ETS: Energy-Guided Test-Time Scaling

• Energy-Guided Test-Time Scaling (ETS) is a family of training-free, inference-time strategies that leverage energy-based principles to improve model robustness, calibration, alignment, and compute efficiency.
• ETS applies transformations, scaling, and reward-guided sampling at test time to reduce adversarial risk and adapt pre-trained models to new data distributions without retraining.
• By utilizing methods like projected gradient descent and contrastive adaptation, ETS achieves improved accuracy and efficiency, with empirical gains such as up to an 11% robust accuracy increase.

Energy-Guided Test-Time Scaling (ETS) refers to a family of training-free, inference-time methods grounded in energy-based principles designed to enhance the robustness, generalization, calibration, or alignment of modern neural models. Across vision, language, and multimodal domains, ETS leverages the notion of “energy” as an implicit confidence or reward surrogate, using this signal to guide either transformations (defensive input adaptation), scaling (adaptive parameter re-tuning), or sampling (test-time RL-style guidance) to the benefit of the downstream task without re-training model weights.

1. Energy-Based Model Foundations and ETS Paradigm

Energy-based modeling underpins ETS. In most deep classifiers and sequence models, output logits $f_\theta(x)$ implicitly define an energy function, often via the log-partition or log-sum-exp (LSE) operator. For a $K$-way classifier, the energy of sample $x$ is $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$ where low energy is associated with naturally occurring (on-manifold) data, while off-manifold or adversarial examples have higher energy (Mirza et al., 27 Mar 2026, Yuan et al., 2023). This framework enables a principled basis for adaptation and calibration at test time—without access to training data or the need for model retraining—by aligning the empirical distribution of test samples with the energy landscape of the pre-trained model.

2. ETS for Robustness: Energy Minimization and Input Purification

In the context of adversarial robustness, ETS—exemplified by ET³ (“Energy-Guided Test-Time Transformation”)—acts as a plug-and-play inference-time defense that systematically reduces the energy of input samples subject to a norm-bounded perturbation budget (Mirza et al., 27 Mar 2026). Formally, given an input $x$ and classifier $f_\theta$, ETS solves:

$$x' = \arg\min_{\tilde{x}: \|\tilde{x} - x\| \le \epsilon} E(\tilde{x})$$

where $E(\tilde{x})$ is as above, and $\epsilon$ is the defense budget controlling the allowed input shift. This problem is efficiently approached via projected gradient descent:

• $x^{(0)} = x$
• $K$0 with $K$1 as the norm-ball projection, step size $K$2, and steps $K$3 (typically $K$4–$K$5).

Theoretical guarantees (for $K$6 classes, under local linearity and gradient-dominance) show that one step of ETS suffices to recover the correct class label for a purified input, provided the energy and logit gradients satisfy certain conditions. These properties extend to multi-step and multi-class settings by induction.

3. ETS for Alignment: Test-Time Reward-Guided Sampling

ETS can be extended to reinforcement learning alignment, enabling test-time direct sampling from an RL-optimal (“Boltzmann”) policy for language or sequential models (Li et al., 29 Jan 2026). Here, the energy functional is $K$7 with $K$8 the reward and $K$9 the temperature regularizer. The ETS backward-sampling algorithm uses online Monte Carlo to estimate this energy, applying importance sampling accelerations and batching to avoid the cost of iterative RL-based re-training. Resultant sampling quality converges to the target policy in total variation distance at rate $x$0 in the number of samples per guidance step.

ETS achieves higher pass@1 accuracy (Table 1 in (Li et al., 29 Jan 2026)) than both best-of-N and beam search, and outperforms RLHF-trained competitors in domains where reward proxies (e.g., self-consistency voting) are effective.

