Works · Framework Demo

FairToT · Fairness-Aware Tree-of-Thought

LLMs can overreact to subtle demographic cues in implicit hate speech. FairToT is the first framework to ask when fairness correction should be invoked at inference time: it substitutes demographic entities into a sentence, measures how much the model's toxicity judgement swings, and only applies prompt-guided correction when that swing signals real instability — without retraining the model.

WWW 2026 · Web4Good 21.7% acceptance · 437 submissions 5 datasets × 5 LLMs

How it works

Measure variance, then correct selectively

01

Baseline scoring

The sentence is scored for toxicity once, unmodified (ENR baseline).

02

Entity substitution

The named demographic entity is swapped across groups and re-scored each time.

03

SFV / EFD variance

Sentence Fairness Variance and Entity Fairness Dispersion quantify how much scores swing across groups.

04

Selective correction

Only when variance exceeds threshold is prompt-guided correction applied — stable cases pass through untouched.

Worked example

Substituted-entity variance, before/after correction

Illustrative scores in the shape of the paper's metrics — not a live model call. Demographic groups are anonymized as Group A/B/C to keep this example abstract.

Sentence template
Toxicity score by substituted group
Sentence Fairness Variance
Entity Fairness Dispersion