AI Programming Capability Leaps at Fixed Model Scale: Public Evidence and a Testable Hypothesis for ES-DSI as a Societal-Level Mechanism of Intergenerational Evolution

Abstract

Current progress in AI capabilities is commonly explained in terms of scaling, improvements in training algorithms, self-training, and inference-time compute. Yet software engineering exhibits a phenomenon that warrants separate study: even without increasing parameter count or architectural capacity—and, in some cases, while retaining the same pretrained base model—post-training can still produce gains of tens of percentage points on software-engineering benchmarks. In this paper, **fixed model scale** specifically means that parameter count and architectural capacity do not increase. It does not imply that model weights, effective capabilities, post-training compute, or inference strategies remain unchanged. Building on the Knowledge–Structure–Constraint (KSC) framework's distinction between existing knowledge and crystallized structures (K) and new structure generation (S), this paper provides complete definitions and operational accounts of DSI and ES-DSI. **Distributional Structural Internalization (DSI)** describes how the distribution of transferable problem-solving structures in a training environment is translated into a model's initial generative tendencies when it faces new, structurally related tasks. **External Selection–Distributional Structural Internalization (ES-DSI)** describes the intergenerational process through which candidate structures undergo differential retention in the real world, propagate, receive effective exposure in subsequent training, and may thereby alter the initial structural tendencies of later models. The empirical analysis uses progressively controlled, within-family generational comparisons: it controls parameter scale, architecture, base model, and selected training conditions in sequence and observes the capability gains that remain under those controls. The accompanying descriptive dataset currently contains 21 models or versions and 12 within-family longitudinal comparisons. In a representative case, DeepSeek V4 Flash retained the same architecture and size, and its developer explicitly stated that only post-training was rerun; in DeepSeek's official longitudinal DeepSWE evaluation, its score rose from 7.3 to 54.4, a gain of 47.1 percentage points. GLM-5.3 uses the same base model as GLM-5.2; its official DeepSWE v1.1 score rose from 46.2 to 66.9. As an order-of-magnitude reference, adjacent versions of several frontier models in DataCurve's standardized evaluation—for example, Gemini Flash and Claude Opus—showed DeepSWE gains of approximately 11–18 percentage points. Because longitudinal results across model families do not all come from the same evaluation implementation, and because this paper has not yet defined a “typical frontier-version gain” from a complete sample distribution, these cross-family differences serve only as order-of-magnitude references, not as strict cross-model performance comparisons. Further controlled studies show that, even when varying degrees of control are imposed on the model, training pipeline, or compute budget, differences in task construction, execution environments, and problem-solving trajectories can still substantially affect generalization to unseen tasks. These results are consistent with DSI's prediction that the distribution of structures in training affects subsequent tendencies in structure generation, but they do not yet directly measure changes in initial structural tendencies. Meanwhile, coding agents already participate in real-world software engineering at scale; observable instances have also emerged for each of the components required by ES-DSI: AI generation, external selection in engineering environments, propagation of digital structures, and subsequent training exposure. On this evidence, the paper advances the following central mechanistic claim: > **ES-DSI should be modeled and tested separately as a societal-level mechanism of AI intergenerational evolution, distinct from scaling, training-algorithm innovation, and closed-loop self-training.** Its basic temporal unit is not a single training run, but “model–real world–subsequent model”: AI systems and humans continually generate candidate problem-solving structures; real environments perform external selection; winning structures change the structural distribution of future digital environments; and subsequent models may convert those changes into new initial structural tendencies through DSI. Importantly, current public evidence has not yet traced the full causal chain—“generation → selection → propagation → subsequent training exposure → change in initial structural tendencies → performance change”—for the same set of structures. Performance leaps at fixed model scale therefore cannot be directly attributed to ES-DSI. These leaps show only that there is room for capability growth without further increases in model scale, while other evidence separately supports the real-world feasibility of each key ES-DSI component. These observations make a stronger hypothesis worthy of direct testing: > **ES-DSI may already be one of the principal drivers of growth in frontier AI coding and agent capabilities.** The direct evidence in this paper comes primarily from software engineering. Whether this claim generalizes to overall AI capability growth requires independent testing in mathematics, scientific reasoning, and other domains. Existing evidence also does not yet permit a decomposition of the respective contributions of ES-DSI, scaling, training algorithms, and inference-time compute. The claim is therefore explicitly presented as an observable, falsifiable, and quantifiable research hypothesis, not as an empirical conclusion already established by this paper.