💬 NLP

Measuring Reasoning Trace Legibility: Can Those Who Understand Teach?

This paper argues that reasoning language models should be evaluated not just on answer correctness but on the legibility of their reasoning traces, introducing "transfer utility" as a metric and revealing that top-performing models often produce less legible traces, with no intrinsic reward for legibility in current training methods.

Original authors: Dani Roytburg, Shreya Sridhar, Daphne Ippolito

Published 2026-03-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dani Roytburg, Shreya Sridhar, Daphne Ippolito

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 are a master chef (a super-smart AI) teaching a young apprentice (a smaller, weaker AI) how to cook a complex dish.

For a long time, we only cared if the final dish tasted good. If the master chef made a perfect steak, we assumed they were a great teacher. But this paper asks a crucial question: Just because the chef made a perfect steak, does that mean they explained how they did it in a way the apprentice can actually learn from?

The authors of this paper, "Measuring Reasoning Trace Legibility," argue that we need to stop just looking at the final answer and start grading the step-by-step thinking process (the "reasoning trace") itself.

Here is the breakdown of their findings using simple analogies:

1. The Problem: The "Genius but Mumbling" Chef

Currently, AI models are trained to get the right answer as fast as possible. They are rewarded for being efficient.

The Result: The smartest models often produce "reasoning traces" that are incredibly dense, full of jargon, or skip steps because they "just know" the answer.
The Analogy: Imagine a math genius who solves a difficult equation in their head in 2 seconds and writes down "42." It's correct, but if you ask them to teach a student, they might say, "Well, you just do the thing," and skip the actual math. The student learns nothing.

2. The New Tool: "Legi-Val" (The Teaching Test)

The researchers built a new testing framework called Legi-Val. Instead of just checking if the answer is right, they test Teachability.

They take the thinking process of a "Master Chef" (a strong AI) and feed it to a "Student" (a weaker AI) piece by piece.

The Test: Does the student get the right answer after reading just the first step? The first half? The whole thing?
The Goal: A "legible" trace is one where the student can follow the logic and arrive at the correct answer because of the explanation, not just because the answer was hidden at the end.

3. The Big Surprise: The "Accuracy Paradox"

This is the paper's most shocking finding. They tested 12 different AI models.

The Finding: The models that got the highest scores on the actual test (the smartest ones) were often the worst teachers.
The Analogy: The "Genius Chef" (like GPT-OSS-120B) made the perfect steak but wrote a recipe that was 10 pages of confusing notes, skipped the crucial "sear the meat" step, and jumped straight to "serve." The apprentice (the weak AI) got lost and failed.
The Twist: Some "average" chefs (weaker models) wrote long, wordy, slightly repetitive recipes that were easy to follow. Their students learned the tricks and got the right answer.

4. The Trade-off: Speed vs. Clarity

The paper found a "Pareto Frontier," which is a fancy way of saying you usually have to choose between two things:

Efficiency: Short, concise, fast thinking. (Good for the AI itself, bad for teaching).
Transfer Utility: Long, detailed, step-by-step thinking that fills in the gaps. (Good for teaching, but "wasteful" for the AI).

The Analogy:

Model A (Efficient): Writes a 3-word note: "Add salt. Done." (Fast, but the apprentice doesn't know how much salt).
Model B (Legible): Writes a 3-page story about why salt is important, how to measure it, and what happens if you forget it. (Slow, but the apprentice learns everything).

5. The "Reward Model" Blind Spot

AI models are trained using a "Reward Model" (a digital judge) that gives them points for getting the right answer.

The Problem: The digital judge is blind to how the answer was reached. It doesn't care if the explanation was confusing. It only cares that the final answer is "42."
The Consequence: Because the judges don't reward "good teaching," the AI models have no reason to learn how to explain themselves clearly. They just learn to be efficient geniuses who mumble.

Why Does This Matter?

This isn't just about AI grades; it's about safety and the future.

Oversight: If we want humans to supervise super-intelligent AI, the AI must be able to explain its logic in a way humans can understand. If the AI is a "mumbling genius," we can't trust it.
Distillation: We want to teach small, cheap AI models to be smart by copying the big ones. If the big ones write "mumbling" notes, the small ones can't learn.
The Future: We need to build AI that doesn't just know the answer, but can teach the answer.

The Bottom Line

The paper concludes that being smart and being a good teacher are two different skills. Currently, our AI is getting very smart but becoming a terrible teacher. To build safe, useful AI for the future, we need to start rewarding the "teaching process" just as much as the final answer. We need AI that can say, "Here is how I got there," in a way that anyone (or any smaller robot) can understand.

1. Problem Statement

Current Reasoning Language Models (RLMs) are increasingly trained to generate long "reasoning traces" (Chain-of-Thought) before outputting a final answer. While these traces are optimized via Reinforcement Learning with Verifiable Rewards (RLVR) to maximize answer correctness, their legibility is largely ignored.

The authors argue that reasoning traces should be viewed not just as intermediate steps to an answer, but as primary artifacts for:

Scalable Oversight: Enabling humans or weaker models to audit and verify the logic of superhuman AI.
Distillation: Transferring reasoning capabilities from strong models to smaller, deployable models.
Caching/Reuse: Reusing atomic reasoning units.

The core problem is that current optimization strategies treat traces as ancillary outputs to be minimized. This leads to a potential paradox where the most accurate models produce traces that are opaque, inefficient, or pedagogically useless for weaker agents. There is no unified framework to measure the "teachability" or legibility of these traces.

