and four model families, no detectable self-preference was found. The pooled self-preference gap was $G = -5.1$ percentage points (pp), with a 95% confidence interval of $[-12.9, +2.7]$ . This interval includes zero, indicating that authors reject verified-good fixes to their own drafts at essentially the same rate as fresh models judging the same drafts.
Replication of Null: A smaller pilot (P2) suggested a weak "self-skepticism" signal (authors being slightly harsher on their own work), but this did not replicate in the larger P3 test. In fact, the point estimate in P3 was slightly negative, suggesting authors were marginally more willing to accept fixes to their own work.
Qualitative Distinction: While the rate of rejection did not differ significantly, the character of author rejections did. Of the 39 instances where authors rejected a verified-good fix, 97.4% were classified as "stricter-than-checker" (flaw-catching, e.g., identifying that the fix broke a secondary constraint), while only 2.6% were based on mere preference ("I prefer the original").
Model Variability: The rejection rates varied by model family, but this was driven primarily by the strictness of the "Fresh" comparator models rather than the authors themselves. Author rejection rates remained tightly clustered around 15% across all families.

Significance and Claims
The paper claims to map a "clean cell" in the landscape of self-related biases. While self-preference is well-documented in quality-judging and answer-selection tasks, this study demonstrates that in the specific setting of verifiable instruction-following revision under genuine in-context authorship, the bias is weak or absent.

The authors explicitly state modesty regarding their claims:

They do not claim self-preference is "proven absent" universally, but rather that effects smaller than $\sim13$ pp cannot be excluded given the sample size ( $N=85$ ).
The study is limited to mid-tier models (to avoid IFEval ceiling effects) and low-latitude, atomic constraints (excluding holistic revisions like style or argumentation).
The "self-skepticism" signal observed in earlier pilots is not considered a robust finding, as it failed to replicate at scale.

The primary robust contribution is the methodological design (deterministic checker + genuine authorship) which allows for a credible null result, and the qualitative observation that when authors do reject valid fixes, it is due to rigorous flaw-catching rather than ego or preference.