⚡ electrical engineering

eCP: Equivariant Conformal Prediction with pre-trained models

This paper proposes eCP, a method that enhances conformal prediction for long-horizon uncertainty quantification by integrating geometric symmetry through group-averaging of pre-trained models, thereby provably reducing non-conformity scores and yielding sharper prediction sets, particularly at high confidence levels.

Original authors: Nikolaos Bousias, Lars Lindemann, George Pappas

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nikolaos Bousias, Lars Lindemann, George Pappas

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 weather forecaster trying to predict the path of a storm. You know the storm is moving, but you aren't 100% sure exactly where it will go tomorrow. To be safe, you draw a giant circle on the map saying, "The storm will definitely be somewhere inside this huge area."

This is a bit like Conformal Prediction (CP). It's a mathematical tool used by AI to say, "I'm 95% sure the answer is in this box." The problem? As the AI tries to predict further into the future (like predicting a storm 5 days out instead of 1), that "safe box" gets so huge it becomes useless. It might say, "The storm could be anywhere in the entire country!" That's technically correct, but not very helpful for planning.

This paper introduces a new trick called eCP (Equivariant Conformal Prediction) to shrink that box without losing its safety guarantee. Here is how it works, using some everyday analogies.

The Problem: The "Confused" AI

Imagine you are teaching a robot to recognize a chair. You show it a chair, then you show it the same chair turned sideways, then upside down.

Standard AI: If you ask the robot, "Is this a chair?" it might say "Yes" for the upright one, but get confused and say "Maybe" for the sideways one, even though it's the exact same object.
The Result: When the robot tries to predict where a pedestrian (a person walking) will go, it gets confused by the angle of the camera or the direction the person is facing. This confusion makes its "safety box" (the uncertainty area) grow unnecessarily large.

The Solution: The "Group Hug" (Symmetry)

The authors realized that many real-world problems have symmetry. A chair is still a chair whether it's rotated 90 degrees or flipped. A person walking north is doing the same thing as a person walking south, just in a different direction.

eCP uses a clever trick called "Group Averaging."

Think of it like this:

The Old Way: The AI looks at a pedestrian walking North and makes a prediction. It also looks at a pedestrian walking East and makes a different prediction. Because the AI is slightly confused by the angle, these two predictions are slightly different, creating a wide, messy cloud of uncertainty.
The eCP Way: Before making the final prediction, eCP says, "Wait a minute. Let's pretend this person is walking in every possible direction at once."
- It takes the AI's prediction.
- It rotates the prediction 90 degrees, 180 degrees, and so on.
- It takes the average of all those rotated predictions.

The Magic Analogy: The "Blur" vs. The "Focus"

Imagine you are trying to take a photo of a spinning fan.

Without eCP: You take a photo, and it's blurry because the fan is moving. You draw a huge circle around the blur to say, "The fan blades are definitely somewhere in here."
With eCP: You realize the fan is perfectly symmetrical. So, instead of looking at one blurry snapshot, you take 10 snapshots of the fan spinning, rotate them so they all line up, and average them out. The blur cancels itself out, and you get a sharp, clear image of the fan's center.

By "averaging out" the confusion caused by angles and directions, eCP removes the "noise" from the AI's uncertainty.

Why This Matters

The paper proves two amazing things:

It's Safer: The prediction box still guarantees that the real answer is inside it (just like the old method).
It's Smaller: Because the AI is less confused by angles, the box shrinks dramatically. In their experiments with predicting where pedestrians will walk, the "safety box" got 28% smaller while staying just as safe.

The Bottom Line

eCP is like giving a super-smart AI a pair of glasses that helps it see the "shape" of the world, regardless of which way things are facing. It takes a method that was already good at being safe but too conservative (making huge boxes), and makes it precise without sacrificing that safety.

It's a "post-hoc" fix, meaning you don't have to retrain the AI or build a new robot. You just take the AI you already have, give it this "symmetry glasses," and suddenly, its predictions become much sharper and more useful for real-world tasks like self-driving cars or robotics.

1. Problem Statement

Conformal Prediction (CP) is a widely used framework for uncertainty quantification that provides finite-sample, distribution-free coverage guarantees under the assumption of data exchangeability. However, a critical limitation of CP is efficiency: as the confidence level increases (e.g., 99%) or the prediction horizon extends (e.g., long-horizon trajectory forecasting), the resulting prediction sets (uncertainty regions) often become excessively large, rendering them practically uninformative.

Current CP methods treat each data sample in isolation. They fail to leverage geometric symmetries (such as rotation, translation, or reflection) inherent in many real-world problems (e.g., pedestrian movement, robotics). Consequently, symmetry-equivalent inputs may receive different non-conformity scores simply due to arbitrary coordinate frame choices, leading to inflated uncertainty estimates.

The paper addresses the challenge of improving the efficiency (tightness) of conformal prediction sets without sacrificing coverage guarantees or retraining the underlying predictive model.

2. Methodology: Equivariantized Conformal Prediction (eCP)

The authors propose eCP, a post-hoc procedure that infuses geometric information into standard CP by leveraging group symmetries. The core mechanism involves group-averaging the non-conformity scores of a pre-trained predictor.

