Test-Time Scaling in Multimodal Foundation Models: A Comprehensive Survey of Generation and Reasoning

Conference: ACL 2026
arXiv: 2606.08231
Code: TBD
Area: Multimodal VLM / Test-Time Scaling / Survey / Multimodal Reasoning / Multimodal Generation
Keywords: Test-Time Scaling, Multimodal Foundation Models, Sampling, Feedback, Search, Survey

TL;DR

The first survey dedicated to Test-Time Scaling (TTS) in Multimodal Foundation Models (MFM): it unifies various methods of "dynamically allocating compute at the inference stage" into a framework of \(\pi^*=\arg\max_\pi \mathbb{E}[U(x,y)]\) s.t. compute budget constraints. It categorizes these into three paradigms: Sampling-based, Feedback-based, and Search-based, covering both multimodal generation and reasoning tasks, and provides a roadmap of representative methods, benchmarks, and open challenges.

Background & Motivation

Background: The capabilities of foundation models have primarily relied on scaling parameters, data, and compute during pre-training (scaling laws). However, marginal returns from training-side scaling are diminishing, shifting the focus to "extracting the potential of pre-trained models during the inference stage." TTS (Test-time Scaling) follows this path: instead of updating parameters, it trades additional compute at inference (multiple samplings, searching, verification, iterative refinement) for higher performance. While proven effective in LLMs through search/sampling/verification, it is now expanding into Multimodal Foundation Models (MFM).

Limitations of Prior Work: MFM TTS research is exploding, yet it lacks a unified taxonomy and theoretical framework. Existing surveys almost exclusively cover LLM TTS. Multimodal components (image/video generation, multimodal reasoning) are scattered with inconsistent terminology, leaving new researchers without a clear map.

Key Challenge: Multimodal TTS appears to mirror LLM TTS but is inherently more difficult. While text requires compute allocation along a single modality's reasoning chain, MFMs must scale compute across perceptual evidence, spatial grounding, and temporal context. Evaluating intermediate steps requires strict cross-modal consistency (not just internal textual logic, but faithfulness to visual and spatial relations), and the modality gap often necessitates additional VLMs or reward models for scoring.

Goal: (1) Provide the first systematic survey of TTS on MFMs; (2) propose a unified taxonomy to clarify mechanisms and applicability; (3) summarize benchmarks and identify open challenges as a roadmap for future research.

Key Insight: The authors formalize TTS as "selecting an inference pipeline \(\pi\) for a fixed model to maximize task utility within a compute budget," then converge diverse methods into three main branches categorized by application (multimodal generation vs. multimodal reasoning).

Core Idea: Organize all MFM TTS work using three mechanisms of "how to spend test-time compute": Sampling-based, Feedback-based, and Search-based.

Method

Overall Architecture

The survey formalizes TTS as: given fixed parameters \(\theta\), select an inference pipeline \(\pi\) to query the model, maximizing expected utility within a test-time budget \(B\):

\[\pi^*=\arg\max_{\pi}\ \mathbb{E}_{y\sim\pi(\cdot\mid x,\theta)}[U(x,y)]\quad \text{s.t.}\ C(\pi,x)\le B,\ \theta\ \text{fixed}.\]

Where \(x\) is the input, \(y\) is the output, \(U\) is the task utility, and \(C(\pi,x)\) is the compute cost of applying \(\pi\) to \(x\). This formula highlights that TTS scales the inference process rather than model parameters, allowing adaptive compute. The authors distinguish between three test-time resources: Compute (the core of TTS), Memory/State (retrieval banks, episodic memory, KV cache), and Weights (test-time training/adaptation via gradients). This survey focuses on compute-centric, parameter-static methods, treating memory and weight updates as auxiliary. Two base architectures support MFM TTS: MLLMs (autoregressive processing of vision/audio/video tokens, supporting CoT and multi-step verification) and Diffusion Models (iterative denoising + CFG, allowing trade-offs between sampling steps/candidates and fidelity).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multimodal TTS<br/>Fixed Params · Inference Compute Allocation"] --> B["Sampling-based<br/>Parallel Candidates + Aggregation/Selection"]
    A --> C["Feedback-based<br/>Auxiliary Signals for Guidance/Correction"]
    A --> D["Search-based<br/>Structured Exploration of Solution Space"]
    B --> B1["Best-of-N + Majority Voting"]
    C --> C1["Reward Models ORM/PRM + Iterative Refinement"]
    D --> D1["Beam Search + Tree Search/MCTS + Heuristic Adaptive"]

Key Designs

1. Sampling-based: Parallel Candidates + Aggregation/Selection

The most direct way to "spend compute"—generating multiple solutions in parallel to explore a larger space. Two sub-paths: Best-of-N (BoN) uses a scoring function or "MLLM-as-a-judge" to evaluate \(N\) candidates and pick the best (e.g., TTGen uses CLIP to pick optimal latents per diffusion step; SANA-1.5 uses tournament comparison; CoDe uses local BoN to save cost; UniGen uses CoT verification). Majority Voting aggregates candidates to find the most consistent output without an external verifier (e.g., CoT-Vid uses path clustering for video reasoning consistency; Video-RTS uses multi-path voting; RoboMonkey votes on VLA actions under Gaussian noise).

2. Feedback-based: Guidance via Auxiliary Evaluation Signals

Filters, guides, or corrects outputs using external signals during inference. Two sub-paths: Reward Models are split into Outcome Reward Models (ORM) and Process Reward Models (PRM)—the former evaluates final results (e.g., Guo et al. using LLaVA-OneVision as a zero-shot ORM), while the latter provides step-by-step feedback for search (e.g., Athena using consistency labels, RoVer refining 6D orientation via PRM, VisualPRM as a BoN verifier). Iterative Refinement emphasizes a "Generate-Evaluate-Refine" loop (e.g., Reflect-DiT combines VLM feedback with history to guide image editing; CyberV corrects attention drift via sensor-controller feedback; GenPilot iteratively rewrites prompts via visual verification).

