C.40:11 - SoTA-Echoing
| Practice question and selected line | Serious alternative, mutation and limit | Reopen condition |
|---|---|---|
| How can retained material support a useful next attempt before its ultimate contribution is settled? The selected line is examined variation with warranted intermediate retention. | Adapt the branching-discovery argument in Stanley and Lehman’s Why Greatness Cannot Be Planned (2015, especially chapters 3–6) in sections 4.1–4.2. The historical argument supplies an action-changing alternative to keeping only the present winner; its anecdotes do not establish universal institutional policy or inevitable progress. A sufficient attainable objective remains a serious cheaper choice. | Reopen if actual receiving use shows that retention cannot support the named continuation or a cheaper direct Method supplies the whole result. |
| Can branching serve a fixed contribution rather than require new-problem generation? | Darwin Gödel Machine v3, section 3 supplies a concrete evaluator-guided countercase: coding-agent variants branch from retained predecessors under supplied task evaluation. Adapt non-incumbent continuation to section 4.2; reject a universal ban on objectives. This computational evidence does not demonstrate human or organizational search effectiveness. | Reopen when the selected material cannot be changed or examined as assumed, or when incumbent-only development is sufficient at lower burden. |
| What extra operation makes changing problems and cross-problem reuse productive? | POET v3, sections 3 and 6 supplies the paired environment/agent and target-transfer construction. Adapt meaningful problem variation and target examination in section 4.3 and C.40.CD; fixed-problem branching remains sufficient where no changed problem is needed. Reject universal numeric thresholds, oldest-member removal and treating a person’s capability as a copyable policy vector. The paper is a computational instantiation, not proof of transdisciplinary effect. | Reopen when the proposed target problem is invalid, target examination is unavailable, or the coupled entry needs a different operative core. |
| How can task generation, local improvement and reuse avoid repetitive or unattainable development? | Adapt the task/solution cycle and transfer recheck of POET v3, the repertoire-relative characterization of Enhanced POET v1, and the distinct attainability/interestingness and successful/failed archive roles in OMNI v3 and OMNI-EPIC v3, into §4.6. OMNI-EPIC’s long generation runs simulate learning; its short learning runs produce specialists, not a demonstrated universal agent. Dreaming in Code v2 (2026) supplies a serious different choice: generate training worlds for one fixed receiving task and one shared policy, using local and receiving feedback. Its task engine and extra computation remain consequential limits. Generated evaluation still needs its own grounds. | Reopen when tasks become redundant after transfer, a changed repertoire invalidates their comparison, local success fails to transfer, or constructing and evaluating tasks costs more than a sufficient fixed suite. |
| How can changing opponents expose useful weaknesses without making a rising local score look like general progress? | Adapt the interaction construction and retained opponent ensemble in Tang, Tan and Harada, Learning Agile Locomotion via Adversarial Training, v1 (2020), §§III–IV, into §4.5. Its controller encodings, rewards and simulator remain domain-specific. Nolfi and Pagliuca, Global progress in competitive co-evolution (2025), §§2–4, supplies the later distinction between local, historical and estimated broader progress, and alternatives for retaining and selecting opponents. Reject an unconditional progress guarantee: their comparison concerns simulated pursuit and evasion, and fixed representative challenges remain cheaper when sufficient. §5.4 is a separate constructed transfer, not empirical evidence from either paper. | Reopen when the permitted interaction, observations, required outcome, opponent family or comparison cost changes; an archive alone does not settle the renewed claim. |
| How can a search develop the obtaining way rather than only its final artifact? | Adapt the two-level evaluation in MEA (2015), TaylorGLO (2021), AutoML-Zero (2020) and PANGAEA (2022) in §4.7: apply a candidate, judge its produced result, preserve the operative construction and check transfer. Lion (2023), §§2–3 adds progressively larger selection tasks and checked simplification after search. Direct derivation, tuning a known form and finite comparison remain serious alternatives. The sources’ computational mechanisms and experiments do not establish effectiveness for human learning. | Reopen when the representation excludes a needed operation, candidate and result are conflated, or a selected way fails at the receiving scale. |
