Test-Time Compositional Generation

• Test-time compositional generation is a paradigm that dynamically combines independently trained models at inference to manage complex, multi-attribute constraints.
• It leverages techniques such as energy-based composition, modular attention, and product-of-experts to integrate concepts across visual, textual, and scientific domains.
• Despite its promise in improving generalization, challenges persist in scalability, memory efficiency, and data accuracy when handling tightly coupled or rare attribute combinations.

Test-time compositional generation refers to the algorithmic paradigm in which multiple models—often corresponding to distinct concepts, aspects, modules, or system components—are trained or acquired independently, and then combined at inference (test) time to satisfy rich compositional constraints, generalize to unseen combinations, or solve complex multi-constraint tasks. This paradigm is motivated by the observation that large neural models commonly fail to handle rare or combinatorial mixtures of attributes, relations, primitives, or control factors despite strong empirical performance on in-distribution examples. Test-time compositional strategies have been developed for visual, textual, scientific, and structured domains, leveraging modular training, energy-based combinations, optimization, and dynamic reasoning techniques.

1. Foundational Principles of Test-Time Composition

The central technical principle underlying test-time compositional generation is modularity: separate models or subcomponents encode distinct atomic concepts, attributes, aspects, or primitives. At inference, these pieces can be jointly deployed, combined, or adapted—often without retraining—so that the output satisfies the aggregate of user-specified constraints or attributes. This approach is deployed in several ways:

• Energy-Based Composition: In diffusion-based models, each concept can be represented as an energy-based model (EBM). The joint distribution over multiple concepts is formed via the product-of-experts identity, aggregating the energies, resulting in composite score functions for guidance in sampling (Liu et al., 2022).
• Modularized Attention and Routing: For structured reasoning, dynamic modular self-attention mechanisms select, route, and assemble relevant input features or reasoning steps into functional modules during inference; this enables out-of-distribution composition of logical chains or trees (Fu et al., 2023).
• Score, Critic, and Verifier Integration: In image/text/video domains, composite alignment objectives are constructed by evaluating global and element-wise similarity across models (vision-language critics, CLIP-style verifiers) to drive iterative correction, prompt re-writing, or synthesis loops (Sameti et al., 27 Sep 2025, Jaiswal et al., 21 Jan 2026, Qu et al., 9 Oct 2025).
• Product-of-Experts in Scientific Surrogates: In surrogate modeling for coupled scientific PDEs, separate field-models are combined through score-based alternating updates, leveraging decoupled training data to reproduce joint dynamics (Dhulipala et al., 23 Oct 2025).

2. Mathematical Formalizations and Algorithms

Test-time compositional generation is implemented via well-defined mathematical operations. For instance:

Diffusion Model Composition (Visual Generation):

Given conditional denoising networks $\epsilon_\theta(x_t, t|c_i)$ for concepts $c_i$, and an unconditional $\epsilon_\theta(x_t, t)$, the joint composite score for $K$ conjunctions is:

$$\hat{\epsilon}(x_t, t) = \epsilon_\theta(x_t, t) + \sum_{i=1}^K w_i [ \epsilon_\theta(x_t, t|c_i) - \epsilon_\theta(x_t, t) ]$$

This is used as the denoiser in the reverse diffusion step, yielding images faithful to all constituent concepts. Negative constraints are handled analogously by score subtraction (Liu et al., 2022).

Score-Based Coupling in PDE Surrogates:

Two field-specific models predict noise or interpolation variables (via $\epsilon$- or $v$-parameterization). At each reverse-diffusion step, alternating updates via

$$u_1^{t,k+1} = u_1^{t,k} + \lambda ( \hat{u}_1^0(u_1^{t,k}, u_2^{t,k}) - u_1^{t,k} )$$

allow recovery of joint (coupled) field dynamics from sides trained only on decoupled data (Dhulipala et al., 23 Oct 2025).

Modularized Structured Reasoning:

Reasoning trees are generated via a sequence-to-sequence model with dynamic module selection and input filtering. Each module (attention head $h_i$) routes only relevant inputs via learned representations and cosine similarity masks:

$$m_{i,j} = \begin{cases} \exp(1 - \cos(rep_{h_i}, s_j^{l-1})) & \text{if } \cos(rep_{h_i}, s_j^{l-1}) > \tau \ 0 & \text{otherwise} \end{cases}$$

This dynamic selection at test-time enables assembly of novel reasoning trees, unseen during training (Fu et al., 2023).

3. Domains and Representative Applications

Test-time compositional generation is documented across modalities:

Domain	Composition Mechanism	Citation
Visual generation	EBM/diffusion energy sum	(Liu et al., 2022)
Text-to-image/image	Concept-level CLIP scoring, iterative refinement	(Sameti et al., 27 Sep 2025, Jaiswal et al., 21 Jan 2026)
Surrogate modeling (PDEs)	Alternating fieldwise DDPM, PoE	(Dhulipala et al., 23 Oct 2025)
Reasoning/Explanation	Modular attention, dynamic routing	(Fu et al., 2023)
Data-to-text	Predicate clustering, clusterwise sentence generation	(Xu et al., 2023)
Multi-aspect text	Meta-learning simulating compositional generalization splits	(Zhong et al., 2024)
Video generation	Test-time LoRA optimization, spatiotemporal layout alignment, memory reuse	(Qu et al., 9 Oct 2025)
Industrial system test	Deductive spatial and temporal compositional proof schemes	(Ishii et al., 2021)

Specific applications include zero-shot scene synthesis, attribute disentanglement, relational scene generation, compositional review/text production, compositional video synthesis, rigorous modular test generation for industrial synchronous systems, scientific surrogate simulation under coupled PDEs, and modular compositional reasoning.

