🤖 machine learning

Mirror Horizon: Viable Path Entropy as a Measure of Bounded Reflection

This paper introduces Mirror Theory and its operational measure, Viable Path Entropy (VPE), to define intelligent capability as the structure of verified, diverse continuations reachable under bounded reflection, demonstrating through language model experiments that this metric captures accessible reasoning capacity more effectively than parameter count or one-shot accuracy.

Original authors: Tiantian Zhang (Crystal)

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tiantian Zhang (Crystal)

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 magical mirror that doesn't just show your face, but shows every possible version of your future self that could exist if you kept talking, thinking, or solving a problem. Most people look at a mirror and ask, "Did I get the answer right?" This paper asks a much more interesting question: "How many different right answers can this mirror find, and how many of those answers actually survive the test?"

The authors call this idea Mirror Theory. Instead of treating an AI like a static encyclopedia that just stores facts, they treat it like a living mirror that has to survive a process called "reflection." Think of reflection not as a single flash of light, but as a long, winding path where the AI tries out many different ways to solve a problem.

The Main Discovery: Bigger Isn't Always Better

The paper's biggest surprise is that bigger isn't always better. In the world of AI, we usually assume that a model with more "brain power" (parameters) is automatically superior. But the authors tested this using a specific setup: they asked three different versions of an AI (named Qwen2.5) to solve 30 math problems. They gave them a limited amount of "thinking time" (measured in tokens, which are like word-pieces).

Here is what they found:

When they gave the models a short thinking time (96 tokens), the smaller models struggled.
When they increased the thinking time to 160 tokens, something cool happened. The 1.5-billion-parameter model (the middle-sized one) became the champion. It found more correct answers, and more importantly, it found them in more different ways than the others.
The 3-billion-parameter model (the biggest one) actually did worse than the middle-sized one in this specific test. It got some answers right, but it got stuck in a rut, finding the same few solutions over and over, while the middle model explored a wider variety of successful paths.

The paper suggests that "capability" isn't just about how big the model is; it's about how many viable paths (correct, diverse solutions) it can actually reach within a limited budget of time and energy.

The "Viable Path Entropy" Scorecard

To measure this, the authors invented a new score called Viable Path Entropy (VPE). Imagine you are a judge at a talent show.

Old Way (Pass@k): You just check if at least one contestant got a perfect score. If yes, you give them a gold star. If no, they get nothing.
New Way (VPE): You check two things:
1. Reachability: Did they find any correct solution at all?
2. Diversity: If they did, how many different types of correct solutions did they find? Did they solve it using a clever shortcut, a long calculation, and a visual diagram? Or did they just guess the same answer three times?

The paper shows that the middle-sized model (1.5B) had the highest VPE score. It didn't just get more answers right; it got them right in more creative, distinct ways. The biggest model (3B) had a lower score because, even though it was smart, it was less "exploratory" under these specific rules.

What This Paper Rules Out

The authors are very clear about what this is not.

It is not a rule that says "bigger models always win." The experiment explicitly shows that a larger model can have a smaller "mirror horizon" (less accessible success) than a smaller one if the conditions aren't right.
It is not a magic formula that predicts exactly how AI will scale forever. The authors argue that there is no single, simple line that says "more parameters = more capability." Instead, capability depends on the specific rules of the game (the prompt, the time limit, the verifier).
It is not just about getting the right answer once. The paper argues that getting the right answer in only one rigid way is less impressive than finding many different valid ways to get there.

How Sure Are They?

The authors are careful not to claim they have solved the mystery of AI forever. They describe their results as measured and observed in a specific experiment, not as a universal law of the universe.

They ran 32 samples (attempts) for each of the 30 math problems.
They used a specific "verifier" (a strict checker that looks for the right number) and a specific "mode map" (a way to group similar solutions).
They admit that their "mode map" is a bit rough (like sorting solutions by length or simple math symbols) and that future tests might use more complex ways to group answers.
They suggest that their findings support the idea of Mirror Theory, but they don't claim to have proven it beyond all doubt. They say their results "show that" the measure works and "supports the claim," but they leave the door open for more testing with different tasks and models.

The Takeaway

Think of the AI not as a robot that memorizes a textbook, but as an explorer with a limited supply of fuel. The paper suggests that the best explorer isn't necessarily the biggest one; it's the one that can use its fuel to find the most distinct, correct paths through the forest. When you give the explorer more fuel (increasing the token budget from 96 to 160), the middle-sized explorer suddenly becomes the most capable, finding a rich variety of solutions that the bigger, heavier explorer missed.

The paper concludes that to truly understand an AI's intelligence, we shouldn't just ask, "Did it get the answer?" We should ask, "How many different, valid ways could it have found the answer, and how many of those did it actually discover?" That is the new "mirror horizon."

