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

This paper investigates whether large language models (LLMs) exhibit a 'self-preference' bias when revising their own writing. While prior research has shown that models tend to favor their own outputs when acting as judges, it remains unclear if this bias extends to rejecting valid, externally verified corrections to their own drafts. The authors seek to determine if models are systematically resistant to improving their own work compared to a neutral 'fresh' model.

The researchers utilize the IFEval dataset, which provides deterministic, program-based verification for instruction-following tasks. By using this objective checker, they avoid the circularity of using an LLM to judge the quality of its own revisions. The study compares two conditions: an 'author' model, which attempts to revise its own previously generated draft, and a 'fresh' model, which evaluates the same draft and correction without having seen the generation history. The study spans four mid-tier model families and 85 unique draft-correction pairs.

The study finds no statistically significant evidence of self-preference. Across the four model families, the author models accepted verified-good fixes at rates essentially identical to those of fresh models (a gap of -5.1 percentage points, with a 95% confidence interval spanning zero). Furthermore, the authors observed that when models did reject a verified-good fix, 97% of their stated reasons were based on 'flaw-catching'—arguing that the fix was insufficient or broke other constraints—rather than a preference for their original output. This suggests that the rare instances of rejection are driven by perceived quality concerns rather than ego-driven bias.

These findings suggest that the self-preference bias observed in free-form quality-judging tasks does not necessarily translate to verifiable, constraint-based revision. For developers building agentic pipelines, this is a positive result: it implies that models can be effectively integrated into self-correction loops for verifiable tasks without being inherently handicapped by an inability to accept their own corrections. However, the authors caution that this does not preclude the existence of such biases in more subjective, high-latitude writing tasks.