Self-Error Adjustment (SEA) Methods

• SEA is a framework that leverages explicit error signals—such as residuals, execution failures, or deviations—to adjust predictions and improve system performance.
• It is applied across diverse domains including spatiotemporal traffic forecasting, iterative feedforward refinement in robotics, and execution-aware corrections in language models.
• SEA also extends to ensemble learning and self-evolving agents, balancing individual performance with diversity and enabling regulated, auditable self-modification for persistent adaptation.

Self-Error Adjustment (SEA) is not a single standardized technique but a recurring research pattern in which a system uses structure in its own errors, residuals, execution failures, or historical trajectories to modify subsequent behavior. Across the works collected under this label or under closely related interpretations, SEA appears as spatiotemporal residual correction in traffic forecasting, iterative feedforward refinement in compliant robot control, execution-grounded repair in Text-to-SQL, error-generalized data synthesis for mathematical reasoning, accuracy–diversity balancing in ensembles, and certified self-modification in self-evolving agents (Liu et al., 25 Oct 2025, Li et al., 2024, Zou, 7 Aug 2025, Sengupta, 1 Jul 2026).

1. Terminological scope and recurrent structure

The acronym “SEA” is overloaded in the cited literature. In "Error Adjustment Based on Spatiotemporal Correlation Fusion for Traffic Forecasting" it denotes a direct residual-correction mechanism, implemented as Spatiotemporally Autocorrelated Error Adjustment (SAEA) (Liu et al., 25 Oct 2025). In "Self-Error Adjustment: Theory and Practice of Balancing Individual Performance and Diversity in Ensemble Learning" it denotes a loss-level ensemble-learning framework (Zou, 7 Aug 2025). In "SEA-SQL: Semantic-Enhanced Text-to-SQL with Adaptive Refinement" it names a Text-to-SQL pipeline whose components are Adaptive Bias Elimination and Dynamic Execution Adjustment (Li et al., 2024). In the robotics papers, “SEA” primarily denotes Series Elastic Actuators, but the described iterative learning procedures are explicitly interpreted as self-error-based adjustment mechanisms (Banka et al., 2017, Yan et al., 2020). In the agent literature, “SEA” denotes Self-Evolving Agent or Self-Evolving Agents with Anytime-Valid Certificates (Jiang et al., 10 Apr 2026, Sengupta, 1 Jul 2026).

Domain	Error signal	Adjustment mechanism
Traffic forecasting	Autocorrelated residuals	VAR-based correction and adjusted loss
Robot control	Past tracking error	Iterative feedforward update
Text-to-SQL / LLM reasoning	Execution outcomes or bad cases	Refinement, repair, or targeted retraining
Ensemble learning	Self-error and inter-learner interaction	Adjustable per-learner loss
Self-evolving agents	Failure attributions and task histories	Persistent memory and gated self-modification

Taken together, these works suggest a common pipeline: an error signal is first made explicit, then parameterized or summarized, and finally reintroduced into training, inference, or control. The signal may be a residual vector, a failed SQL execution, a trajectory-tracking deviation, a clustered error type, or a cross-task failure attribution. What varies is the object being updated: predictions, inputs, base-learner losses, prompts and tools, or persistent memory.

A common misconception is that SEA is necessarily a post-hoc heuristic layered on top of a fixed system. The cited works show several alternatives. In traffic forecasting, the error model is embedded directly into the cost function and the inference equation (Liu et al., 25 Oct 2025). In ensemble learning, the decomposition is itself the training objective (Zou, 7 Aug 2025). In certified self-evolving agents, self-adjustment is admitted only through an anytime-valid gate that emits an auditable certificate against a fixed error budget (Sengupta, 1 Jul 2026).

