Neo Framework: Unified Learnable Architectures

• Neo Framework is a suite of advanced, learnable architectures that integrate heterogeneous data modalities to enable end-to-end optimization across databases, vision-language tasks, and orbit classification.
• It leverages deep neural networks, tree-structured convolutions, and modular abstractions to transform traditional query planning, native multimodal processing, and rapid NEO classification.
• Empirical benchmarks show that Neo systems achieve state-of-the-art performance with enhanced runtime efficiency, sample efficiency, and robustness, driving next-generation system interoperability.

The term "Neo Framework" refers to several advanced, technically rigorous software architectures and systems in distinct research domains, unified by an emphasis on learning-based approaches, end-to-end differentiability, and native integration of heterogeneous data modalities. The most prominent Neo frameworks in the literature encompass: (1) end-to-end learned query optimization for database systems (Marcus et al., 2019), (2) unified, dense native vision-language modeling (Diao et al., 16 Oct 2025), and (3) high-throughput short-arc orbit classification in Solar-System object surveillance via digest2 (where "NEO" denotes Near-Earth Object) (Keys et al., 2019). Each instantiation represents a paradigm shift in its respective field, implementing architectures characterized by modular abstraction, model-driven inference, and robust data-driven adaptation.

1. Neo in Learned Database Query Optimization

Neo as a learned query optimizer formalizes every component of classical Selinger-style query optimization—query representation, cost modeling, plan enumeration, and cardinality estimation—using deep neural networks trained to predict query plan costs directly from query-plan embeddings and observed runtime experience (Marcus et al., 2019). The input encoding exploits both query-level (join graph, predicates) and plan-level (one-hot join types, scan encodings per relation, subtree structure) featureizations. Queries and partial execution plans are mapped to high-dimensional vector embeddings, concatenated with query context representations via a small MLP.

The core cost model employs a stack of tree-structured convolution ("TreeConv") layers over the plan forest, followed by dynamic pooling and a final MLP head to estimate the "predicted best remaining cost" for a given partial plan. Bootstrapping is performed against expert optimizers such as PostgreSQL, with further refinement through continual online reinforcement, where the model is updated using observed execution latencies via value iteration on the plan search lattice.

The plan enumeration engine uses best-first (anytime) search guided by the neural cost model $M(P)$ within strict time budgets, employing "hurry-up" mode if exhaustive search is infeasible. Neo integrates with any target DBMS by treating the engine as a black box, using plan hints for join order, operators, and scan type. Empirical benchmarks demonstrate that Neo, even when initialized on simple open-source optimizers, can reach or surpass the runtime performance of state-of-the-art commercial optimizers such as Oracle and MS SQL Server, with substantial gains in adaptation to new workloads and data distributions. The model is resilient to cardinality-estimation errors and adapts rapidly to unseen query structures after several online learning epochs.

2. Neo Framework for Native Vision-Language Modeling

NEO in multimodal AI denotes a family of natively unified vision-LLMs, distinguished by structurally dense, modality-agnostic, single-backbone systems where pixel and word representations are integrated from the earliest embedding layers (Diao et al., 16 Oct 2025). In contrast to modular VLMs relying on late-fusion or adapter strategies, the NEO framework enforces native co-processing by:

• Employing a lightweight Patch Embedding Layer (PEL) using strided convolutions and 2D sinusoidal positional encodings for images, in parallel with a Word Embedding Layer (WEL) for tokenized text.
• Concatenating these streams into a shared decoder-only Transformer backbone, which is segmented into modality-agnostic "pre-Buffer" and "post-LLM" stacks during pre-training and later collapsed to a monolithic stack for inference.
• Augmenting Transformer blocks with Multi-Head Native Attention (MHNA) incorporating distinct temporal, height, and width attention heads; Native Rotary Position Embeddings (RoPE) that disentangle index frequencies across T/H/W axes; and a mixed-mask mechanism that applies causal masking to text tokens but enables fully bidirectional attention for visual patches.

The pre-training pipeline involves freezing LLM weights and updating only embedding, pre-Buffer, and enlarged Q/K projections using an autoregressive next-token loss over approximately 390 million image-text pairs, with mid-training and supervised fine-tuning on curated, high-fidelity annotated corpora.