4. Experimental Benchmarking and Empirical Findings

Across each domain, benchmarking protocols assess compositional generalization—the ability to accurately generate, synthesize, or solve for attribute/constraint combinations never seen during training:

• Visual/T2I Benchmarks: GLIDE/CompBench/DrawBench/ConceptMix, with multi-object/rich-attribute prompts. Compositional diffusion summation and/or iterative refinement yield substantial accuracy/FID improvements over baseline diffusion, DALL-E2, or parallel sampling. E.g., iterative refinement attains +16.9% all-correct rate on 7-concept binding (Jaiswal et al., 21 Jan 2026); EBM composition yields +30% accuracy over vanilla EBM (Liu et al., 2022).
• Scientific Modeling: Compositional DDPM recovers coupled trajectory RMSE within 1–2 orders of magnitude of fully supervised FNO, using only decoupled training data (Dhulipala et al., 23 Oct 2025).
• Multi-aspect Text: Meta-MCTG meta-learning raises compositional-set attribute accuracy in 94.4% of settings (+3.64%) while maintaining fluency (constant perplexity). Benchmark splits using hold-out, ACD, and few-shot protocols, scoring compositional gap (Zhong et al., 2024).
• Structured Reasoning: MORSE outperforms competitive baselines (+3–5 points F1, +4–7 all-correct) especially for unseen reasoning tree lengths and shapes, with dynamic routing proven critical (performance drops -6.9 F1 when disabled) (Fu et al., 2023).
• Data-to-Text: CG-RL predicate clustering model achieves 31% relative improvement on faithfulness vs. T5 baselines, sharply reducing hallucinations and omissions (Xu et al., 2023).
• Compositional Video: TTOM TTO + memory mechanism achieves up to +82.6% relative accuracy gains for compositional motion on VBench; memory enables transfer of compositional layouts across prompt stream (Qu et al., 9 Oct 2025).
• Industrial Testing: Compositional deductive proof generation scales better than monolithic BMC or Simulink Design Verifier, with successful tests for complex systems (up to 281 blocks) unreachable by others (Ishii et al., 2021).

5. Limitations and Trade-offs

Despite compelling compositional generalization, current techniques exhibit notable limitations:

• Data Efficiency vs. Accuracy: Compositional surrogates using decoupled/synthetic modules trade off absolute accuracy relative to fully supervised joint training (FNO, T5), especially under tightly coupled or highly nonlinear dependencies (Dhulipala et al., 23 Oct 2025, Xu et al., 2023).
• Scalability in Reasoning: Combinatorial explosion may occur in structured reasoning with deep trees or lengthy chains, stressing dynamic routing/mask sparsity (Fu et al., 2023).
• Memory and Adaptation Constraints: In TTOM, parametric memory capacity and key abstraction bottleneck cross-task transfer; streaming protocols may force eviction of useful compositional patterns (Qu et al., 9 Oct 2025).
• Benchmarks and Evaluation: Current benchmarks (CompBench, ConceptMix, Fyelp, EntailmentBank) measure compositional gap but may not exhaustively cover systematicity/productivity/productivity axes for real-world complexity (Zhong et al., 2024, Fu et al., 2023).
• Few-shot Failure: Meta-MCTG cannot form pseudo-compositional splits when observed combinations are too sparse (Few-Shot setting) (Zhong et al., 2024).

A plausible implication is that further progress in compositional modeling is contingent on scalable module design, hierarchical abstraction in memory, adaptive coupling schemes, and more robust evaluation across combinatorial attribute spaces.

6. Cross-domain Extensions and Future Directions

The compositional paradigm is being continually generalized:

• Semantic Parsing, Multi-hop QA, Multi-document Summarization, Instruction Generation: Latent clustering, decomposition, and dynamic module assembly at test-time are being extended to more complex sequence and knowledge-graph tasks (Xu et al., 2023).
• Energy-based Compositionality: The product-of-experts framework enables plug-and-play composition across arbitrarily pretrained diffusion backbones, facilitating transfer and extrapolation beyond the training distribution (Liu et al., 2022).
• Memory-augmented Streaming: TTOM’s key-value LoRA parameter store suggests future multi-modal compositional generative architectures with lifelong memory, enabling both fast adaptation and generalization (Qu et al., 9 Oct 2025).
• Meta-Learning and Adaptation: Simulating compositional splits during training (Meta-MCTG) allows models to anticipate OOD compositions, raising generalization at test-time; this could be integrated with continual learning or federated multi-attribute learning (Zhong et al., 2024).

This suggests that test-time compositional generation will continue to be a unifying framework underpinning scalable, modular, and adaptable generative models across artificial intelligence research.