| Which cheaper evaluation changes the claim being made? | Adapt the prediction/actual-trial distinction from AQuaSurF (2023), the continued-state alternative from EPBT (2021), and the waiting/selection distinction from AES (2024). Guijt et al. (2023), §§2–6 shows why asynchronous variation, selection and stale inputs can change search behavior. Reject treating fewer trials, inherited progress or faster return as equivalent evidence of a better fresh-start way. Full independent trials remain the simpler baseline. | Reopen when forecast errors alter selection, state compatibility fails, slow candidates disappear without examination, or total development burden outweighs the benefit. |
| How can a model and diverse expert material make action-policy search useful? | Adapt the predictor/prescriptor/actual-feedback construction of ESP (2020), §§3–5 and the Define–Gather–Distill–Evolve construction of RHEA (2024), §§2–5 and Appendices B–D, into §4.8. The NPI study (2021), §§V–VIII supplies a recurrent forecast and human trade-off use; its policy effects remain model-based. Land-use prescription (2025) supplies bounded seeding, constraint and aggregation consequences in §5.9. Direct trials, an exact policy calculation, an expert selector or a sufficient fixed policy remain serious alternatives. Reject treating model agreement, imitation rank, ancestry or a predicted front as demonstrated intervention benefit. | Reopen when the policy uses unavailable information, search reaches unsupported actions, translation loses a consequential expert distinction, or a new source response reverses the comparison. |
| How should predictive uncertainty affect the next action? | Adapt RIO (2020), §§3–5 and Appendix D.2 as one residual-correction realization, alongside bounds, direct source return and simpler correction. Its empirical predictive gains and conditional theoretical results are not arbitrary-shift coverage guarantees. The 2025 chlorination report, §§2–4 provides a consequential caution: only two of five objectives were optimized and reported violation variation suggested little learning; an attractive front need not establish the whole requested improvement. Symbolic policies in Shahrzad and Miikkulainen (2025) are a current readable-representation alternative; readability alone supplies neither causal validity nor safety. | Reopen when interval behavior, persistent disturbances, omitted outcomes or the burden of uncertainty estimation changes which policy or test is worthwhile. |
The best-known line selected for these bounded questions combines the actual shared operations with conditional problem construction and target testing. Task generation for an expanding repertoire and curriculum generation for a fixed receiving use have different selection grounds. Later generator, representation and transfer-scheduling variants do not make one implementation universal. A broader claim of present algorithmic superiority would require its own current comparison.
For comparison-guided search, adapt DNESG (2023), §§4.1–4.2.5’s learned pairwise comparator and tournament selection and CRSEA v2 (2025), §§3–3.3’s population-relative aggregation and return to actual function evaluations in :4.9. The latter makes a usable derived order; it does not validate every comparison, recover distances or establish equality from a score tie. Direct comparisons and numerical surrogates remain alternatives when they answer the receiving question more cheaply. Duel-Evolve (2026), §2 and Appendix A.1 supplies a current sparse-preference alternative: fit a Bayesian Bradley–Terry model, direct comparisons toward plausible contenders and generate from selected parents. Adapt the explicit comparison acquisition and revision, while retaining the model and approximation assumptions. Its same-model judgements can amplify shared bias (§5); more self-comparisons do not independently validate a generated answer. Reopen when judge, context, reference set, missingness, tie meaning or the cost of actual evidence changes the useful construction.
For human contributions in :4.10, adapt the task-dependent choice among demonstrations, advice and environment shaping in Karpov, Valsalam and Miikkulainen (2011), §§4–7. This historical implementation shows different useful entry points; its sixteen-participant study does not rank them for every practice. Chapter 8 of Risi, Tang, Ha and Miikkulainen’s Neuroevolution: Harnessing Creativity in AI Agent Design (2026) adds shared continuation and participation burden. Its §8.5 also develops preparation of initially uninformative material before useful human continuation. The study by Lehman and Miikkulainen (2013), §§2.3–4.2 supplies the historical image-generation and paid-selection construction; :4.10 retains the preparation-to-human-examination connection without adopting a universal aesthetic measure or a market implementation. Preserve those operations while testing the contribution in the receiving use.