Technical Summary: Mirror Horizon and Viable Path Entropy

Problem Statement
Current evaluation metrics for large language models (LLMs)—such as loss, accuracy, pass@k, and benchmark scores—collapse a system's internal capacity into single outcome statistics. These metrics fail to distinguish between a model that solves a problem once and fails repeatedly, a model that solves fewer problems but reveals diverse solution strategies, and a model with high raw output entropy but little verified structure. The paper argues that capability should not be defined merely by one-shot success or parameter count, but by the structure of viable continuations a system can sustain under repeated reflection.

Methodology
The paper operationalizes "Mirror Theory" through a new measure called Viable Path Entropy (VPE), defined as a finite-budget measure of verified continuation capacity.

Theoretical Scaffold:
- Mirror as Continuation Law: A mirror is defined not as a static representation (e.g., a hidden vector) but as the internal structure that induces a law of viable continuations.
- Components: The theory integrates intuition (local constraints from evidence), taste (selection pressure preserving valued invariants), and geometry (learned structure ensuring stable selection).
- Constructive Compatibility: A possibility is "constructively compatible" if it increases the finite-budget viable continuation capacity without rupturing invariants.
The Measure (VPE):
- Given a budget $B$ (rollouts), horizon $T$ (tokens), a verifier $V$ , and a mode map $\psi$ , VPE is defined as:
  $H_{B,T}(M) = \log \Pr[V=1] + H(\psi \mid V=1)$
- Term 1 (Verified Reachability): The probability of reaching a viable continuation ( $\Pr[V=1]$ ).
- Term 2 (Verified-Mode Diversity): The entropy of the semantic modes reached among successful rollouts.
- This decomposition separates breadth (can the model solve the problem?) from depth/diversity (how many distinct, verified ways can it solve it?).
Experimental Instantiation:
- Task: GSM8K math reasoning problems.
- Models: Qwen2.5-Instruct family (0.5B, 1.5B, and 3B parameters).
- Protocol: 30 problems, 32 sampled rollouts per problem, temperature 0.8, top-p 0.95.
- Variables: Two reflection horizons were tested: 96 tokens and 160 tokens.
- Verifier: A strict regex-based extractor checking for numeric correctness against gold answers.
- Mode Map: A heuristic function grouping verified solutions by length bin, operator signature, equation count, and reasoning style.

Key Contributions

Formal Path-Space Theory: A version of Mirror Theory where mirrors are defined as continuation laws rather than static states.
Viable Path Entropy (VPE): A definition of mirror horizon that decomposes bounded capability into verified reachability and verified-mode diversity.
Theoretical Connection: Linking abstract concepts (intuition, taste, geometry, constructive compatibility) to the measurable VPE object.
Empirical Validation: Experiments demonstrating that reflection budget expands verified continuation capacity and that mirror horizon is distinct from parameter count.

Results

Budget Expansion: Increasing the token horizon from 96 to 160 substantially improved all metrics for the 0.5B and 1.5B models. Verified probability rose, zero-verified fractions dropped, and verified-mode entropy increased.
Model Performance at 160 Tokens:
- Qwen2.5-1.5B achieved the strongest mirror horizon across all tested models, dominating in verified probability, pass@32, zero-verified fraction, and verified-mode entropy.
- Qwen2.5-3B did not outperform the 1.5B model. While it had a higher average verified probability than the 0.5B model, it exhibited lower breadth (lower pass@32) and significantly lower verified-mode entropy than the 1.5B model.
Coverage Curves: Empirical coverage curves showed that the 1.5B model reached a higher expected number of distinct verified modes across all rollout budgets ( $b \in \{1, \dots, 32\}$ ), confirming that its "mirror horizon" unfolds more effectively with increased budget.

Significance and Claims
The paper claims to provide a measure-level account of capability, arguing that a system's strength lies in the structure of viable continuations it can make reachable under a bounded reflection protocol, rather than in raw parameter count or single-answer accuracy.

Protocol-Conditioned Horizon: The results demonstrate that mirror horizon is not an intrinsic scalar of a model but depends on the protocol (budget, verifier, mode map). A larger model (3B) can have a smaller accessible mirror horizon than a smaller model (1.5B) under specific constraints.
Beyond Pass@k: VPE reveals distinctions hidden by pass@k. A model might have high pass@k but low mode diversity (solving problems in a single brittle way), whereas VPE captures the robustness of having multiple verified solution modes.
Operationalizing Reflection: The experiment validates the theoretical claim that increasing the reflection horizon (token budget) allows the internal structure of the mirror to unfold, revealing more viable continuations.

The paper remains modest, acknowledging that the current mode map is coarse and the experiments are a pilot. It does not claim a universal scaling law but establishes that the verified continuation law is a useful, measurable object for studying bounded reflection in intelligent systems.