When and why does improving worst‑case or distributional robustness require real sacrifices in standard performance, versus being avoidable through better data, model capacity, training objectives, or evaluation choices?

Adversarial trade-offs

Adversarial training generally seeks to reduce error under malicious or worst-case input changes, but several theoretical and empirical studies find that this can increase error on unperturbed inputs. The trade-off is not universal or fixed: it varies with data quality and amount, model size and architecture, perturbation type and strength, overparameterization, and the training method. Some methods and settings report improvements in both robustness and standard performance, including robust self-training and selected neural-network experiments, while application studies such as robot learning report net performance costs. The main disagreement is whether robustness–accuracy tension is a fundamental constraint in practical systems or mainly a consequence of particular data, models, attacks, and evaluation choices.

10 sources

Robustness vs accuracy

Several theoretical and empirical studies find that adversarial training often improves robustness to worst-case perturbations while reducing accuracy on unperturbed inputs, especially with conventional methods. Other research argues that the trade-off is not inherent: suitable data geometry, locally adaptive robustness definitions, local Lipschitz methods, or alternative training procedures may preserve both objectives. The main disagreement is whether observed losses in clean accuracy reflect a fundamental limitation of robustness or limitations in the robustness definition, model class, data assumptions, and training algorithms.

9 sources

Robustness testing

Robustness testing examines whether software continues to behave acceptably under invalid inputs, faults, failures, or stressful conditions—not merely on its intended “happy path.” Supporters emphasize fault injection, fuzzing, model-based generation, and automated exploration; critics caution that tests can miss important behaviors and cannot by themselves establish correctness.

11 sources

Adversarial training

Adversarial training incorporates deliberately perturbed inputs into model training, typically by solving a min-max problem: generate perturbations that increase loss, then update the model to reduce that loss. The mainstream research account holds that adversarial training can substantially improve robustness to specified attack classes, with PGD-based training among the most effective established defenses. Important qualifications include dependence on attack quality and threat model, possible computational and clean-accuracy costs, and evidence that adversarial training can hurt robust generalization when data are scarce. The central disagreement is not whether adversarial training can help, but how reliably it generalizes beyond the attacks and data conditions used during training, and when its costs outweigh its benefits.

10 sources

Overparameterization effects

The sources report that highly overparameterized models can interpolate training data yet generalize well, sometimes showing double or triple descent, faster optimization, and improved test performance as width increases. The main disagreement is whether these effects reflect broadly applicable principles or results that depend strongly on data, architecture, optimization, and carefully chosen assumptions.

9 sources

Data quality and augmentation

The mainstream account treats data quality improvement and augmentation as complementary tools: cleaning, monitoring, transformation, and carefully designed additional examples can expand useful training data and support better ML systems. Dissenting analyses stress that generated data can distort distributions, fail on rare or complex cases, or add little value when real classes are already separable. The main disagreement is whether synthetic data is a reliable scalable solution or a task-dependent risk requiring strong external validation.

11 sources

Robustness in robotics

Robustness in robotics generally means maintaining task success despite uncertainty, variation, disturbances, and imperfect perception. Research points toward combining diverse data, formal safety and stability tools, robust control, hardware design, and evaluations that test distribution shifts—not relying on benchmark performance alone. A continuing disagreement concerns where robustness should come from: larger and more diverse datasets, principled control and certification, biologically inspired architectures, or changes to training procedures. The evidence suggests these approaches address different failure modes, while current benchmarks can overstate real-world reliability.

9 sources