For inspectable policies, adapt the executable grammar, variation and explicit action-selection alternatives of EVOTER, §§3 and 6. Direct construction and ordinary programming remain alternatives. The newer NeuroRule (2026), §§3–4 and 7 develops fitting rule sets to an existing network and balancing agreement with condition count. Adapt that obtainable copying method; qualify the query domain, lost behavior and distinct subject correctness. Both sources leave a convenient expert-intervention facility as future work. Their transparency and complexity claims do not establish human understanding, an effective intervention, fairness or safe deployment. A changed action-selection rule, consequential rare case or unavailable human contribution reopens the relevant comparison.
For representation and change in :4.11, adapt the connected representation, correspondence, developmental opportunity and incremental-construction argument of NEAT (2002), §§2.2–3.4. Its historical innovation markers align inherited structure; matching is not behavioral equivalence. HybrID (2009 manuscript), §§2–4 supplies a consequential alternative: transfer generated structures to a direct encoding for local adjustment. Its difficulty can be finding an expressible exception, rather than an inability to represent one. Adapt that distinction and the explicit transfer; retain direct construction where it is sufficient. Neither historical benchmark comparison is a general superiority claim.
The newer line can develop the generator and its changes as well as its outputs. DDE (2020), §3 and Discussion connects a learned encoding, archived examples and a mixture of exploratory operations. Meta-evolved NCA (2024), Methods and Discussion evaluates a developing decoder through an inner search over encoded inputs; its learned regularities can also limit that search. Self-Referential Graph HyperNetworks (2025), method and Discussion lets inherited parameters control the variation of copied networks while policy outcomes guide selection. Its demonstrated parameter variation leaves architecture mutation as future work and incurs substantial generation cost. Adapt the supported feedback and change relations, without making neural implementation or automatic improvement a Core requirement. Reopen when a changed generator alters retained meanings, loses a needed continuation, or costs more than the direct alternative.
For useful retained diversity in :4.12, adapt the production, local comparison and archive-reuse connection in MAP-Elites (2015), §3 and Cully and Demiris’s modular QD account (2017), §III. The latter separates collection organization from parent selection and repairs an archive replacement that can erode the retained range. These historical algorithms supply constructions, not a universal requirement for grids or behavioral distance. AURORA, §IV, extends the line to learned descriptors and rebuilding the affected collection. Extract-QD (2025), §§2–3, adds budgeted reexamination, removal and reinsertion under uncertain quality and descriptors. Adapt those consequential returns; the numerical parameters remain implementation choices. Lin et al. (2026), §§2–3 and 5 supplies a different set-optimization construction when differentiable quality and descriptors are available. A finite direct construction, or one conditioned way with its needed planner as in Batra et al. (2024), §§2 and 4, remains a serious cheaper alternative. The controller’s storage bound does not establish the whole planner’s cost. Reopen when changed descriptors, noise, lost intermediates or complete cost changes which collection and continuing operation are useful.
For actual use of a population, adapt Pardoe, Ryoo and Miikkulainen (2005), §§2–3 and 6’s ensemble construction, its state/history limit and the proposed development of components by outcomes of their combinations. Adapt Tansey, Feasley and Miikkulainen (2012), §§2–3, 5.2 and 6’s learning from locally useful examples with an explicit acceptability rule, learning groups and choice of whether acquired weight changes are inherited. Its foraging result does not rank inheritance arrangements for every changing environment. These historical neural implementations establish neither universal ensemble benefit nor transfer of a person’s capability. Sections 4.7, 4.8 and 4.10 provide the relevant state, expert and human contribution conditions; :4.12 connects retained differences to the actual operation that uses them. Reopen when selector inputs arrive too late, component state no longer follows actual actions, teaching suppresses useful variation, or one sufficient way makes the combined arrangement unnecessary.