3. Search-based: Structured Exploration in Solution Space

Models inference/generation as an explorable structure, using pruning, backtracking, or dynamic scheduling. Three sub-paths: Beam Search maintains the Top-K trajectories (e.g., LLaVA-CoT generates per-stage candidates with backtracking; MindJourney uses world models for spatial reasoning). Tree Search / MCTS expands reasoning into trees with self-rewards (e.g., VReST uses MCTS; Visuothink performs vision-language tree search with rollback; ZoomEye uses hierarchical search for perception; VLA-Reasoner optimizes actions via MCTS + world models). Heuristic / Adaptive Search provides flexibility (e.g., evolutionary search for gradient-free alignment; adaptive diffusion cycles that allocate compute as needed).

Application & Tasks

The survey cross-cuts the three paradigms across two task categories: Multimodal Generation (Image/Video generation, alignment, UI-to-Code, where TTS manifests as candidate filtering and iterative prompt/image refinement) and Multimodal Reasoning (Math, spatial, video, embodied QA, VLA, where TTS involves process rewards, tree search, and consistency voting).

Key Experimental Results

This is a survey; no new experiments were conducted. The tables below summarize representative methods and mechanisms.

Comparison of the Three Paradigms

Paradigm	Core Mechanism	Verifier/Reward Required?	Representative Methods
Sampling-based	Parallel candidates + aggregation	Scoring for BoN; None for Voting	TTGen, SANA-1.5, CoDe, UniGen, CoT-Vid
Feedback-based	Guidance via evaluation/correction	Requires ORM/PRM or VLM feedback	Reflect-DiT, CyberV, Athena, RoVer, VisualPRM
Search-based	Structured exploration (Pruning/Backtracking)	Often pairs with PRM or World Models	LLaVA-CoT, MindJourney, Visuothink, VLA-Reasoner

TTS Focus: Generation vs. Reasoning

Task Category	Typical Scenarios	Common TTS Techniques	Evaluation Focus
Multimodal Generation	Image/Video Gen, Alignment, UI/Charts	BoN on latents, Iterative prompt/image refinement	Cross-modal fidelity & alignment
Multimodal Reasoning	Math/Spatial/Video/Embodied QA, VLA	Process rewards + Tree Search, Voting	Cross-modal faithfulness of steps

Key Findings

• Multimodal TTS is harder than pure text TTS: It requires scaling compute across perception, spatial grounding, and temporal context simultaneously. Intermediate verification requires cross-modal faithfulness, often necessitating external VLMs.
• Paradigms are not mutually exclusive: Many works use hybrid approaches (e.g., Tree Search with PRM, BoN with CoT verification).
• Diffusion and MLLMs are the primary carriers: Diffusion's iterative nature supports step/candidate trade-offs, while MLLM's autoregressive CoT facilitates multi-step verification and search.

Highlights & Insights

• Unified Formalization \(\pi^*=\arg\max_\pi \mathbb{E}[U]\) s.t. \(C\le B\): Frame's test-time compute as an inference pipeline selection problem under budget constraints, providing a common coordinate system.
• Clear Resource Boundaries: Explicitly distinguishes "parameter-invariant compute scaling" from memory-augmented and test-time training methods.
• Decomposition of "Multimodal Difficulty": Breaking down difficulty into perception, grounding, and temporal dimensions provides a checklist for designing new methods.
• Two-dimensional Organization: Allows readers to find methods either by mechanism (Sampling/Feedback/Search) or by task (Generation/Reasoning).

Limitations & Future Work

• Open Challenges Identified by Authors: Lack of unified multimodal TTS benchmarks, difficulty in acquiring high-quality verification signals, and lack of theoretical characterization for the compute-utility trade-off.
• Self-identified Limitations: The survey focuses on categorization; it lacks quantitative horizontal comparisons of the three paradigms on a single benchmark. Heuristic search categories remain somewhat heterogeneous.
• Future Directions: Developing a unified evaluation to quantify "utility gain per unit of compute" and exploring the synergy between memory-augmented and compute-centric TTS.

Related Work & Insights

• vs. LLM TTS Surveys (e.g., Zhang et al. 2025c, Ji et al. 2025): These only cover text-based LLMs. This is the first systematic survey for Multimodal Foundation Models.
• vs. Single-paradigm Papers: This survey integrates scattered techniques (BoN, PRM, MCTS) into a unified Sampling-Feedback-Search map.
• vs. Test-Time Training: This paper clearly excludes parameter updates from the core definition of "compute-centric TTS," refining the conceptual scope.

Rating

• Novelty: ⭐⭐⭐⭐ First MFM TTS survey with a unified formalization and three-paradigm taxonomy.
• Experimental Thoroughness: ⭐⭐⭐⭐ Extensive coverage of methods and benchmarks (though lacks unified quantitative comparison).
• Writing Quality: ⭐⭐⭐⭐ Clear formalization, distinct resource boundaries, and intuitive structure.
• Value: ⭐⭐⭐⭐ Provides a clear roadmap and self-check criteria for the rapidly expanding MFM TTS field.

Related Papers

• [CVPR 2026] Scaling Spatial Intelligence with Multimodal Foundation Models
• [NeurIPS 2025] TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning
• [CVPR 2026] TTRV: Test-Time Reinforcement Learning for Vision Language Models
• [CVPR 2026] Controllable Federated Prompt Learning at Test Time
• [CVPR 2026] Test-Time Attention Purification for Backdoored Large Vision Language Models