🤖 machine learning

EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample

EVA-0 is a minimal zeroth-order adaptation framework that enables efficient test-time model evolution with only two forward passes per sample, overcoming key optimization obstacles without backpropagation to achieve superior performance and significant speed-ups over existing methods.

Original authors: Guohao Chen, Shuaicheng Niu, Geng Li, Yunbei Zhang, Shilin Shan, Chunyan Miao, Jianfei Yang

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Guohao Chen, Shuaicheng Niu, Geng Li, Yunbei Zhang, Shilin Shan, Chunyan Miao, Jianfei Yang

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 highly trained AI assistant, like a smart navigation app, that was taught to drive perfectly on sunny, clear highways. But now, you've deployed it into a chaotic city where it suddenly starts raining, the roads are covered in snow, and the signs are blurry.

Most AI systems are "frozen" once they leave the lab. They can't learn from these new, messy conditions. To fix them, engineers usually have to recall the car, retrain it in a massive garage (using backpropagation, which is like a heavy, expensive, memory-hogging process), and then send it back out. This is slow, expensive, and impossible for small devices like a smartwatch or a drone.

Some newer methods try to let the AI learn on the fly without that heavy retraining. But they often require the AI to "think" about the same problem dozens of times to figure out the right direction, which is still too slow for real-time use.

Enter EVA-0: The "Two-Guess" Genius

The authors of this paper propose a new method called EVA-0. Their big idea is simple but powerful: Can an AI learn and adapt effectively by looking at a problem only twice?

They say "Yes." Here is how they made it work, using some creative analogies:

1. The Problem: The "Shortcut" Trap

Imagine you are trying to lose weight (improve the AI).

The Shortcut: You could just stop eating entirely. The scale (the loss function) will show you are losing weight immediately, but you aren't actually getting healthier; you're just starving. In AI terms, the model finds a "shortcut" where it fakes a good answer by making its confidence numbers huge, rather than actually understanding the image.
The EVA-0 Fix: The authors designed a special "diet plan" (a scale-invariant objective) that prevents the AI from cheating. It forces the AI to actually learn the right features, not just inflate its confidence numbers to look good.

2. The Problem: Getting Lost (Weight Drift)

Imagine you are hiking in a foggy forest. If you take random steps to find the path, you might accidentally wander miles away from the trail. In AI, taking random "guesses" to improve can cause the model to drift so far from its original training that it forgets everything it knew.

The EVA-0 Fix: They introduced an Anchor. Think of this as a sturdy rope tied to a tree back at the trailhead. The AI is allowed to wander and explore to find the new path, but the rope gently pulls it back if it starts to drift too far. This keeps the AI stable while still allowing it to learn.

3. The Problem: Guessing the Direction

Usually, to know which way to go, you might try walking forward, see if it's better, then walk backward, and compare.

The Old Way: Some methods use a "one-sided" guess (just walk forward and hope). This is noisy and unreliable.
The EVA-0 Trick: They use a Symmetric Two-Sided approach. They ask the AI to look at the problem twice: once with a tiny "nudge" to the left, and once with a tiny "nudge" to the right.
- The Magic: By averaging the results of these two nudges, they get a very clear picture of what the answer would have been without any nudges (so they can make a prediction).
- The Efficiency: By comparing the difference between the left-nudge and right-nudge, they figure out exactly which way to move to improve.
- The Result: They get both a prediction and a learning signal from the exact same two looks. No extra steps needed.

Why This Matters

The paper tested this on images that were corrupted (blurry, noisy, snowy) using two types of AI models (ViT and ResNet).

Speed: It is 14 times faster than previous "no-backpropagation" methods because it doesn't need to guess 28 times like the old methods; it only guesses twice.
Memory: It uses a tiny fraction of the memory (about 822MB vs. 5GB for other methods), meaning it can run on small, edge devices like phones or sensors.
Performance: It actually works better than some methods that do use the heavy, expensive backpropagation, and it is the first method to do this without needing any "source data" (it learns purely from the messy new world).

In a nutshell: EVA-0 is a way to let an AI learn from its mistakes in real-time, using only two quick glances at the problem, without getting lost, without cheating, and without needing a supercomputer to do it. It turns a "frozen" model into a "living" one that can adapt to the real world on the fly.

Technical Summary: EVA-0

Problem Statement

Modern AI models are increasingly deployed in dynamic environments subject to distribution shifts (e.g., sensor variations, changing user behavior). While Test-Time Adaptation (TTA) aims to enable models to learn from unlabeled test-time experience, most existing methods rely on backpropagation (BP). BP incurs substantial memory overhead, making it incompatible with resource-constrained edge devices, quantized models, specialized accelerators, or black-box model settings.

Recent efforts have explored BP-free TTA using zeroth-order (ZO) optimization. However, existing ZO methods face significant limitations:

Efficiency vs. Performance: Methods like FOA require many forward passes (e.g., 28) for stable signals, while methods like ZOA, which operate under a strict two-forward budget, suffer from coarse and noisy update directions.
Optimization Challenges: ZO optimization is prone to learning "shortcut solutions" (e.g., reducing entropy via logit-norm inflation rather than improving decision boundaries), suffers from uncontrolled weight drift, and struggles with effective update direction estimation under tight computational constraints.

