Dual-level Semantic Construction (DSC)

• Dual-level Semantic Construction (DSC) is a framework that integrates explicit fine-grained attribute extraction with holistic, high-level summaries for robust multimodal understanding.
• It leverages techniques like LLM-based attribute extraction, iterative template selection, and RL-gated fusion to enhance few-shot vision-language learning and neural radiance field synthesis.
• The approach unifies symbolic grammar rules with distributional semantic representations, enabling both rigid compositional processing and flexible, graded similarity evaluations.

Dual-level Semantic Construction (DSC) defines a class of architectures, algorithms, and formalisms across language, vision-language, and neural rendering that explicitly represent and process semantics at two distinct but complementary levels: a local/fine-grained attribute or supervision level, and a global/high-level summary or integration level. This approach is motivated by the inadequacy of methods relying on only a single semantic abstraction—either missing crucial nuanced cues or lacking coherent holistic structure. DSC modules, in diverse instantiations, have been shown to enhance few-shot vision-LLMs, improve neural radiance field synthesis in sparse regimes, and provide fine-grained, psycholinguistically plausible models for compositional and non-compositional language understanding (Li et al., 31 Jan 2026, Zhong et al., 4 Mar 2025, Blache et al., 2024, Lewis et al., 2016).

1. Foundational Principles and Motivation

DSC arose from the convergence of two needs: (1) to balance discriminative, instance-grounded local features with abstract, robust global representations, and (2) to unify symbolic and distributed representations in multimodal and language processing. In the vision-language domain, early methods incorporated only class-level text embeddings or attribute lists, leading either to missed subtle visual differences (if only global) or context fragmentation (if only local). DSC, as formalized in "DVLA-RL" (Li et al., 31 Jan 2026), addresses these issues by extracting both low-level discriminative attributes and high-level class descriptions, integrating them adaptively with vision features for refined grounding and holistic understanding.

Similarly, in neural rendering for few-view NeRF, the use of rendered semantics as both supervision and feature-level codebook guidance constitutes a form of DSC, achieving generalization from minimal data (Zhong et al., 4 Mar 2025). In linguistic modeling, frameworks such as Distributional Construction Grammars and DisCo models achieve DSC by unifying feature-structure grammars (symbolic) with vectorial or tensor-based distributional semantics, thus supporting both rigid composition and flexible, similarity-based reasoning (Blache et al., 2024, Lewis et al., 2016).

2. Formal Structure and Mathematical Workflows

DSC is realized through system-specific but structurally analogous workflows:

Vision-Language Few-Shot Learning

• Attribute Extraction: Given a support class $C^{\mathrm{sup}}$ and images $\{x_k\}$, a multimodal LLM generates short, fine-grained attributes $$\mathcal{A}^{C^{\mathrm{sup}}} = \mathcal{L}_e(P_{\mathrm{dis}}(C^{\mathrm{sup}}))$$.
• Progressive Selection: Attributes are iteratively scored via cosine similarity in a CLIP-based semantic space against an evolving template $T^{(i)}$, selecting top-k to form $\widehat{\mathcal{A}}^{C^{\mathrm{sup}}}$.
• Prompt Formation: Each selected $a_j$ is wrapped in a cross-modal prompt for shallow vision transformer layers: "A photo of a {CLASS}, which has {attribute}."
• Global Summary: The top-k attributes are summarized into a paragraph description $D^{C^{\mathrm{sup}}}$ via the LLM with a summarization prompt.

All steps are defined by explicit formulas: $$s_j^{(i)} = \cos\left(\mathrm{CLIP}(T^{(i)}),\,\mathrm{CLIP}(a_j)\right)$$ DSC outputs both $\widehat{\mathcal{A}}$ and $D$, which feed into an RL-gated fusion module. There is no DSC-specific training loss; integration is end-to-end (Li et al., 31 Jan 2026).

Dual-level Semantic Guidance for NeRF

• Supervision Level: Teacher NeRF renders dense-view semantic maps $\hat{S}_j$ which, after filtering by bi-directional geometric verification, are used as pseudo-labels for student NeRF training. Only "verified" pixels (validity mask $w(\hat{\mathbf{r}})$) contribute to the semantic loss.
• Feature Level: A codebook of learnable vectors is embedded in the student MLP. For each point, per-point features $\mathbf{f}$ attend to this codebook to form a semantic-relevant enhancement $\mathbf{f}_{sr}$, which is added to $\mathbf{f}$ before final predictions.

The total loss comprises RGB reconstruction, semantic cross-entropy (with BDV-masked pseudo-labels), and an optional depth penalty (Zhong et al., 4 Mar 2025).

Linguistic and Categorical Models

• Symbolic Level: Extended feature-structure or pregroup-grammar signatures encode morphosyntactic and logical dependencies, supporting classical unification and composition (Blache et al., 2024, Lewis et al., 2016).
• Distributional Level: Each sign or construction is additionally assigned a real-valued embedding (vector or tensor). Distributional similarity modulates activation and cue-based scoring in both parsing and interpretation: $$A_i = B_i + \sum_{c \in \text{cues}(i)} W_c\,F_c\,S_{c,i}$$

$$\text{Score}(i) = B_i + \sum_{c\in\mathrm{cues}(i)} W_c\,\mathrm{sim}(v_{\mathrm{inst}(c)},v_{\mathrm{proto}(c)})\,(\mathrm{MAS}-\ln(\mathrm{fan}_{c,i})) - P_u(i)$$

Integration with functorial mappings (e.g., from pregroup reductions to tensor contractions in FdVect, as in DisCo) enables composition of both grammatical and semantic meaning, with harmony scores measuring well-formedness (Lewis et al., 2016).