For restarting a converged search, adapt the saved-base, change-population and repeated-rebasing construction in Gomez and Miikkulainen, technical report AI96-248, §4.2, the earlier report behind their 1997 incremental-evolution treatment. Its ESP realization applies changes to neuron weights and permits occasional larger changes; the numerical distribution and experimental gains are not Core requirements. Independent restarts, changed operators and a sufficient direct construction remain alternatives. Reopen when the retained base, allowed changes, comparison conditions or cost no longer support this continuation.
For learning on the histories produced by a developing controller in :4.7, adapt the iterative estimator-training construction of RMA (2021), §III. Its privileged simulation targets and history-conditioned deployment have different available inputs; its timing and disturbance limits remain implementation conditions. The newer SCOUT preprint (September 2026), §§3 and 5, Appendix A supports conditional teacher development on learner histories. Adapt that feedback connection within its access, outcome-checking and resource limits; long-horizon agent use remains unestablished.
For operation choice in :4.11, adapt SNAP-NEAT (2012), §§II–III and V’s connection between initial examination, continued structural development and later operator credit. Its operator repertoire and numerical settings are historical realizations. Hanna, Blot and Petke (2025), §§4 and 11 supplies a newer contrasting use: improved intermediate repair outputs did not establish more bugs repaired than uniform selection. Retain that receiving-result comparison. Reopen the allocation when delayed effects, changed conditions, sparse examination or total cost defeat its useful continuation.
For developing persistence and recovery in :4.7, adapt the repeated-use and state-return construction of Growing Neural Cellular Automata (2020), Experiments 1–3. Its longer-trajectory alternative explains the memory trade-off behind pooled starts; the pool does not establish recovery from arbitrary damage. Horibe et al. (2022), §§2–5 connects discovery of a functioning body, learning from original and damaged starts, and simulated function after recovery. Adapt that connection and its observed separation between shape similarity and locomotion, without attributing the whole comparison to optimizer choice. A sufficient direct rule and long trials with repeated examination remain alternatives.
The newer Smart cellular bricks study (2026), Damage detection and recovery and Discussion adds a consequential ambiguity: damage can make the intended shape indistinguishable from another legitimate shape; a supplied target embedding changes that information. Adapt this return to the missing target and means of repair. Its physical shape-recognition results and simulated directional detection/regrowth have different grounds. Moving and attaching physical material remain additional operations. Reopen the development arrangement when state compatibility, missing target information, tested disturbance range, actual function or complete cost changes the supported continuation.
For comparison through further variation in :4.11, adapt the joint quality/behavior account in Tarapore and Mouret, Evolvability signatures (2015), §3, the distribution update from sampled offspring in Evolvability ES (2019), §§3–6, and the explicit quality/variation comparison in Quality Evolvability ES (2021), method and main experiments. Their simulated tasks and particular descriptors do not establish universal adaptability. The meta-evolved NCA construction above supplies the distinct inner-search trial. Reject equating diverse retained answers, immediate offspring variety and useful further search. Direct construction and a sufficient fixed mapping remain alternatives. The Cost of Becoming (September 2026 preprint), §§3–7 adds a bounded trade-off among generativity, viable variation and inheritance in one compact controller simulation; equal genome dimensions and evaluation counts do not equalize mapping cost or the initial phenotype distribution. Its excluded lifetime-learning experiment supplies no adaptation evidence. Reopen when the task family, descriptors, permitted continuation or total cost changes.
For repeated contraction in :4.12, adapt the survivor–regrowth–next-event relation from Extinction Events Can Accelerate Evolution (2015), Methods and the divergent-search study (2015), §§3.3–6. The first keeps all members of five selected occupied niches; the second uses ten individual survivors and a temporary refill rule. These are different realizations, not interchangeable settings of one unstated algorithm. AURORA-XCon (2025), §§2–6 adds contraction with a protected elite and a learned behavior descriptor whose training distribution can change; it does not directly train the genotype-to-phenotype mapping. Adapt that consequential descriptor return, without claiming universal superiority or measured acquisition of future variability. Continued search, direct offspring trials and an adequate fresh start remain serious alternatives. Reopen when survival, grouping, regrowth, retained answers or rebuilding cost changes the useful effect.