Validation-Free Evolution in Adaptive Systems

• Validation-free evolution is a design principle that replaces traditional external validation with intrinsic, traceable signals and first-principles benchmarks.
• It is applied in domains like software evolution, few-shot adaptation, federated learning, face recognition, and metrology using proxy indicators.
• It leverages iterative workflows—constructing internal references, evolving candidates, and enforcing controlled acceptance—to ensure robust system performance.

Searching arXiv for recent and foundational papers related to validation-free evolution across software evolution, verifier-free inference, few-shot adaptation, federated learning, dataset assessment, and broader evolvability frameworks. Validation-free evolution denotes a family of adaptive procedures in which iterative improvement is performed without the conventional external artifact that would ordinarily authorize or score change: a pre-existing regression suite, a held-out validation split, hidden tests, certified reference materials, or an external verifier. The common move is not to eliminate control, but to replace it with intrinsic, traceable, or comparison-based signals such as point-of-interest traces in program executions, model-intrinsic confidence and answer diversity, one-shot hold-out calibration, proxy-embedding geometry, cross-round parameter alignment, paired trajectory auditing, or first-principles metrological quantities (Insa et al., 2017, Maheswaran et al., 9 Apr 2026, Vorster et al., 4 Mar 2026, Chen et al., 28 May 2026, Wu et al., 2024, Gao et al., 12 Jun 2026, Hönicke et al., 2019). The literature uses the expression in several non-equivalent senses, ranging from practical engineering workflows that avoid explicit validation data to theoretical accounts of evolution and learning that proceed without labeled examples or, in a stronger sense, without variation and selection in the Darwinian form (Fidalgo et al., 24 Jul 2025, Gabora et al., 2024).

1. Semantic scope and domain-specific meanings

In software evolution, validation-free evolution refers to checking whether a new release preserves the intended behavior of an old release when no pre-existing regression test suite exists. "Erlang Code Evolution Control" (Insa et al., 2017) formulates the problem as automatic generation of a comparison-oriented suite from the old version itself: a programmer-selected point of interest (POI) is traced, the values observed there become the behavioral reference, and the same tests are replayed on the new version.

In few-shot CLIP adaptation, the term denotes strict avoidance of validation-set selection and test-set tuning for the blending ratio $\alpha$. "Hold-One-Shot-Out (HOSO) for Validation-Free Few-Shot CLIP Adapters" (Vorster et al., 4 Mar 2026) removes exactly one labeled example per class from the support set, trains the adapter on the remaining examples, and uses the one-shot hold-out cache only to learn the blending-ratio logit. The method is therefore validation-free in the sense of not requiring additional labeled data beyond the few-shot budget.

In federated learning, validation-free means contribution assessment without access to any validation data. "CoAst: Validation-Free Contribution Assessment for Federated Learning based on Cross-Round Valuation" (Wu et al., 2024) treats representative validation datasets as often unavailable in practical FL scenarios and replaces them with a cross-round similarity analysis between client updates and subsequent global updates.

In large-scale face recognition, the term refers to evaluating the inherent potential of a dataset to produce high-performance models without full-scale training and without a clean held-out validation set. "Efficient, Validation-Free Intrinsic Quality Estimation for Large-Scale Face Recognition Datasets" (Chen et al., 28 May 2026) estimates trainability from proxy embeddings using a Neighbor-Consistency Score and Global Representation Subspace Complexity.

In LLM-based inference and skill improvement, closely related expressions are verifier-free and ground-truth-free. "Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution" (Maheswaran et al., 9 Apr 2026) assumes no external verifier and routes evolutionary stages by model-intrinsic confidence or answer diversity. "SkillAudit: Ground-Truth-Free Skill Evolution via Paired Trajectory Auditing" (Gao et al., 12 Jun 2026) assumes the evolution loop has no hidden tests, reference solutions, validation scores, or oracle rewards, and instead contrasts trajectories produced with and without the candidate skill.