2. Methodology: The Legi-Val Framework

The paper introduces Legi-Val, a unified framework that decomposes legibility into two orthogonal dimensions: Efficiency and Transfer Utility.

A. Efficiency Dimension

This measures the "cost" of the trace in terms of cognitive load and verbosity. It includes three metrics:

Length: Token count and step count (normalized by median lengths for correct solutions).
Redundancy: Semantic redundancy measured by computing sentence embeddings (using MiniLM-L6-v2) and calculating the maximum cosine similarity between a step and all preceding steps. A threshold ( $\tau=0.8$ ) identifies repetitive steps.
Backtracking: The frequency of strategy switches or reversals (e.g., "Wait," "But," "Going back"), detected using an LLM judge. High backtracking indicates exploratory reasoning but can impede linear comprehension.

B. Transfer Utility Dimension

This measures the pedagogical effectiveness of a trace by testing if a weaker "student" model can arrive at the correct answer when guided by the trace.

Setup: A strong "teacher" RLM generates a trace. A weaker student model (e.g., Phi-3.5-Mini or LLaMA-3.2-1B) is prompted to continue the reasoning process given progressively longer prefixes of the teacher's trace.
Metrics:
1. First-Order Transfer Utility (FOTU): The area under the curve of student accuracy vs. prefix length. High FOTU means the student learns quickly and stays correct.
2. Second-Order Transfer Utility (SOTU): Measures the uniformity of information density. It calculates the entropy of the distribution of the "first correct step." High SOTU implies information is distributed evenly; low SOTU implies the answer is "front-loaded" or "sandbagged" (buried at the end).
3. Regression Rate (RR): The frequency with which adding more steps causes the student's accuracy to decrease, indicating confusing or contradictory reasoning.

3. Experimental Setup

Models: 12 open-source Reasoning Language Models (e.g., GPT-OSS-120B, DeepSeek-R1, Qwen3, Gemma-3, LLaMA-Nano).
Datasets: Three diverse domains:
- MATH: Competition math problems.
- GPQA: Graduate-level science questions.
- Connections: NYT word puzzle clustering.
Scale: Evaluation of 99,528 reasoning traces.
Student Models: Microsoft Phi-3.5-Mini (3.8B) and Meta LLaMA-3.2-1B.

4. Key Results

The study reveals significant tensions between accuracy, efficiency, and teachability:

The Accuracy-Legibility Paradox: The highest-performing models (e.g., GPT-OSS-120B with 82% accuracy) often rank among the lowest for transfer utility. Conversely, lower-capability models (e.g., LLaMA-Nano) often rank highest in transfer utility.
- Implication: Current optimization strategies produce traces effective for the model itself but pedagogically opaque for others.
No Universal "Legible Reasoner": No single model dominates all metrics. There is a Pareto Frontier between efficiency and transfer utility:
- Efficiency-focused models (e.g., Gemma-3) produce concise, low-redundancy traces but struggle to scaffold weaker models.
- Transfer-focused models (e.g., DeepSeek-R1) produce verbose, redundant traces that effectively teach students but are inefficient.
Reward Models Ignore Legibility: Scalar rewards from standard Reward Models (RMs) correlate with transfer utility only when driven by accuracy. When isolating correct traces, the correlation between RM scores and legibility metrics vanishes ( $\rho \approx 0.05$ ). RMs do not intrinsically reward legibility.
Domain Specificity: Legibility is task-dependent. Models that perform well on STEM tasks (MATH/GPQA) do not necessarily perform well on word puzzles (Connections), suggesting different pedagogical requirements for different reasoning domains.
Trade-offs: There is a negative correlation between First-Order and Second-Order Transfer Utility ( $\rho = -0.68$ ), indicating that models that get students to the answer quickly often do so by front-loading the answer rather than distributing information evenly.

5. Key Contributions

Legi-Val Framework: A novel, scalable framework for quantifying reasoning trace legibility, moving beyond simple token counts to include pedagogical transfer metrics.
Empirical Evidence of the Paradox: Demonstrating that state-of-the-art accuracy does not guarantee legibility, and that current RLVR training fails to optimize for multi-agent utility.
Identification of Trade-offs: Mapping the Pareto frontier between efficiency and transfer utility, showing that optimizing for one often degrades the other.
Critique of Reward Models: Showing that current reward models are blind to the intrinsic value of legibility, treating traces merely as instrumental means to an end.

6. Significance and Future Directions

This paper fundamentally shifts the perspective on reasoning traces from "internal scratchpads" to "communication artifacts."

Safety & Oversight: For scalable oversight, AI systems must produce traces that weak verifiers (humans or small models) can understand. Current models fail this test.
Distillation: To effectively distill reasoning capabilities into smaller models, the "teacher" must produce traces with high transfer utility, not just high accuracy.
Future Training: The authors propose incorporating Legi-Val metrics directly into RL training (e.g., via multi-objective rewards or using student models as part of the rollout process) to explicitly optimize for legibility.
Multi-Agent Future: As AI systems scale, the flow of information between models (and between models and humans) becomes the bottleneck. Legi-Val provides the metrics necessary to measure and optimize this flow.

In conclusion, the paper argues that reasoning traces are primary artifacts that require their own optimization objectives, distinct from final answer correctness, to enable safe, interpretable, and transferable AI systems.