2. Spatiotemporal residual modeling in traffic forecasting

The traffic-forecasting formulation is the most explicit realization of SEA as “learn how your own errors evolve, then use that learned structure to correct future predictions.” The problem begins with a one-step predictor

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \mathbf{e}_t,$$

where standard MSE- or MAE-based training implicitly assumes that prediction errors are independent over time and often independent over space. The paper argues that this assumption does not hold for traffic data because of temporality, spatiality, shared exogenous factors, congestion propagation, and model misspecification. Empirically, it reports strong off-diagonal structure in spatial error covariance matrices, clear cross-lag covariance patterns, and significant residual autocorrelation functions, indicating that the residuals are not white noise (Liu et al., 25 Oct 2025).

To model this structure, SAEA replaces the i.i.d. residual assumption with a spatiotemporal VAR process. For VAR(1),

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$

with $\Phi \in \mathbb{R}^{N \times N}$ encoding how past residuals propagate across sensors and across time. Substituting this into the forecasting model yields an adjusted equation,

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$

and, after approximation, a transformed input–target relation in which both inputs and targets are shifted by $\Phi$-weighted past states. The corresponding training loss is

$$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$

When $\Phi = 0$, this reduces to standard MSE training; this is why the method is described as a strict generalization of MSE.

A major design component is the coefficient matrix $\Phi$. For one-step prediction, $\Phi_{ij}$ is the influence of sensor $j$ at time $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$0 on the adjusted state or error at sensor $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$1 at time $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$2. For multi-step prediction, the method uses a horizon-specific matrix $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$3 for each step $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$4. To align the learned dependencies with road-network structure, the paper introduces structurally sparse regularization. With adjacency matrix $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$5, degree matrix $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$6, Laplacian $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$7, and mask

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$8

the regularizer is

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$9

The intent is to favor coefficients compatible with physical connectivity and to avoid overfitting arbitrary cross-sensor correlations.

At inference time, SEA appears as an explicit correction term. For one-step forecasting,

$\Phi \in \mathbb{R}^{N \times N}$0

For multi-step forecasting, each horizon $\Phi \in \mathbb{R}^{N \times N}$1 uses a separate model $\Phi \in \mathbb{R}^{N \times N}$2 and coefficient matrix $\Phi \in \mathbb{R}^{N \times N}$3,

$\Phi \in \mathbb{R}^{N \times N}$4

The paper emphasizes that this is a single forward pass rather than iterative test-time optimization.

The empirical results are reported on PeMS and METR-LA with input length 12 and direct 15, 30, and 45 minute prediction horizons. Across five base forecasting models—SVR, FC-LSTM, STGCN, Graph WaveNet, and AutoSTG—the proposed method improves RMSE and MAPE in almost all cases. Table 1 reports an average accuracy improvement of about $\Phi \in \mathbb{R}^{N \times N}$5 across five models, larger gains at 30 and 45 minutes, and better performance for structurally sparse $\Phi \in \mathbb{R}^{N \times N}$6 than scalar, diagonal, generic sparse, and low-rank variants. One example on PeMS at 45 minutes reports STGCN improving from RMSE $\Phi \in \mathbb{R}^{N \times N}$7 without adjustment to RMSE $\Phi \in \mathbb{R}^{N \times N}$8 with SAEA, an approximately $\Phi \in \mathbb{R}^{N \times N}$9 reduction. The paper also reports that VAR(1) outperforms VAR(2) in its toy comparison, and that error correlation matrices become much more diagonal after adjustment.

3. Iterative feedforward refinement in robots with Series Elastic Actuators

In the robotics literature considered here, self-error adjustment is formulated as iterative improvement of a feedforward command for robots driven by Series Elastic Actuators. The control problem is precision trajectory tracking for compliant, low-impedance systems whose safety and force-control advantages come at the cost of lower positioning precision and bandwidth. The 2017 paper studies a 2-DOF robotic arm and updates the input across repetitions by

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$0

or, equivalently,

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$1

so that the next feedforward command is computed directly from the previous iteration’s tracking error. The paper treats the 2-DOF SEA robot as a time-varying, locally linear MIMO system, learns local frequency-domain models with complex-valued Gaussian Process Regression, and uses those models to invert the local dynamics and update the command (Banka et al., 2017).

