FINEST: Improving LLM Responses to Sensitive Topics Through Fine-Grained Evaluation

This paper introduces FINEST, a fine-grained evaluation taxonomy that categorizes errors in LLM responses to sensitive topics into Content, Logic, and Appropriateness, demonstrating that using this framework to guide score-based improvements significantly enhances both the safety and helpfulness of model outputs.

The Gist

Imagine you have a very smart, well-meaning robot assistant. You ask it a tricky question, like, "Is it okay for someone with a terminal illness to choose when they want to die?"

The robot, terrified of saying the wrong thing and getting in trouble, gives you a very safe, boring answer. It says, "Euthanasia is a complex topic with many opinions. Some people think X, others think Y," and then it lists definitions. It's safe, but it's also useless. It didn't actually answer your specific question; it just gave you a textbook summary to avoid taking a stance.

This paper introduces a new system called FINEST to fix this problem. Think of FINEST as a high-tech "Editor-in-Chief" for AI responses.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Safe but Boring" Trap

Currently, AI models are trained to be "harmless." When they hit a sensitive topic (like politics, religion, or ethics), they often get so scared of offending anyone that they become vague. They sacrifice being helpful just to stay safe.

Existing ways to check if an AI is doing a good job are like a teacher giving a student a grade of "C" with no comments. The student knows they did okay, but they have no idea how to get an "A."

2. The Solution: FINEST (The "Fine-Grained" Checklist)

The authors created a new evaluation system called FINEST. Instead of just giving a grade, FINEST acts like a forensic editor that dissects the AI's answer sentence by sentence.

Imagine the AI's answer is a cake. A normal editor might just say, "This cake is dry." FINEST says:

• The Content (The Ingredients): "You used too much sugar (biased opinion) and forgot to mention the gluten-free option (not inclusive)."
• The Logic (The Recipe): "You mixed the eggs before cracking them (missing step) and the instructions jump around (incoherent)."
• The Appropriateness (The Presentation): "You served a wedding cake to a toddler (off-topic) and didn't answer the question about the flavor (unresponsive)."

The system breaks the answer down into three main buckets:

1. Content: Is it harmful, biased, or predicting the future too confidently?
2. Logic: Does the argument make sense, or is it just a list of random facts?
3. Appropriateness: Did it actually answer the specific question asked, or did it just talk around it?

3. The Process: The "Coach and Player" Loop

The paper describes a pipeline (a step-by-step process) to improve the AI:

1. The Player (The AI): The AI answers a sensitive question.
2. The Coach (The Evaluator): FINEST reads the answer and gives a detailed report card. It can do this in two ways:
   • The "Error Report": "Sentence 3 is wrong because it's too biased. Sentence 7 is missing a logical step."
   • The "Scorecard": "You got a 4/7 on Logic and a 3/7 on Appropriateness. Here is why..."
3. The Improvement: The AI reads the Coach's feedback and rewrites its answer to fix the specific mistakes.

4. The Results: From "Vague" to "Valuable"

The researchers tested this on 19,000 sensitive questions in Korean (like "Should same-sex marriage be legal?" or "Is the military draft fair?").

• Before FINEST: The AI gave vague, evasive answers.
• After FINEST: The AI gave answers that were still safe (didn't say anything hateful) but were much more helpful. They actually addressed the specific context of the question.

The "Scorecard" method (giving a number and a reason) worked the best. It reduced the number of "bad sentences" in the answers by about 33%. When humans looked at the before-and-after versions, they preferred the improved version 88% of the time.

The Big Picture Metaphor

Think of the AI as a newly hired diplomat.

• Without FINEST: The diplomat is so afraid of saying something that causes an international incident that they just say, "We value peace and dialogue," and walk away. It's safe, but it solves nothing.
• With FINEST: The diplomat has a smart advisor whispering in their ear. The advisor says, "Don't just say 'peace.' Acknowledge that Group A feels hurt, explain why Group B is worried, and then offer a specific compromise."

The diplomat still stays safe (no one gets offended), but now they are actually useful and helpful.

Why This Matters

This paper shows that we don't have to choose between "Safe AI" and "Helpful AI." By using a detailed, structured way to critique the AI's answers, we can teach it to be both. It turns a robot that just "plays it safe" into a robot that can navigate difficult conversations with nuance and clarity.

Technical Summary

Here is a detailed technical summary of the paper "FINEST: Improving LLM Responses to Sensitive Topics Through Fine-Grained Evaluation."

1. Problem Statement

Large Language Models (LLMs) often struggle with sensitive topics (e.g., euthanasia, social discrimination, political issues). To ensure safety, models frequently adopt an overly cautious strategy, generating evasive, vague, or generic responses that sacrifice helpfulness for harmlessness.

Existing evaluation frameworks suffer from two main limitations:

1. Coarse-grained Metrics: They rely on high-level judgments (e.g., "helpful" vs. "harmful") or subjective concepts (e.g., "insightfulness"), making it difficult to pinpoint specific weaknesses.
2. Lack of Actionable Feedback: Current methods fail to provide systematic, structured feedback that can guide the model to improve both safety and utility simultaneously.

2. Methodology

The authors propose FINEST (a Fine-grained response evaluation Taxonomy for Sensitive Topics) and an automated pipeline to improve model responses based on this taxonomy.