3. Supervision, Selection, and Integration Algorithms

DSC frameworks typically alternate, or interleave, symbolic or explicit attribute selection with graded, distributional, or data-driven integration. The mechanisms include:

• Iterative Template-based Selection: Progressive extraction and scoring of candidate attributes, refining semantic relevance at each step (Li et al., 31 Jan 2026).
• Bi-directional Verification: Geometry-based filtering of supervision signals, guarding against label noise and hallucination in teachers' outputs (Zhong et al., 4 Mar 2025).
• Attention over Codebooks: In neural rendering, codebooks at the feature level, equipped with attention, serve as inductive priors for expressing semantic regularities amid sparse supervision (Zhong et al., 4 Mar 2025).
• Activation/Unification Heuristics: In parsing, activation-based scoring guides the instantiation of symbolic constructions, with penalties for incomplete unifications but softening via distributional similarity (Blache et al., 2024).
• Harmony-based Grading: The DisCo model couples symbolic category reductions with vector-based computation, assigning a real-valued harmony as a graded judgment of compositionality and well-formedness (Lewis et al., 2016).

4. Representative Applications and Empirical Results

DSC advances multiple modalities:

Domain	Low Level	High Level	Integration Mechanism
Vision-language FSL (Li et al., 31 Jan 2026)	LLM-generated attributes	Synthesized class paragraph	RL-gated attention fusion
NeRF sparse-input (Zhong et al., 4 Mar 2025)	Per-pixel semantic labels	Semantic codebook	Masked loss + codebook attn
Distributional grammar (Blache et al., 2024)	Frame/role fillers, cues	Event or construction AVMs	Unification + vector sim.
DisCo/Harmony (Lewis et al., 2016)	Pregroup contractions	Sentence vector in V_s	Functorial mapping, H score

In "DVLA-RL", ablations isolate the impact of dual-level strategy: use of only attributes improves one-shot miniImageNet by 7.06%; addition of class description further increases accuracy; progressive selection yields an additional gain (+1.1% on CUB). Qualitative analysis (t-SNE plots) demonstrates better intra-class clustering and inter-class separation versus single-level baselines (Li et al., 31 Jan 2026).

In "Empowering Sparse-Input Neural Radiance Fields", feature-level guidance augments PSNR on ScanNet++ by +1.04 dB, outperforms InfoNeRF, DietNeRF, and FreeNeRF, and yields visually sharper boundaries and better color fidelity (Zhong et al., 4 Mar 2025).

Distributional Construction Grammar frameworks support incremental parsing with both compositional and non-compositional mechanisms, with activation-based thresholds enabling "fast-path" idiom recognition and soft constraint satisfaction by vector similarity (Blache et al., 2024).

5. Theoretical and Computational Implications

The DSC paradigm realizes a spectrum between compositionally rigorous, symbolic processing and context-adaptive, graded, distributional inference:

• Compositional vs. Non-Compositional Meaning: Symbolic unification and activation-based instantiation models both stepwise compositional build-up and direct, high-activation non-compositional retrieval (idioms, idiomatic patterns) (Blache et al., 2024).
• Gradient-based Evaluation: Harmony scores in DisCo models permit fine discrimination of nearly grammatical or ill-formed utterances, supporting gradient optimization in both grammar induction and learning (Lewis et al., 2016).
• Inductive Priors: Semantic codebooks and class descriptions serve as priors in vision-language and rendering, biasing learning towards transferable and robust representations even in data-scarce settings (Li et al., 31 Jan 2026, Zhong et al., 4 Mar 2025).
• Symbolic-Distributed Unification: The explicit coupling of AVM (Attribute-Value Matrix) feature structures or categorical grammars with vector spaces implements a form of integrated connectionist/symbolic computation.

6. Extensions and Open Directions

Prominent extensions include:

• Richer Algebraic Structures: Incorporating Frobenius algebras in the DisCo framework to encode complex compositional mechanisms (e.g., relative pronoun structures) (Lewis et al., 2016).
• Adaptive Fusion Policies: Reinforcement learning-based gates control layer-specific integration of dual-level semantics in vision transformers, enabling depth-aware alignment (Li et al., 31 Jan 2026).
• Threshold and Penalty Design: Flexible thresholds and penalties can control the balance between hard symbolic requirements and soft distributional matching, a crucial design axis in parsing and interpretation (Blache et al., 2024).
• Enhanced Semantic Supervision: Use of bi-directional geometric or contextual verification to filter pseudo-labels and attributes further mitigates the risk of hallucinated or irrelevant cues, especially as teacher models scale (Zhong et al., 4 Mar 2025).

A plausible implication is that DSC will continue to play a central role in architectures where compositionality, robustness to scarce data, and cross-modal integration are required. The paradigm also aligns with psycholinguistic findings on incremental and context-sensitive meaning construction and may inform future developments in grounded cognition and multi-agent communication protocols.