The modeling layer is central. For scalar systems, the paper places a GP prior on $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$2; for the nonlinear, time-varying robot it uses inputs $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$3 and a squared-exponential kernel with ARD. For MIMO identification, each transfer element $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$4 is modeled as an independent GP and combined through a weighted kernel derived from the active inputs. The corrected time-domain law is written as

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$5

The paper also develops convergence conditions. For the MIMO case, convergence requires all eigenvalues of

$$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$6

to have magnitude less than $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$7, and derives diagonal-gain constraints via Geršgorin-disc arguments.

The reported 2-DOF experiments use HEBI X5-4 SEAs with torsional stiffness approximately $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$8. On a slow, large-range trajectory, initial maximum errors $$\mathbf{G}_t = f(\mathbf{G}_{t-1}, \dots, \mathbf{G}_{t-H}; \theta) + \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$9 rad at iteration $\Phi$0 are reduced to $\Phi$1 rad after 20 iterations, about $\Phi$2 of the range and about $\Phi$3 relative reduction. On a faster trajectory, initial maximum errors $\Phi$4 rad are reduced to $\Phi$5 rad after 20 iterations, about $\Phi$6 of the range and about $\Phi$7 reduction. The paper states that convergence is essentially achieved after only 5 iterations in the slow regime.

The 2020 paper extends the same logic to a 3-DOF SEA robot using MIMO ILC with an input-weighted complex kernel. The control layer updates reference joint trajectories by

$\Phi$8

where $\Phi$9 is the local MIMO frequency response. The paper introduces the input-weighted complex kernel

$$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$0

to identify local MIMO models from multi-input excitation, then bounds uncertainty and derives Geršgorin-theorem-based sufficient conditions on the iteration gains $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$1 (Yan et al., 2020).

The experimental setting is a 3-DOF low-profile arm with HEBI X5-4 SEAs performing reciprocating brush motion for pilot-hole cleaning. Without ILC, mean joint errors grow substantially as the period decreases from $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$2 s to $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$3 s. After about 10 iterations at the fast $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$4 s setting, the paper reports reductions from $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$5, $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$6, and $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$7 rad to $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$8, $$\mathcal{L} = \left\| \mathbf{G}_t - \Phi \mathbf{G}_{t-1} - f(\mathbf{G}_{t-1} - \Phi \mathbf{G}_{t-2}, \dots; \theta) \right\|_2^2 + \alpha R.$$9, and $\Phi = 0$0 rad, close to the repeatability noise floor. The abstract summarizes the outcome as around $\Phi = 0$1 improvement in positioning precision and a 10-times increase in operating speed. In both robotics papers, a plausible implication is that self-error adjustment is most effective when the task is repeatable, local linearization is adequate, and feedback alone is bandwidth-limited by compliance.

4. Execution-grounded refinement in LLMs

In language-model systems, SEA is instantiated as execution-aware refinement rather than residual dynamics. SEA-SQL is a three-stage Text-to-SQL pipeline consisting of a Semantic-Enhanced Schema, Adaptive Bias Elimination, and Dynamic Execution Adjustment. The task is to generate a SQL query $\Phi = 0$2 from question $\Phi = 0$3 and schema $\Phi = 0$4. First, a zero-shot GPT-3.5 call produces a pre-SQL $\Phi = 0$5 from a semantic-enhanced schema that augments tables, columns, and foreign keys with retrieved column values. Second, a fine-tuned Mistral-7B bias eliminator $\Phi = 0$6 maps

$\Phi = 0$7

where $\Phi = 0$8 is the execution result or error of $\Phi = 0$9. Third, Dynamic Execution Adjustment executes $\Phi$0, captures error messages, asks GPT-3.5 to reflect on the cause, and iteratively corrects the SQL until a version capable of successful execution is obtained (Li et al., 2024).

The semantic-enhancement stage uses BM25 indexing over database contents, retrieves top-$\Phi$1 candidate values per column with respect to the question, and retains top-$\Phi$2 values via the bridging strategy. The paper reports that removing semantic-enhanced values causes a severe drop on BIRD: EX decreases from $\Phi$3 to $\Phi$4, a $\Phi$5-point decline. The adaptive-bias stage is trained not simply to rewrite queries into canonical gold form, but to preserve $\Phi$6 when it is execution-equivalent to the gold SQL. The dynamic-adjustment stage maintains a history of prior SQLs, error messages, reasons, and corrected SQLs so that the model can avoid repeating failed variants.