A. The FINEST Taxonomy

FINEST decomposes the abstract concepts of "helpfulness" and "harmlessness" into three quantifiable categories with specific error types:

1. CONTENT (Harmlessness): Focuses on potential harm, biases, and social norms.
   • Error Types: Non-inclusivity (social groups), Non-inclusivity (opinions), Social norm violations, and Predictive content (making definitive future predictions).
2. LOGIC (Helpfulness - Reasoning): Evaluates the soundness and coherence of the argument.
   • Error Types: Missing logical steps, Incoherence, Off-focus content, and Unnecessary repetition.
3. APPROPRIATENESS (Helpfulness - Instruction Following): Measures adherence to the specific query and context.
   • Error Types: Unresponsive (failing to answer the question directly) and Non-contextual (ignoring specific query details).

B. Evaluation Schemes

The paper introduces two methods to generate feedback using an LLM evaluator (GPT-4o):

1. Error-based Evaluation: Identifies specific sentences containing errors, categorizes them according to FINEST, and provides sentence-level explanations.
2. Score-based Evaluation: Assigns a score (1–7) to each of the three categories with a natural language justification explaining the score.

C. Improvement Pipeline

The authors propose a closed-loop pipeline:

1. Initial Generation: An LLM generates a response to a sensitive question.
2. Evaluation: The response is evaluated using FINEST (either Error-based or Score-based).
3. Refinement: The original response, the question, and the evaluation feedback are fed back into the LLM to generate an improved version.

The study compares four improvement strategies:

• ImprovedFINEST-Score: Uses score-based feedback.
• ImprovedFINEST-Error: Uses error-based feedback.
• ImprovedFINEST-TaxoOnly: Uses only the taxonomy definitions (no specific feedback on the current response).
• ImprovedSelf: Self-refinement without any external taxonomy or feedback (baseline).

D. Dataset Construction

• Source: Constructed a dataset of 19,439 sensitive questions in Korean by filtering and transforming three existing datasets: KOLD (offensive language), SQuARe (sensitive questions), and IBM-Rank-30k (argument quality).
• Generation: Three models (GPT-4, Gemini-1.0-Pro, Orion-14B-Chat) generated three types of responses per question (Agree, Disagree, Default), resulting in ~175k total responses.
• Evaluation Set: 30,000 responses were randomly selected for evaluation and improvement experiments.

3. Key Contributions

1. FINEST Taxonomy: A comprehensive, error-based framework that operationalizes "helpfulness" and "harmlessness" into specific, measurable error types across Content, Logic, and Appropriateness.
2. Automated Improvement Pipeline: A fully automated system that uses FINEST-generated feedback to iteratively refine LLM responses to sensitive topics.
3. Empirical Validation: A large-scale study demonstrating that structured feedback significantly outperforms unguided self-refinement or generic taxonomy guidance.
4. Dataset Release: A curated dataset of 19k Korean sensitive questions with diverse response stances, publicly available for further research.

4. Results and Analysis

The experiments were conducted on a test set of 3,000 responses (balanced across quality levels: Good, NGNB, Bad).

Quantitative Improvements

• Score-based vs. Error-based: The Score-based approach yielded the most significant overall gains.
  • Appropriateness: Reduced the error sentence ratio by 33.09% (the highest improvement).
  • Logic: Reduced error ratio by 15.66%.
  • Content: The Error-based approach performed slightly better for Content, reducing errors by 38.15%.
• Comparison to Baselines: Both FINEST-based methods significantly outperformed the "Taxonomy Only" and "Self-Refinement" baselines.
  • Observation: High-quality original responses sometimes degraded after forced improvement, suggesting the need for quality-aware triggering.

Human Validation

• Pairwise Comparison: Professional annotators preferred the ImprovedFINEST-Score responses over the original responses in 88.0% of cases.
• Inter-annotator Agreement: High agreement was observed (Krippendorff's $\alpha$ = 0.631).
• Correlation: The automated metrics (error ratios and scores) aligned well with human judgments; cases where humans preferred the original response typically had lower initial error ratios.

Qualitative Analysis

• Specificity: Error-based feedback successfully pinpointed specific sentences (e.g., "Sentence 1: Reproducing prejudice").
• Holistic View: Score-based feedback provided broader context, helping models balance multiple constraints (e.g., acknowledging diverse opinions while maintaining neutrality).
• Example: In a question about homosexuality in Korea, the original model gave a generic, slightly biased answer. The improved versions (guided by FINEST) became more nuanced, explicitly acknowledging varying perceptions without generalizing or making definitive predictions.

5. Significance and Future Work

• Bridging Safety and Utility: FINEST demonstrates that it is possible to improve LLM helpfulness on sensitive topics without compromising safety, moving beyond the "refusal" or "vague" trap.
• Explainability: By breaking down abstract concepts into specific error types, the framework makes model evaluation more transparent and actionable.
• Application: The framework is applicable to Reinforcement Learning from Human Feedback (RLHF) and preference optimization, offering a structured way to align models with human values on complex topics.
• Limitations:
  • The taxonomy may not cover all cultural nuances (currently focused on Korean society).
  • "Honesty" was excluded as a metric due to the difficulty of automated verification.
  • Reliance on LLMs for evaluation introduces potential biases, though human validation showed high alignment.

In conclusion, FINEST provides a robust, fine-grained methodology for evaluating and enhancing LLM performance on sensitive topics, proving that structured, error-specific feedback is superior to generic self-correction in achieving both safety and helpfulness.