4. ETS as Test-Time Adaptation: Energy-Guided Recali­bration

Test-time energy adaptation (TEA) (Yuan et al., 2023) frames ETS as a means of re-calibrating fixed pre-trained models to new data distributions via adjustment of a small set of scaling parameters (e.g., normalization-layer gains/biases). Here, adaptation is performed by contrastive-divergence, moving the scaling parameters $x$1 to minimize the energy of authentic test samples while raising the energy of negative samples synthesized via short-run Langevin dynamics:

• $x$2 updated via $x$3

This drives alignment of the model’s marginal $x$4 to the test data’s marginal, closing the covariate-shift gap and improving calibration (reducing ECE/MCE) and generalization error.

5. ETS for Resource Allocation: Test-Time Compute Scaling

Energy-Guided Test-Time Scaling is also synonymous with dynamic inference-time compute allocation—Test-Time Compute (TTC)—particularly in LLMs (2505.14733). Here, ETS strategies consist of:

• Majority Vote (MV): Multiple independent samples (K), each requiring a full forward pass, with post hoc aggregation.
• Reasoning Tokens (RT): Allowing the model to extend its output (decode) length to trigger self-refinement processes.

Both strategies deliver substantial accuracy improvements on reasoning tasks (math, code), often exceeding those achievable by simply increasing model size. Energy–accuracy trade-off analyses show that on many tasks, computing longer at test time yields higher accuracy per Joule spent versus scaling parameter count. For example, on Math500, RT enables a $x$5 accuracy gain on a $x$6 B LLM at similar energy, whereas parameter scaling yields $x$7 for the same energy change.

Dynamic difficulty prediction and early exit heuristics (length-based) enable further efficiency: the model adaptively spends more energy on harder queries, reducing total deployment cost.

6. Practical Usage, Hyperparameters, and Limitations

Key hyperparameters of ETS methods are domain-specific. For adversarial robustness, $x$8–$x$9, $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$0–$$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$1, and step size $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$2 yield robust gains with only $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$3–$$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$4 compute overhead (Mirza et al., 27 Mar 2026). In RL alignment, $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$5 (candidates), $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$6 (MC samples), and $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$7 (guidance steps) control the latency/quality tradeoff, with $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$8, $$E_\theta(x) = -\log \sum_{k=1}^K \exp(f_\theta(x)_k)$$9, $x$0 or $x$1 recommended (Li et al., 29 Jan 2026). For compute allocation, a balance between batch size (to amortize prefill cost) and decode length (to avoid bandwidth saturation) is crucial (2505.14733).

Limitations include:

• Overhead from additional forward/backward passes (vision/language encoder),
• Dependence on energy landscape smoothness and local linearity assumptions,
• No guarantees for universal consistency in TEA or for noisy/ambiguous reward proxies in RL alignment,
• Hardware or architectural limitations (e.g., GPU/TPU memory BW bottlenecks).

7. Empirical Impact and Future Prospects

Across domains, ETS consistently improves robustness, alignment, and efficiency. In adversarially robust classification, ETS outperforms prior test-time transformation and prompt-based methods by up to $x$2 mean robust accuracy gain (Mirza et al., 27 Mar 2026). In RL alignment, ETS matches or exceeds RLHF-style methods in sample quality and accuracy, with faster inference (Li et al., 29 Jan 2026). In LLM deployment, TTC/ETS yields substantial per-Joule improvements for tasks requiring extended computation, supporting greener and more adaptive inference pipelines (2505.14733).

Potential extensions include: adaptive hyperparameter tuning (e.g., step size, defense budget), application to audio/text modalities, OOD detection, and joint optimization over compute and scaling parameters for deployment-aware tradeoffs.

---

• "A Provable Energy-Guided Test-Time Defense Boosting Adversarial Robustness of Large Vision-LLMs" (Mirza et al., 27 Mar 2026)
• "ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment" (Li et al., 29 Jan 2026)
• "TEA: Test-time Energy Adaptation" (Yuan et al., 2023)
• "The Energy Cost of Reasoning: Analyzing Energy Usage in LLMs with Test-time Compute" (2505.14733)