Empirical evaluation reveals that NEO, at both 2B and 8B parameter scales, matches or exceeds the capabilities of top modular VLMs (Qwen2-VL, InternVL2.5, Qwen2.5-VL, InternVL3), frequently utilizing orders-of-magnitude less data. NEO's design emphasizes reusability, with primitives (patch/word embeddings, mixed attention kernels, RoPE modules, pre-Buffer checkpoints) open-sourced for rapid research extension and interoperability. This architecture promotes democratic accessibility to unified, intrinsically multimodal modeling.

3. Neo in Solar System Object Classification (digest2)

In the context of the digest2 software package, "NEO" stands for Near-Earth Object and the framework refers to a robust, high-throughput classifier for Solar-System small-body orbit determination from observational tracklets (Keys et al., 2019). digest2’s main output, the D₂ score, is a Bayesian-analog pseudo-probability that classifies tracklets into orbit classes (NEO, Mars-Crosser, Main-Belt, Trojan, etc.) based on population models binned in perihelion distance $q$, eccentricity $e$, inclination $i$, and absolute magnitude $H$. The classifier supports rapid, parallelized processing (O(10²–10³) orbit solves per tracklet) with per-tracklet runtimes on the order of ms.

The architecture synthesizes two-point "motion vectors" from an arc by percentile endpoint selection and dithering within astrometric uncertainties, then samples admissible orbital elements via recursive binary-subdivision in (ρ, α) to identify population bins. For each class $C$, the D₂ score is calculated as $$100 \cdot \Sigma_C / (\Sigma_C + \Sigma_{\text{notC}})$$, where $\Sigma_C$ sums the population bins for in-class candidate orbits. This enables robust triage of new tracklets for NEO candidate status, underpinning processes such as the MPC Near-Earth Object Confirmation Page.

Performance metrics demonstrate 14% false negative and 1.5% false positive rates at threshold D₂=65 for real discovery data. The framework is sensitive to rate of motion, ecliptic latitude, solar elongation, and apparent magnitude. Recommendations for future upgrades focus on continuous population models, advanced uncertainty propagation, and seamless integration via Python APIs and containerization.

4. Architectural and Methodological Innovations

Across all instantiations, the Neo framework is characterized by three essential methodological advancements:

• End-to-End Learning: Each framework replaces legacy expert-driven heuristics, decoupled pipelines, or hand-tuned modules with unified, differentiable models—neural networks (Neo/Neo-VLM), Bayesian pseudo-probabilities (digest2), or deep equilibrium surrogates (in variants such as NEO-Grid, not detailed here).
• Hybrid Data Representation: The frameworks are designed for direct integration of heterogeneous or multi-domain features, such as fusing imagery and text at the lowest representational level (NEO VLM), or amalgamating query graphs and execution plans in dense spectral spaces (Neo DB optimizer).
• Extensible, Performance-Driven Abstractions: The public exposure of reusable primitives—embedding kernels, attention schemes, dynamic masking strategies, modular population models—facilitates algorithmic extensibility and rapid adaptation to new research tasks.

5. Empirical and Benchmark Performance

Empirical results for Neo architectures consistently demonstrate rapid convergence to or beyond state-of-the-art across industry-standard benchmarks:

Domain	System/Task	Key Metrics / Results	Reference
Database	JOB, TPC-H, Corp workloads	Neo yields 0.6× PostgreSQL latency, matches/conquers Oracle/MS SQL after bootstrapping	(Marcus et al., 2019)
Vision-Language	MMMU, MMB, AI2D, DocVQA, ChartQA	NEO (2B,8B) rivals InternVL2.5, Qwen2.5-VL with less data	(Diao et al., 16 Oct 2025)
Solar System	NEO triage, D₂≥65 threshold	94%–99% NEO detection completeness, 1.5% false positives	(Keys et al., 2019)

A plausible implication is that the Neo design approach—a shift toward natively unified, learnable primitives—yields not only theoretical elegance but practical gains in sample efficiency, generalization, and system interoperability.

6. Limitations and Future Directions

Neo-based systems, though performant, retain limitations endemic to data-driven, end-to-end architectures. For database optimization, recovery from new query distributions initially requires several online epochs; in VLMs, OCR-specific accuracy can lag specialized decoders, though mitigated by end-to-end fine-tuning. In orbit classification, discrete population bins can induce "sampling jitter," motivating a transition to continuous probabilistic density modeling.

Ongoing recommendations target (a) full uncertainty propagation with covariance tracking, (b) plugin APIs for rapid workflow integration, and (c) broadening coverage through richer population priors or multi-modal fusion mechanisms. The release of codebases and modular checkpoints fosters community-driven evolution of both theory and practical solutions.