💻 computer science

Reproducing FACTER: Fairness via Conformal Thresholding and Prompt Repair

This paper presents a reproducibility study of the FACTER framework for fair LLM-based recommendations, revealing that while it effectively reduces adaptive-threshold violations, its iterative prompt repair mechanism offers limited additional benefit over static fairness instructions in constrained re-ranking scenarios and suffers from utility divergence due to underspecified evaluation in the original study.

Original authors: Oscar Miró López-Feliu, Daimy van Loo, Xanthos Kekkos, Mikel Blom, Clara Rus

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Oscar Miró López-Feliu, Daimy van Loo, Xanthos Kekkos, Mikel Blom, Clara Rus

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 have a very smart, but slightly biased, digital librarian (a Large Language Model) who suggests movies to people. You want this librarian to be fair: if two people have the same taste in movies, they should get the same suggestions, regardless of whether one is a man, a woman, young, or old.

A new method called FACTOR was proposed to fix this. It works like a two-step safety net:

The Alarm System: It sets a "fairness score." If the librarian suggests something that looks too much like a stereotype (e.g., only suggesting romance movies to women), the alarm rings.
The Repair Crew: When the alarm rings, the system writes a note to the librarian saying, "Hey, stop doing that specific thing," and adds it to a list of rules to avoid in the future. It also adjusts the alarm's sensitivity to be a bit more lenient if it keeps ringing too often.

The Goal of This Paper
The authors of this paper wanted to see if FACTER actually works as advertised. They tried to rebuild the system from scratch using the original instructions and code, like a mechanic trying to rebuild a car engine using only the owner's manual. They tested it on two things:

The Original Test: Asking the librarian to just "make up" movie titles from scratch.
A New Test: Giving the librarian a fixed list of 40 movies and asking them to pick the best 10 from that list (like a re-ranking task).

Here is what they found, explained simply:

1. The "Magic" Didn't Work in the Original Test

In the original setup (where the librarian had to invent movie titles from thin air), the results were a disaster. The librarian couldn't even remember the names of the movies correctly. It was like asking someone to name a movie, and they kept saying "The one with the guy and the dog" instead of "The Lion King."

Because the librarian was so bad at just naming movies, the "fairness" numbers looked weird. The authors realized that the original study might have been using a very loose way of checking if the movie names matched, which made the results look much better than they actually were.

2. The "Repair Crew" Was Just Moving the Goalposts

The most interesting finding was about how FACTER reduced the number of "alarms" (violations).

The Claim: FACTER was supposed to make the librarian behave better by learning from its mistakes.
The Reality: The system didn't actually make the librarian behave better. Instead, it just lowered the bar for what counts as a mistake.

Imagine a teacher grading a test. If the student keeps getting questions wrong, the teacher could either:
A) Teach the student better (fix the behavior).
B) Lower the passing grade so the student passes anyway (adjust the threshold).

FACTER mostly did Option B. It kept the "alarm" sensitivity low enough that fewer things triggered the alarm, but the librarian's actual suggestions didn't get much fairer. When the authors checked with a strict, unchanging standard, the "fairness" improvements mostly disappeared.

3. A Simple Note Works Just as Well as the Complex Robot

The authors created a simpler version of the system: a "Fair Zero-Shot" baseline. This is just a librarian with a single, static note on their desk that says, "Be fair to everyone."

They found that in the "fixed list" test (where the librarian picks from 40 movies), this simple note worked just as well as the complex, self-correcting FACTER system. The fancy "repair crew" that constantly updates the rules didn't add any extra value. It was like hiring a team of engineers to fix a leaky faucet when a simple wrench would have done the job.

4. The "Black Box" Problem

The study highlights a major issue with these AI systems: they are "black boxes." We can't see inside their brains to fix them directly. We can only talk to them (prompting) and watch what they say.

The authors found that when they tried to force the AI to be fair, the AI often prioritized being "confident" in its wrong answers over actually being fair.
The system's "fairness penalty" (the part that punishes bias) actually got worse over time, even though the system was trying to fix it. The AI just learned to ignore the fairness rules to keep its confidence high.

The Bottom Line

The paper concludes that while FACTER is a clever idea, it doesn't quite work the way the original authors claimed.

It reduces the number of "alarms" it triggers, but mostly by making the rules easier to pass, not by making the AI smarter or fairer.
In many cases, a simple, static instruction to "be fair" works just as well as the complex, self-updating system.
The original claims about how well the system works in an open-ended setting (making up movie titles) couldn't be reproduced; the system struggled to even name the movies correctly, let alone be fair.

In short: The "self-correcting" magic trick of FACTER is mostly an illusion. It looks like it's fixing the problem by adjusting the scoreboard, but the player (the AI) isn't actually playing the game any fairer.

Technical Summary: Reproducing FACTER

Problem Statement

Large Language Models (LLMs) are increasingly deployed as recommendation systems due to their ability to generate personalized suggestions in natural language. However, these models risk amplifying societal biases present in their training data, leading to systematically different recommendations for users from different demographic groups. While retraining models under fairness constraints is a common mitigation strategy, it is often impractical for black-box LLMs deployed via APIs. This necessitates model-agnostic, post-hoc interventions.