The core question addressed is: Can a deployed model stably and effectively evolve from unlabeled test-time experience using only two forward passes per sample without backpropagation?

Methodology: EVA-0

The authors propose EVA-0 (Evolutionary Adaptation with Zeroth-order gradients), a minimal framework designed to achieve stable test-time model evolution under a strict two-forward budget. EVA-0 addresses the identified challenges through three core design principles:

1. Sample-wise Symmetric Two-Sided Perturbation (SSD)

To overcome the noise and inefficiency of one-sided ZO estimators while adhering to the two-forward constraint:

Mechanism: For each sample $x_i$ , the model evaluates two symmetric perturbations: $f_{\theta_t + \mu z_{t,i}}(x_i)$ and $f_{\theta_t - \mu z_{t,i}}(x_i)$ .
Dual Utility: These two forward passes serve a dual purpose:
- Inference: The outputs are averaged to approximate the clean model prediction (canceling first-order perturbation terms via Taylor expansion), eliminating the need for a third "clean" forward pass.
- Adaptation: The difference in losses between the two symmetric outputs provides a lower-variance, two-sided SPSA gradient estimate.
Sample-wise Independence: Unlike batch-level perturbations, EVA-0 applies independent perturbations to each sample within a mini-batch. This increases the diversity of explored update directions, significantly improving alignment with true BP gradients compared to shared batch perturbations.

2. Shortcut-Resistant Objectives

To prevent ZO optimization from exploiting shortcuts (e.g., inflating logit norms to minimize entropy without learning):

Scale-Invariant Entropy: EVA-0 employs a scale-invariant entropy objective ( $E_{SR}$ ). It normalizes the logits of perturbed outputs by their average norm and applies decentering (subtracting an online center). This prevents the model from reducing entropy simply by inflating the magnitude of logits.
Sample-wise Feature Alignment: When source statistics are available, EVA-0 uses a sample-wise feature alignment loss ( $\ell_{SWA}$ ). Instead of estimating variance from a single noisy sample, it tracks first and second moments of features online via Exponential Moving Average (EMA) to reconstruct a stable variance estimate for alignment.

3. Anchor-Guided Trajectory Control

To mitigate the accumulation of weight drift common in random ZO updates:

Anchor-Guided Perturbation: A reliable anchor (e.g., pretrained weights or a slowly moving EMA of online parameters) is maintained. The perturbation distribution $D_t$ is adaptive: a subset of samples explores random directions, while others are guided by the vector pointing toward the anchor.
Weight Relaxation: After the ZO update, a weak relaxation step blends the updated weights with the anchor ( $\theta_{t+1} = (1-\gamma)\theta'_{t+1} + \gamma\theta_{anc}$ ). This acts as a soft bound, denoising the trajectory and preventing the model from drifting too far into unstable regions while retaining plasticity for adaptation.

Key Contributions

Strict Two-Forward Setting: The paper establishes and investigates the challenging regime of test-time model evolution with only two forward passes per sample and no backpropagation, pushing TTA toward highly efficient real-world deployment.
Systematic Analysis of ZO Limitations: The authors reveal three specific obstacles in zeroth-order TTA: susceptibility to shortcut solutions, uncontrolled weight drift, and ineffective update direction estimation under strict budgets.
EVA-0 Framework: A unified method integrating scale-invariant objectives, anchor-guided trajectory control, and sample-wise symmetric perturbations.
Comprehensive Validation: Extensive experiments across CNNs (ResNet50-GN), Vision Transformers (ViT-Base), quantized models (6-bit/8-bit), continual streams, and black-box settings.

Experimental Results

Evaluated on ImageNet-C (15 corruptions, severity level 5):

Performance: EVA-0 outperforms both BP-based methods (e.g., DeYO) and BP-free baselines (FOA, ZOA, FOZO). On ViT-Base, EVA-0 achieves 66.7% average accuracy, surpassing DeYO (65.7%) and ZOA (61.3%).
Efficiency: EVA-0 achieves a 14× speed-up over FOA (which requires 28 forward passes) and reduces memory usage significantly compared to BP-based methods (822MB vs. 5,350MB for DeYO on ViT-Base).
Quantized Models: EVA-0 enables a 6-bit quantized ViT to outperform a non-adapted 32-bit ViT by 4.4% in OOD accuracy, demonstrating its viability for edge deployment.
Continual Adaptation: In long-term continual TTA settings where models evolve across domains without resetting, EVA-0 maintains stability (45.6% accuracy on ResNet50-GN), whereas other methods (e.g., FOA, DeYO) suffer significant degradation.
Black-Box Adaptation: EVA-0 effectively adapts input prompts in black-box settings, outperforming FOA by 5.3% under the two-forward constraint.

Significance and Claims

The paper claims that forward-only model evolution is not merely a compromise for restricted scenarios but a promising and scalable path toward adaptive deployed intelligence.

By unifying shortcut-resistant objectives, trajectory control, and efficient direction estimation, EVA-0 demonstrates that effective, stable, and efficient model evolution is possible under extremely restrictive deployment constraints (two forward passes, no BP). This work suggests that models can meaningfully learn from their own unlabeled experience in real-world, resource-limited environments without the heavy computational cost of backpropagation.