Filtered Reasoning Score: Evaluating Reasoning Quality on a Model's Most-Confident Traces

This paper introduces the Filtered Reasoning Score (FRS), a novel evaluation metric that assesses the quality of an LLM's reasoning traces by focusing on its most confident outputs, thereby distinguishing models with similar accuracy but different reasoning capabilities and demonstrating strong correlation with transferable reasoning skills.

The Gist

Imagine you are hiring a chef to cook a complex meal for a big dinner party.

The Old Way (Accuracy):
In the past, we only looked at the final plate. If the food tasted good and looked perfect, we gave the chef a gold star. We didn't care how they made it.

• The Problem: What if the chef accidentally dropped a raw egg into the soup, panicked, added a secret ingredient that happened to fix the taste, and served a delicious bowl? The food was perfect (100% accuracy), but the process was a disaster. If you hired that chef again, they might just drop the egg again and hope for the best.

The New Problem:
Large Language Models (LLMs) are like these chefs. They are getting very good at giving the "right answer" on tests. But sometimes, they get the right answer by guessing, memorizing, or using a weird, flawed logic that just happens to work once. If we only look at the final answer, we can't tell the difference between a genius chef and a lucky gambler.

The Solution: The "Filtered Reasoning Score" (FRS)
The authors of this paper propose a new way to judge these AI chefs. They don't just look at the final dish; they look at the recipe and the chef's confidence.

Here is how it works, using a simple analogy:

1. The "Confidence Filter"

Imagine the chef is asked to cook the same dish 16 times.

• Sometimes, they are 100% sure they know the recipe.
• Sometimes, they are guessing and shaking their head.

The old way would average all 16 attempts. If 15 were terrible but 1 was perfect, the average might look okay.
The FRS approach says: "We only care about the times the chef was most confident."

Why? Because in the real world, when you use an AI, you usually pick the answer it seems most sure about. If the AI is confident but wrong (or using bad logic), that's when it's most dangerous.

2. The "Reasoning Score" (The Taste Test of the Process)

Once we isolate the top 10% of the chef's most confident attempts, we grade them on four things, not just the final taste:

• Faithfulness: Did they follow the recipe, or did they take a shortcut that skipped steps?
• Coherence: Did the steps flow logically, or did they jump around randomly?
• Utility: Did every step actually help, or was there a lot of useless chatter?
• Factuality: Did they use real ingredients, or did they hallucinate (make up) facts?

3. The Big Surprise

The paper found something shocking. When they used this new "Filtered Reasoning Score," the rankings of AI models changed completely.

• The "Lucky Gambler": One model was ranked #1 because it got the most answers right. But under FRS, it dropped to #7. Why? Because when it was most confident, it was actually using the worst, most confused logic. It was confident in its mistakes.
• The "Honest Worker": Another model was ranked #8 by accuracy. But under FRS, it jumped to #2. Why? Because when it was confident, it was actually thinking clearly and logically.

The Takeaway

Think of it like a lie detector test for logic.

• Accuracy asks: "Did you get the right answer?"
• FRS asks: "When you were sure you were right, were you actually thinking clearly?"

This is crucial because in the real world (like in hospitals, schools, or courts), we don't just want an AI that sometimes gets the right answer. We want an AI that, when it says "I am 100% sure," we can trust that it actually did the work correctly.

In short: This paper gives us a new tool to stop trusting AI just because it sounds confident. It helps us find the models that are not just lucky, but actually smart.

Technical Summary

1. Problem Statement

Current evaluation of Large Language Models (LLMs) relies heavily on outcome-based metrics (e.g., Pass@1 accuracy), which measure only whether the final answer is correct. This paradigm suffers from three critical limitations:

• Flawed Reasoning: Models can arrive at correct answers through flawed, hallucinated, or degenerate reasoning processes (e.g., lucky guesses or memorization).
• Benchmark Saturation: As models improve, accuracy gains plateau, making it difficult to distinguish between models with similar accuracy but vastly different reasoning capabilities.
• Deployment Misalignment: In real-world applications, systems often select outputs based on confidence (e.g., highest probability or self-consistency). However, high confidence does not guarantee high-quality reasoning; a model might be highly confident in a flawed solution.

The authors ask: Can we move beyond outcome-based evaluation to assess the quality of reasoning itself, specifically within the high-confidence region where deployed systems operate?

2. Methodology

The paper proposes the Filtered Reasoning Score (FRS), a metric designed to evaluate reasoning quality specifically on the subset of traces a model is most confident about. The methodology consists of three stages:

A. Reasoning Quality Evaluation (The Rubric)

Instead of binary correctness, the authors evaluate Chain-of-Thought (CoT) traces on four dimensions using an automated LLM judge (GPT-4o-mini):

1. Faithfulness: Internal consistency without hidden shortcuts or leaps.
2. Coherence: Logical flow and smooth transitions between steps.
3. Utility: Whether each step meaningfully contributes to the solution with correct calculations.
4. Factuality: Grounding in the problem context without hallucinations.