A distinct metrological usage appears in "Reference-free GIXRF-XRR as a methodology for independent validation of XRR on ultrathin layer stacks and a depth-dependent characterization" (Hönicke et al., 2019). There, the analysis is not validation-free in the sense of dispensing with validation altogether; rather, it is validation-independent in the sense that the absolute benchmark does not depend on conventional reference materials.

2. Replacement signals and intrinsic benchmarks

Across these literatures, the decisive technical question is what replaces the missing validation artifact. The replacements differ sharply by domain, but they share the property that they are observable within the evolution loop itself or derived from first principles.

Setting	Missing artifact	Replacement signal
Erlang code evolution	Pre-existing regression suite	POI trace
Few-shot CLIP adaptation	Validation split or test-set tuning	One-shot hold-out cache
Federated learning	Validation dataset	Cross-round similarity
Face-recognition datasets	Full training and clean validation set	IQ from proxy embeddings
Verifier-free inference	External verifier	Group confidence or answer diversity
Skill evolution	Hidden tests or oracle rewards	Paired trajectory auditing
XRR thin-film analysis	Certified standards	Reference-free mass deposition

In SecEr, the output of a test case is the sequence of values obtained by evaluating the POI during execution, and the sequence is explicitly called a trace. Preservation of behavior is therefore defined as trace equality between old and new versions for the same input (Insa et al., 2017).

In Squeeze Evolve, the routing signal is model-intrinsic. Group confidence is computed from token log-probabilities, and when logprobs are unavailable the framework falls back to answer diversity,

$$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$

Low group confidence or high $D$ indicates uncertainty or disagreement and is used to decide whether a group should be handled by a cheaper model, a stronger model, or a lightweight non-LLM aggregator (Maheswaran et al., 9 Apr 2026).

In IQ, the intrinsic score is

$$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$$

with default setting

$\alpha = 0.2,\quad \beta = 0.8.$

Here $c$ is the dataset-level Neighbor-Consistency Score and $\tilde r_{\mathrm{ent}}$ is the normalized Effective Rank of the embedding covariance. The first is a local label-agreement signal, while the second is a global subspace complexity signal (Chen et al., 28 May 2026).

In HOSO-Adapter, the replacement for validation-set tuning is a one-shot internal calibration mechanism. The final feature is

$$\tilde{v} = (1-\alpha)\,v + \alpha\,v_{\text{adapt}},$$

with

$$\alpha = \sigma(\alpha_{\text{logit}})\cdot 0.8 + 0.1.$$

The adapter parameters $\gamma$ are optimized on $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$0, whereas $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$1 is optimized on the held-out cache $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$2 (Vorster et al., 4 Mar 2026).

In CoAst, the local update is

$$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$3

then ternary quantization keeps only the most important parameter directions, and the contribution score is based on signed agreement with cumulative future global updates:

$$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$4

This converts later optimization trajectory information into a validation-free contribution signal (Wu et al., 2024).

In reference-free GIXRF-XRR, the central absolute benchmark is elemental mass deposition,

$$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$5

measured from first principles with radiometrically calibrated instrumentation, fundamental atomic parameters, and Sherman's equation. The quantified mass depositions can then be used to independently validate any XRR modeling result (Hönicke et al., 2019).

3. Recurrent algorithmic structure

Despite large differences in application domain, the practical systems exhibit a recurrent workflow. First, they construct an internal reference from the current artifact, proxy model, or calibrated instrument. Second, they evolve candidates against that reference. Third, they constrain acceptance through a fixed rule, uncertainty threshold, or rollback mechanism. This suggests a shared architecture of internalized evaluation, although the formal guarantees and observables differ across fields.

SecEr is explicitly organized as a three-phase pipeline: type analysis and initial test-input generation; trace-recording instrumentation and test-suite construction; and comparison against a new version. The workflow is to infer types, generate candidate concrete calls, instrument the code, store each distinct trace with the inputs that produce it, mutate inputs to explore more traces, and then replay the resulting suite on the new version (Insa et al., 2017).

