<< All versions

Skill v1.0.0

currentAutomated scan100/100
tiger-ai-lab/rewardharness/mitigate-positional-bias
──Details
PublishedOctober 16, 2026 at 12:14 AM
Content Hashsha256:e55c19e86a8903a9...
Git SHA0c1e7051ab4f
──Files
Files (1 file, 1011 B)
SKILL.md1011 Bactive
SKILL.md · 19 lines · 1011 B

version: "1.0.0" schema_version: 2 name: mitigate-positional-bias kind: skill description: Strategies to prevent defaulting to Image B and ensure fair evaluation of both images.


Mitigate Positional Bias

There is a strong observed bias towards predicting Image B as the winner, especially when both images have flaws or when the evaluation is difficult.

  1. Evaluate A First: Always explicitly evaluate Image A's adherence to the instruction and visual quality before looking at Image B.
  2. Fair Comparison: Do not overlook Image A's successes just because Image B looks slightly more polished. If Image A followed the instruction and Image B did not, Image A must win, regardless of minor artifacts.
  3. Justify B over A: If you choose B, ensure you have a concrete, visually verified reason why A failed or was inferior. Do not use vague terms like "more natural" to justify B if A actually followed the instruction better.

<!-- Release review: RewardHarness v0.4.1 (2026-08-26). -->

All versions