Key Technical Components:

Group Symmetry & Orbits: The method assumes the data and the prediction task exhibit invariance under a symmetry group $G$ (e.g., planar rotations $SO(2)$). A sample $(x, y)$ is treated not as an isolated point but as a representative of its orbit $O(x, y) = \{(\phi_g(x), \psi_g(y)) \mid g \in G\}$ .
Symmetrization Operator ( $\Pi_G$ ): Instead of using the raw non-conformity score $s(f_\theta(x), y)$ , eCP defines a symmetrized score by averaging over the group:
$\Pi_G[s; f](x,y) := \mathbb{E}_{g \sim \mu_G} \left[ s\left( f_\theta(\phi_{g^{-1}}(x)), \psi_{g^{-1}}(y) \right) \right]$
where $\mu_G$ is the Haar measure on the group $G$ .
Equivariantized Predictor: The method effectively constructs a new predictor $f^G_\theta$ by averaging the predictions of the pre-trained model over the group orbit:
$f^G_\theta(x) := \mathbb{E}_{g \sim \mu_G} \left[ \psi_g \circ f_\theta \circ \phi_{g^{-1}}(x) \right]$
This projection forces the predictor to respect the symmetry structure, even if the original model $f_\theta$ was only approximately equivariant.
Post-Hoc Application: Crucially, this process requires no retraining or architectural modification of the base model. It operates entirely on the calibration set and the test input using the pre-trained model.

3. Theoretical Contributions

The paper provides rigorous theoretical proofs demonstrating that eCP yields tighter prediction sets while maintaining validity.

Validity (Coverage Guarantee): Under the assumption of $G$ -invariant exchangeability, the paper proves (Theorem 1) that the equivariantized conformal prediction set maintains the standard finite-sample coverage guarantee:
$P(Y_{test} \in C^{(G)}_{1-\alpha}(X_{test})) \geq 1-\alpha$
Stochastic Ordering (Contraction): The central theoretical result (Theorem 2) establishes that the non-conformity scores of the equivariantized model are contracted in increasing convex order ( $S_{f^G} \preceq_{icx} S_f$ $S_{f^{G}} ⪯_{i c x} S_{f}$ ) relative to the original model.
- This implies that for any increasing convex function $\nu$ , $\mathbb{E}[\nu(S_{f^G})] \leq \mathbb{E}[\nu(S_f)]$ .
- Specifically, this leads to reduced variance and improved tail behavior.
Tail Bounds & Quantile Contraction:
- CVaR Contraction: The Conditional Value at Risk (CVaR) of the equivariantized scores is strictly lower (Theorem 3), implying tighter bounds on extreme quantiles.
- Chernoff & Hoeffding Bounds: The paper demonstrates in the appendices that eCP improves exponential tail bounds (Chernoff) and concentration inequalities (Hoeffding/Bernstein), leading to faster convergence of empirical quantiles.
- Volume Reduction: The expected volume of the conformal prediction set is proven to shrink: $\mathbb{E}[\text{Vol}(C^{(G)}_{1-\alpha})] \leq \mathbb{E}[\text{Vol}(C_{1-\alpha})]$ .

4. Experimental Results

The authors evaluated eCP on pedestrian trajectory prediction, a domain characterized by strong geometric symmetries and long-horizon uncertainty accumulation.

Datasets: ETH–UCY, Stanford Drone Dataset (SDD), and SportVU NBA movement data.
Base Models: Two distinct pre-trained predictors were used:
1. SocialVAE: A stochastic, multimodal latent-variable model.
2. TUTR: A deterministic transformer-based model.
Symmetry Groups: Experiments tested groups of increasing richness: $C_4$ (90° rotations), $C_8$ (45° rotations), and $SO(2)$ (continuous planar rotation).
Key Findings:
- Significant Tightening: eCP consistently reduced the calibration quantile ( $Q_{0.05}$ , the 95% conformal radius) by approximately 28% compared to standard CP.
- High Confidence Gains: The improvements were most pronounced at high confidence levels (e.g., 99%), where standard CP is typically overly conservative.
- Coverage Preservation: Despite the significant reduction in set size, empirical coverage remained close to the nominal target (e.g., ~95% or ~99%), validating the theoretical guarantees.
- Group Richness: Performance improved monotonically with the richness of the symmetry group ( $SO(2) > C_8 > C_4$ ), confirming that stronger symmetry exploitation yields greater uncertainty reduction.
- Model Agnosticism: The method worked equally well for both stochastic (SocialVAE) and deterministic (TUTR) models.

5. Significance and Conclusion

eCP bridges the gap between geometric inductive bias and post-hoc uncertainty quantification.

Practical Impact: It offers a principled mechanism to make uncertainty estimates in robotics and autonomous systems significantly more actionable without the computational cost of retraining complex models.
Theoretical Insight: It demonstrates that "approximate" or "emergent" symmetry in pre-trained models can be formally exploited to tighten uncertainty bounds, even if the model is not perfectly equivariant.
Limitations: The method requires knowledge of the relevant symmetry group. If the assumed group does not match the data's true invariances, gains may be limited. Additionally, continuous groups (like $SO(2)$) require Monte Carlo approximation, introducing a modest computational overhead.

In summary, eCP transforms the "waste" of symmetry-equivalent information in standard CP into a resource for shrinking uncertainty regions, providing sharper, more reliable predictions for safety-critical, long-horizon tasks.