Squeeze Evolve is likewise iterative. It initializes all $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$6 trajectories with the expensive model $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$7, then in each loop scores the current population, selects groups, computes group fitness, routes groups into $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$8, $$D(g)=\bigl|\{\mathrm{answer}(\tau): \tau \in g\}\bigr|.$$9, or $D$0, aggregates each bin with the assigned operator, and updates the population by replacement or accumulation. A per-problem percentile threshold determines which groups are “easy” enough for the cheaper model (Maheswaran et al., 9 Apr 2026).

HOSO has a simpler but structurally similar separation. The support set is partitioned into the held-out cache $D$1 and the remaining training set $D$2. The adapter learns task-specific features on $D$3, while the blending-ratio logit is trained separately on $D$4. The paper emphasizes that this decoupling prevents co-adaptation and overfitting (Vorster et al., 4 Mar 2026).

CoAst combines two stages: weight quantization or pruning to keep only the most important part of local updates, and cross-round valuation based on the similarity between the current local parameters and the global parameter updates in several subsequent communication rounds. The design explicitly rejects single-round valuation as too sensitive to stochastic optimization (Wu et al., 2024).

SkillAudit formalizes its loop as task interpretation, one-time compilation and locking of an Anchor Verifier, paired execution with and without the skill, aggregation of PACE and verifier outputs, commit or rollback, and early stopping when no actionable surgery targets remain and the structural verifier passes. The two edit pipelines, Refine and Repair, differ in whether the skill is treated as broadly useful but noisy or actively conflicting with the task (Gao et al., 12 Jun 2026).

The metrological hybrid in GIXRF-XRR follows the same general logic of constraint by intrinsic reference. Instead of fitting thickness and density independently for every layer, it first determines the absolute elemental mass depositions from reference-free GIXRF and then distributes those quantities in depth across the stack. The total amount of each element is fixed by the quantification, while the layer model only decides how that material is partitioned vertically and how interfaces are intermixed. The explicit purpose is to reduce the degrees of freedom (Hönicke et al., 2019).

4. Representative systems and empirical performance

The practical significance of validation-free evolution is established less by a single benchmark than by repeated demonstrations that internal or proxy signals can replace conventional validation artifacts without collapsing the task.

In software evolution, SecEr was demonstrated on string:tokens/2 from Erlang/OTP. With the POI set to the output of string:tokens/2, SecEr generated 6030 test cases and reported that both versions produced the same traces for all tests. When the authors introduced a deliberate fault, SecEr generated 6010 test cases, of which 497 mismatched, and reported a concrete failing input (Insa et al., 2017). The same paper reports that, in the happy numbers example, the method detected both a different trace length and different trace values or order.

In validation-free few-shot CLIP adaptation, HOSO-Adapter outperformed the strict validation-free CLIP-Adapter baseline by more than 4 percentage points on average across 11 standard few-shot datasets. On ViT-B/16 in the 16-shot setting, the paper reports 80.33% average for HOSO-Adapter versus 75.82% for the CLIP-Adapter validation-free baseline, a gain of about +4.5 percentage points. In the 8-shot and 16-shot settings, HOSO-Adapter also outperformed CLIP-Adapter even when CLIP-Adapter was allowed to use the test-set-optimal blending ratio (Vorster et al., 4 Mar 2026).

In verifier-free evolutionary inference, Squeeze Evolve reports consistent improvements on the cost-capability frontier. Representative homogeneous open-source results include AIME 2025: 90.7 vs. 89.2 accuracy at \$D$50.94</strong>, <strong>HMMT 2025: 92.0 vs. 89.7 at \$D$60.41</strong>, <strong>GPQA-Diamond: 75.9 vs. 74.0 at \$D$70.57</strong>, and <strong><a href="https://www.emergentmind.com/topics/livecodebench-lcb" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">LiveCodeBench</a> V6: 75.6 vs. 75.9 at \$D$$80.44</strong>. Across reasoning benchmarks, the paper reports a <strong>1.3×–3.3× cost reduction</strong>. On ARC-AGI-V2, a diversity-routed version reaches <strong>97.5$$D$95.93 per task (Maheswaran et al., 9 Apr 2026).