The paper focuses on reproducing FACTER (Fayyazi et al., 2025), a framework designed to enforce fairness in LLM-based recommendations without altering model parameters. FACTER combines conformal prediction to calibrate a nonconformity threshold with an iterative prompt repair mechanism that injects "AVOID" constraints based on detected violations. The original study claimed FACTER could reduce fairness violations without degrading recommendation utility. This work investigates the reproducibility of these claims, addressing ambiguities in the original implementation and evaluating the framework under both the original open-ended generation setting and a controlled re-ranking extension.

Methodology

Framework Implementation

The authors implemented FACTER from scratch, adhering to the original paper's theoretical formalizations while resolving implementation discrepancies found in the public codebase. The framework operates in two phases:

Offline Calibration: An initial acceptance threshold ( $Q^{(0)}_\alpha$ ) is established using split conformal prediction. A nonconformity score ( $S$ ) is computed for calibration instances, combining a predictive error term (deviation from ground truth) and a fairness penalty (semantic distance to recommendations given to "counterfactual neighbors" with different protected attributes).
Online Monitoring: As queries arrive, the system checks if the nonconformity score exceeds the current threshold. If a violation occurs:
- Prompt Repair: The system mines frequent item features associated with the violation and injects explicit "AVOID" constraints into the system prompt (e.g., "Avoid: gender=F $\to$ Romance").
- Threshold Adaptation: The threshold is updated via exponential smoothing to dynamically resize the acceptance region.

Experimental Setup

The study evaluates FACTER across two datasets (MovieLens-1M and Amazon Movies & TV) and three LLM backbones (Llama 2-7B, Llama 3-8B, Mistral-7B).

Tasks: The evaluation covers the original open-ended generation task (where the LLM generates item titles from scratch) and a new constrained re-ranking task (where the LLM ranks a fixed candidate pool).
Baselines: The study introduces a Fair Zero-Shot baseline, which uses static fairness instructions without the iterative repair loop, to isolate the contribution of the dynamic mechanism.
Metrics: Fairness is measured via Sensitive-to-Neutral Similarity Ratio (SNSR) and Counterfactual Fairness Ratio (CFR), alongside Violation Counts. Utility is measured via Recall@k and NDCG@k.

Key Results

1. Reproducibility of Open-Ended Generation

Under strict reproduction of the open-ended generation task, the authors observed a significant divergence in recommendation utility.

Utility Gap: The reproduced model yielded near-zero NDCG@10 and Recall@10 scores, contrasting sharply with the strong results reported in the original study. This gap is attributed to the difficulty of zero-shot generation and underspecified target-set evaluation protocols in the original work (e.g., string-matching tolerances and relevance set construction).
Fairness Trends: Despite the utility gap, FACTER consistently reduced adaptive-threshold violation counts compared to the neutral baseline. However, these reductions did not consistently translate to improvements in global fairness metrics (SNSR, CFR). In some cases, FACTER's CFR was worse than the neutral baseline.

2. Performance in Re-Ranking Extension

In the controlled re-ranking setting, where the candidate set is fixed:

Utility: Recommendation utility was markedly higher than in open generation, confirming that the underlying LLMs can identify relevant items when the search space is constrained.
Fairness: FACTER maintained low adaptive violation counts. However, static fairness instructions (Fair Zero-Shot) achieved comparable semantic-parity outcomes (SNSR, CFR) and fixed-threshold violation counts to the dynamic FACTER loop.
Threshold Dynamics: The reduction in violations was found to be highly dependent on the adaptive threshold. When evaluated against a fixed, population-calibrated threshold, FACTER's violation counts were comparable to the baselines, suggesting the apparent improvement stems from threshold relaxation rather than fundamental behavioral changes.

3. Mechanism Analysis

Component analysis of the nonconformity scores revealed that while the predictive error marginally improved over iterations, the fairness penalty actually increased. This indicates that the injected "AVOID" constraints often failed to override the model's predictive distribution, potentially destabilizing counterfactual stability. The dynamic repair loop provided limited additional benefit over static prompting in the constrained ranking formulation.

Significance and Claims

The paper positions itself as a best-effort reproduction audit rather than a definitive adjudication of the original results, citing ambiguities in the original paper and code. Its primary contributions are:

Documentation of Limits: It highlights that the original open-generation utility claims could not be reproduced under matched settings, likely due to underspecified evaluation protocols.
Isolation of Mechanisms: By introducing a Fair Zero-Shot baseline, the study demonstrates that the iterative prompt repair loop provides limited marginal benefit over static fairness instructions in constrained ranking tasks.
Threshold vs. Behavior: The analysis suggests that FACTER's primary mechanism for reducing violation counts is adaptive threshold relaxation rather than the successful erasure of bias via prompt repair.
Reproducibility Constraints: The work underscores the necessity of transparent evaluation protocols in open-ended generative tasks and the challenges of reproducing model-agnostic fairness frameworks when implementation details (e.g., target-set construction, metric computation) are ambiguous.

The authors conclude that while conformal prediction effectively bounds internal violation counts, the practical necessity of the online repair loop is questionable in constrained settings, where static instructions may offer a more efficient alternative. They call for future research to explore mechanisms that enforce semantic bounds during decoding and to better calibrate fairness penalties against predictive confidence.