On Spider dev, SEA-SQL reports EX $\Phi$7, higher than GPT-3.5 C3 at $\Phi$8 and comparable to GPT-4 DAIL-SQL at $\Phi$9. On BIRD dev, it reports ACC $\Phi_{ij}$0 and VES $\Phi_{ij}$1, higher than GPT-3.5 MAC-SQL’s ACC $\Phi_{ij}$2. The cost analysis reports per-query generation cost $\Phi_{ij}$3. The ablation table shows that SEA-SQL outperforms all variants without ABE, without DEA, without both, and without SES.

A closely related but distinct direction is "Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning". The abstract states that the framework explores a target model on GSM8K and MATH to pinpoint bad cases, generates error keyphrases for these cases based on the instructor model’s analysis, identifies error types by clustering these keyphrases, samples a few bad cases during each generation for each identified error type, synthesizes additional training data using a self-instruct approach, refines the new data through a one-shot learning process, fine-tunes the target model on the curated data, and iteratively repeats the process to enhance performance (Yu et al., 28 May 2025). Because the supplied details explicitly state that the full text was unavailable, only the abstract-level characterization is secure. Even at that level, the method places SEA in a broader class of error-generalization procedures: errors are not only repaired instance by instance, but also converted into a taxonomy that drives targeted retraining.

A common misconception in LLM applications is that post-correction is merely prompt engineering. The cited systems are more structured. SEA-SQL introduces a learned bias eliminator trained on execution outcomes and a runtime Execute–Reflect–Correct loop (Li et al., 2024). Self-Error-Instruct uses clustered error types and curated synthetic data generation rather than isolated bad-case extrapolation (Yu et al., 28 May 2025).

5. Loss-level self-error adjustment in ensemble learning

In ensemble learning, SEA is formulated as a decomposition of ensemble error into self-error and diversity terms, with a continuous control parameter. For $\Phi_{ij}$4 base learners $\Phi_{ij}$5 and ensemble prediction

$\Phi_{ij}$6

the paper rewrites the training-relevant part of the ensemble’s squared error for learner $\Phi_{ij}$7 as

$\Phi_{ij}$8

The first term is the self-error term; the second is the diversity term, measuring how learner $\Phi_{ij}$9’s error interacts with the errors of the other learners. The paper then defines a complementary prediction $j$0 satisfying

$j$1

so that

$j$2

and rewrites the decomposition as

$j$3

SEA introduces an adjustable parameter $j$4 and defines the per-learner loss

$j$5

or, up to a constant,

$j$6

This makes $j$7 an explicit knob controlling the balance between individual performance and diversity (Zou, 7 Aug 2025).

The theoretical analysis distinguishes two conditions: the loss must be optimizable, and optimizing it must reduce ensemble error. For SEA, the Hessian with respect to $j$8 is always positive, so convexity imposes no bound on $j$9. The tighter bound comes from the ensemble-error reduction condition, yielding

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$00

The paper then maps Negative Correlation Learning and NCL$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$01 into SEA’s parameterization and derives tighter theoretical bounds for their parameters than Hessian-only analyses provide. It also derives a diversity measure, the standard deviation of learner predictions,

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$02

and shows that, within the effective range, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$03 is linear in $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$04, whereas the corresponding NCL diversity curve is nonlinear and becomes highly curved for large ensembles.

The reported experiments use 9 regression datasets and 6 classification datasets, with 3-layer MLPs or LeNet-5 as base learners and ensemble sizes $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$05. SEA achieves the best RMSE on all 9 regression datasets and the best accuracy on all 6 classification datasets. The regression table reports examples such as pyrim $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$06 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$07, triazines $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$08 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$09, cpusmall $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$10 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$11, and mpg $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$12 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$13, with reported improvements from $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$14 to $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$15. The classification table reports gains such as ionosphere $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$16 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$17 and german $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$18 vs $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$19. Ablations further report

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$20

indicating that both the broader adjustment range and the specific SEA objective contribute to performance.