In face-recognition dataset assessment, the empirical claim is not one benchmark score but strong predictive alignment. IQ rises with clean scaling from WebFace4M to WebFace12M to WebFace42M, while injected label noise makes downstream accuracy drop, ER increase, and Consis decrease sharply. The paper reports that IQ achieves essentially perfect rank agreement in the controlled settings and very high correlation to downstream performance across clean, noisy, and mixed scale-corruption comparisons (Chen et al., 28 May 2026).

In federated learning, CoAst achieves an average score of 0.9, while the validation-based SV baseline has average 0.855. The paper states that CoAst is comparable to validation-based methods and better than prior validation-free methods in almost all cases, and recommends $\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$0 or $\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$1 for the future-round window (Wu et al., 2024).

In ground-truth-free skill evolution, SkillAudit achieves 73.9% average task reward across 89 containerized tasks spanning 8 professional domains, outperforming an agent without skills (40.9%) and the static expert skill (56.7%). The paper further reports that SkillAudit improves over the static skill in 7 of 8 domains, matches it in Finance/Economics, preserves or improves 92% of tasks where the initial skill already worked, and lifts 43% of low-quality initial skills to passing (Gao et al., 12 Jun 2026).

In nanoscale thin-film characterization, the validation result is negative for conventional XRR alone. The reported XRR fits matched the measured curves well, but comparison to reference-free GIXRF quantification revealed significant discrepancies: for Al the agreement was only within uncertainty for one sample, for oxygen the XRR model systematically gave too much oxygen, and for Hf the agreement was similarly imperfect. In the hybrid analysis of sample S4 and its annealed counterpart, the resulting concentration depth profiles showed increased interfacial intermixing after annealing and growth of the interfacial SiO$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$2 layer (Hönicke et al., 2019).

5. Theoretical formulations of evolution without external validation

The theoretical literature broadens the meaning of validation-free evolution from “no validation set” to “learning from intrinsic performance” and, in one line of work, to cumulative adaptive change without Darwinian selection in the standard sense.

Valiant-style evolvability treats evolution as learning without labeled training examples in the usual supervised-learning sense. "Simulating Evolvability as a Learning Algorithm" (Fidalgo et al., 24 Jul 2025) defines performance over the Boolean hypercube by

$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$3

and uses the empirical estimate

$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$4

Candidate mutations are partitioned into beneficial and neutral sets according to a tolerance parameter, and the mutator selects a beneficial mutation if any exist, otherwise a neutral mutation, otherwise terminates. The empirical study validates known results such as evolvability of monotone conjunctions and non-evolvability of parity, and reports that neutral mutations are essential for escaping fitness plateaus (Fidalgo et al., 24 Jul 2025).

"Evolving a Vector Space with any Generating Set" (Nock et al., 2017) generalizes this line by showing that all it takes to evolve a normed vector space is merely a set that generates the space. Under the paper’s singularity-free assumptions, the class is distribution-free evolvable by any permissible mutator, with

$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$5

in the leading dependence on accuracy, together with agnostic evolvability, stability, and tolerance to target drift. The algorithm mutates by adding or subtracting one generator direction scaled by a mutation magnitude $\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$6, and performance is expressed through a Bregman divergence (Nock et al., 2017).

"Direct From Darwin: Deriving Advanced Optimizers From Evolutionary First Principles" (Grimmer, 6 May 2026) pushes the idea further by arguing that many familiar optimization methods can be interpreted as faithful simulations of Darwinian evolution once one adopts Darwinian Lineage Simulations (DLS). The leading-order DLS update is

$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$7

with DLS noise relation

$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$8

Within this framework, Stochastic Gradient Descent, Natural Gradient Descent, and Damped Newton’s method are DLS-compatible, while Adam requires a projection-based correction because inertial momentum is treated as an evolutionary impossibility (Grimmer, 6 May 2026).