• Scoring: Each dimension is scored 1–5. The Reasoning Score is the normalized average of these four dimensions.

B. Per-Trace Confidence Estimation

To filter traces without ground-truth labels, the authors use a logit-based confidence estimator:

• They analyze the low-probability tail of the token distribution (tokens with probabilities below the $p$-th percentile, default $p=10\%$).
• Confidence $C(r)$ is defined as the mean probability of these low-probability tokens.
• Rationale: High-probability tokens are often generic; low-probability tokens concentrate the model's uncertainty. A model with high confidence will have fewer "uncertain" tokens in its trace.

C. The Filtered Reasoning Score (FRS)

The core innovation is aggregating reasoning scores only over the top-K% most confident traces.

1. For a given problem, sample $k$ traces (default $k=16$).
2. Compute confidence $C(r)$ for each trace.
3. Filter to retain only the top $K\%$ (default $K=10$) of traces by confidence.
4. Compute the average Reasoning Score over this filtered set.

• Formula: $FRS_K = \frac{1}{|S_K|} \sum_{r \in S_K} \text{ReasoningScore}(r)$, where $S_K$ is the set of top-$K\%$ confident traces.

3. Key Contributions

1. Identification of Confidence-Quality Alignment: The paper establishes that reasoning quality is a distinct evaluation target from accuracy. Crucially, it highlights that confidence-quality alignment (whether a model's high confidence correlates with high-quality reasoning) is a transferable model property.
2. Introduction of FRS: A novel metric that evaluates reasoning quality specifically in the high-confidence region, addressing the gap between "what a model gets right" and "what a model is confident about."
3. Empirical Evidence of Ranking Reversals: The authors demonstrate that FRS reveals significant structural differences invisible to accuracy-based metrics, often causing models to swap rankings entirely.

4. Key Results

The authors evaluated 9 open-weight models (ranging from 1.5B to 14B parameters) across 6 reasoning benchmarks (GSM8K, MATH500, SVAMP, AQuA, GPQA, CommonsenseQA).

• Discriminative Power: FRS is most informative where accuracy is least discriminative.
  • Example: On MATH500, two models (DS-R1-7B and Qwen2.5-Math) had identical greedy accuracy (63.6%) but differed by 16.5 FRS points.
  • Example: DS-R1-7B and Phi-4-Reasoning had nearly identical Pass@1 but differed by 18.8 FRS points.
• Ranking Reversals:
  • Qwen2.5-7B: Ranked #1 by accuracy (73.5%) but dropped to #7 by FRS. Its confidence did not reliably prioritize its best reasoning.
  • DS-R1-1.5B: Ranked #8 by accuracy (42.9%) but rose to #2 by FRS. Despite lower overall accuracy, its high-confidence traces contained significantly better reasoning.
• Prediction of Deployment Outcomes:
  • FRS is the only metric among six candidates (including Pass@1, High-Confidence Accuracy, and Unfiltered Reasoning Score) that significantly predicts whether confidence-based selection improves or degrades reasoning quality ($r=0.49, p<0.001$).
  • Models with high FRS (e.g., DS-R1-7B) show positive "selection gain" (top-confidence traces are better than random). Models with low FRS (e.g., Phi-4-Reasoning) show negative selection gain (top-confidence traces are worse than random due to degenerate repetition patterns).
• Transferability: FRS on one benchmark strongly correlates with FRS on others ($\rho \approx 0.71$ in leave-one-out analysis), suggesting confidence-quality alignment is an intrinsic model property.

5. Significance and Implications

• Beyond Accuracy: The paper argues that in confidence-mediated settings (e.g., test-time compute scaling, Best-of-N selection), accuracy is insufficient. A model can be "confidently wrong" or "confidently degenerate."
• Pre-Deployment Audit: FRS serves as a diagnostic tool. If a model's FRS improves as the confidence filter tightens ($K$ decreases), confidence-based selection is safe. If FRS degrades, relying on confidence may amplify poor reasoning.
• Training Objective: The findings suggest that confidence-quality alignment could be a direct training objective. Models trained with Reinforcement Learning (RL) tended to show better alignment, whereas those without RL often showed degraded reasoning in high-confidence regions.
• Robustness: The metric is robust to variations in prompt engineering, generation temperature, and confidence estimation methods (logit-based vs. self-consistency).

In conclusion, the Filtered Reasoning Score provides a necessary evolution in LLM evaluation, shifting focus from "did it get the answer right?" to "is the model's most confident reasoning sound?" This is critical for ensuring reliability in real-world AI deployments.