This formulation is noteworthy because it relocates SEA from residual correction to objective design. Here, “self-error” does not mean temporal residual history but the individual learner’s own deviation from the target. The paper’s central claim is that precise control over the self-error/diversity trade-off is easier when the loss is written directly in those terms.

6. Cross-task self-correction in self-evolving agents

In the agent literature, SEA is expanded from local correction to persistent adaptation across task boundaries. "SEA-Eval: A Benchmark for Evaluating Self-Evolving Agents Beyond Episodic Assessment" defines a self-evolving agent formally as

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$21

where $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$22 is an open-ended digital environment, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$23 is an extensible action space, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$24 is persistent stratified memory, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$25 is an experience-distillation function mapping trajectories to memory updates, and $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$26 is a selective cross-task memory-transfer function. The paper presents the Evolutionary Flywheel as the minimal sufficient architecture and introduces SEA-Eval as the first benchmark designed specifically for evaluating SEAs. Its primary metrics are success rate $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$27 and token consumption $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$28, and its central empirical claim is that identical success rates can conceal very different execution costs: under identical success rates, token consumption differs by up to $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$29 across frameworks, demonstrating that success rate alone creates a capability illusion (Jiang et al., 10 Apr 2026).

SEA-Eval operationalizes this by using correlated sequences to measure evolutionary gain, orthogonal sequences to measure evolutionary stability, and implicit-intent sequences to measure alignment convergence. It defines auxiliary rates such as the Efficiency Evolution Rate $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$30, the $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$31 growth slope $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$32, and the alignment convergence rate $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$33. The paper argues that the sequential convergence of $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$34 is the key criterion for distinguishing genuine evolution from pseudo-evolution. In its case study, OpenClaw and GenericAgent both achieve $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$35 success rate on a static task set, yet OpenClaw consumes about $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$36 tokens versus about $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$37 for GenericAgent; the sequential analysis then shows pseudo-evolution for OpenClaw and genuine evolution for GenericAgent.

"Self-Evolving Agents with Anytime-Valid Certificates" moves from benchmarking to architecture. It defines the deployed agent at round $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$38 as

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$39

with a frozen base model $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$40, a trainable steering adapter $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$41, a mutable versioned harness $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$42, and five loop controllers outside the forward path. Self-modification is confined to the steering adapter and the harness, and every modification is admitted only through an anytime-valid gate that spends from a fixed global error budget

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$43

The gate for harness edits computes a lower confidence bound corrected by a Wasserstein-1 term,

$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$44

and admits the edit only if the corrected lower bound exceeds the tolerance threshold. The architecture couples five loop controllers—PPB-CL, PNMP-A, PA-COCOA, SGM-CS, and SDC-QD—with five verifier-in-the-loop mechanisms: best-of-$$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$45, micro-step search, self-authored reproduction oracles, search-layer control, and self-repair (Sengupta, 1 Jul 2026).

The reported experiments use a 52-instance SWE-bench Verified subset across four base models. The main table reports resolved-instance counts of $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$46 for Gemma, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$47 for Qwen, $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$48 for Gpt-mini, and $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$49 for Gpt under the composite SEA stack. A no-op-composite control isolates the suite’s contribution on two strong bases: Glm 5.2 improves $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$50, and Gpt improves $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$51, with $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$52 reported as the $$\boldsymbol{\eta}_t = \Phi \boldsymbol{\eta}_{t-1} + \boldsymbol{\varepsilon}_t,$$53 best. The paper emphasizes that results are single-run on expensive evaluations and that confirming run-to-run variance is future work.

These agent papers broaden the meaning of self-error adjustment. Instead of correcting a numerical residual or a failed SQL query, the system stores failure attributions, updates procedural memory, restricts changes to auditable components, and accepts those changes only through statistical gates. A plausible implication is that SEA at agent scale becomes less about one-step correction and more about regulated accumulation of experience.