A more radical departure from Darwinian filtering appears in "An evolutionary process without variation and selection" (Gabora et al., 2024). The paper argues that in cultural evolution and earliest life, acquired traits are retained and transmitted horizontally, so the Darwinian problem solved by natural selection is not the right explanatory frame. The proposed mechanism, Self–Other Reorganisation (SOR), is modeled with Reflexively Autocatalytic and Foodset-generated networks. In the community-based high-percolation regime, the authors present an existence proof that evolution is possible in the absence of variation, selection, and competition, and explicitly describe the special case of SOR without variation (Gabora et al., 2024).

These theoretical lines are not identical. Valiant-style evolvability and DLS still employ formal performance or fitness structure, whereas SOR argues that cumulative adaptive change can proceed through self-organization plus exchange rather than validation-like elimination of alternatives. A plausible implication is that “validation-free evolution” names a spectrum of mechanisms rather than a single theory.

6. Constraints, misconceptions, and open directions

A recurrent misconception is that validation-free evolution dispenses with checking. The literature instead shows repeated substitution of conventional validation with internalized control. SkillAudit uses a deterministic Anchor Verifier compiled once from the task specification and then fixed; any single skill_hurt report vetoes the update, and verifier failure triggers rollback (Gao et al., 12 Jun 2026). HOSO reserves exactly one example per class as a micro-validation cache within the original few-shot budget rather than introducing an external validation set (Vorster et al., 4 Mar 2026). In metrology, reference-free GIXRF is explicitly used for an independent validation of XRR, not as an excuse to avoid validation (Hönicke et al., 2019).

Another misconception is that intrinsic signals are equivalent to correctness. Squeeze Evolve is careful that confidence and diversity are only proxies for correctness, and notes that sparse verification or a lightweight correctness classifier could improve performance in the future. The method also depends on logprob access for the confidence-based version and is more limited in pure black-box closed-source settings (Maheswaran et al., 9 Apr 2026). IQ is likewise a proxy for trainability under a fixed FR protocol, not a universal quality score across all architectures, losses, or benchmarks, and it depends on proxy embeddings (Chen et al., 28 May 2026).

The software-evolution line identifies additional operational constraints. SecEr assumes deterministic executions in the same sense as classical regression testing, the current implementation limits the POI to variables, and extremely divergent program versions make POI matching harder. The paper suggests support for nondeterministic or concurrent behavior, richer traces, multiple POIs, and CI integration as future work (Insa et al., 2017).

CoAst acknowledges a tradeoff between assessment reliability and training efficiency: because parameter quantization is used, the training process may converge more slowly than standard FL, even though the pruning is intended only to improve valuation reliability (Wu et al., 2024). HOSO identifies a cache-size tradeoff: using more than one shot for ratio learning hurts because it steals data from adapter training, and the peak occurs at a cache size of one shot per class (Vorster et al., 4 Mar 2026).

The theoretical work also imposes strong conditions. DLS requires an asexual context, frequency-independent selection, a Gaussian small-variance approximation, positivity constraints on the noise covariance, and a projection-based repair for Adam (Grimmer, 6 May 2026). Valiant-style simulations reveal sensitivity to input distribution and the importance of neutral mutations, including a pronounced $$\mathrm{IQ} = \alpha c + \beta \tilde r_{\mathrm{ent}}, \qquad \alpha+\beta=1,$$9 bottleneck for general conjunctions and disjunctions in one empirical study (Fidalgo et al., 24 Jul 2025). SOR, by contrast, is presented as a primitive evolutionary process appropriate where there is no self-assembly code used in the distinct developmental and reproductive ways characteristic of standard biology (Gabora et al., 2024).

Taken together, these works suggest that validation-free evolution is best understood as a design principle for adaptive systems under missing external supervision: replace absent validation artifacts with intrinsic observables, fixed structural constraints, traceable first-principles quantities, or formally justified fitness signals; then evolve only within the limits those internal references can support. The central unresolved question is not whether validation can be removed in the abstract, but which internal signal is sufficiently informative, robust, and domain-faithful to stand in for it.