This survey is motivated by three specific concerns, detailed below:

(1) The definition of complex logical reasoning in multimodal settings is often ambiguous, with unclear boundaries relative to perceptual understanding, commonsense reasoning, and mathematical problem solving [11,25,26].

(2) Meanwhile, existing research lacks a unified evaluation framework. Current benchmarks and methodologies are scattered and typically focus on final-answer accuracy, such as OlympiadBench [27] and TheoremQA [28], offering limited insight into reasoning faithfulness and intermediate inference processes [29–31].

(3) In addition, this field is rapidly evolving, with architectural innovations, prompting strategies, reasoning mechanisms, and evaluation protocols emerging at an accelerating pace [12,32]. However, these techniques are frequently assessed in isolated settings, making systematic comparison across methods difficult [24,29].

To address these issues, this survey focuses specifically on complex logical reasoning in LVLMs, distinguishing it from related areas such as mathematical reasoning, commonsense reasoning, and visual perception [11,26,33]. We define it as multimodal reasoning in which the final answer cannot be obtained by direct perceptual lookup alone, but instead requires at least one inferential transformation such as explicit logical operations, multi-step dependency tracking, abstract relation handling, or cross-source evidence composition [10,23,34–36]. We primarily cover work published from 2022 to early 2026, while retaining foundational earlier studies for historical context [22,33,37–39]. Our survey covers five primary dimensions: conceptual foundations, architectural paradigms, reasoning mechanisms, evaluation benchmarks, and empirical findings, along with discussions of current challenges and future directions [12,24,25].

The main contributions of this survey are summarized as follows: 1.

Problem formalization and conceptual taxonomy. We clarify the scope of complex logical reasoning in LVLMs and distinguish it from visual perception, commonsense reasoning, mathematical reasoning, and factual recall. We further organize existing studies into five recurrent reasoning families under a hierarchical taxonomy: deductive, inductive, abductive, multi-hop, and causal reasoning [10,21,23,34,35].

In addition to the above contributions, this survey also provides extensive references to foundational and state-of-the-art works, serving as a comprehensive resource for researchers entering the field [2,3,50]. We organize our discussion around clear conceptual distinctions and provide examples drawn from recent literature to illustrate key points [12,24,25].

We conducted a structured search across major academic and open-access platforms, including Google Scholar¹1

Google Scholar: https://scholar.google.com

The screening process consisted of three stages: (1) initial collection via database search and reference expansion; (2) title and abstract filtering using the above inclusion/exclusion rules; and (3) full-text review for technical relevance, contribution type, and fit to one or more survey dimensions. The retained studies were then organized into conceptual/foundational works, architecture-oriented works, reasoning-mechanism works, and benchmark/evaluation works. Because this survey was developed as a structured narrative review rather than a preregistered systematic review, and because the earliest candidate pool was expanded iteratively during drafting and backward snowballing, exact stage-wise exclusion rates were not preserved in a form suitable for retrospective reporting. We therefore report the final corpus composition as a transparency summary rather than claiming PRISMA-style stage counts for the earliest drafting stage.

Following this process, the retained corpus comprises 138 final cited references and spans international conferences, journal publications, and a large proportion of preprints, reflecting the fast-evolving nature of the field. Most works were published between 2023 and 2025.

For greater quantitative transparency, we further summarize the retained corpus here. The current manuscript cites 138 unique works in its final reference list. The cited corpus is strongly recent: 101 of the 138 cited works (73.2%) were published between 2023 and 2025. To summarize topical coverage without forcing each paper into a single exclusive bin, we report overlapping section-level counts: 58 architecture-related citations, 42 reasoning-mechanism citations, and 56 benchmark/evaluation citations, in addition to conceptual and survey papers used for problem framing. This composition improves coverage of a rapidly moving field, but it also introduces clear coverage biases: the survey favors recent English-language public papers, benchmark-heavy studies, and open preprints, while closed industrial evaluations, negative results, and papers with limited methodological disclosure are more likely to be underrepresented.

The remainder of this survey is organized as follows. Section 2 introduces conceptual foundations of complex logical reasoning in LVLMs and presents the proposed taxonomy. Section 3 examines LVLM architectures designed for logical reasoning. Section 4 reviews reasoning mechanisms and strategies. Section 5 summarizes benchmarks and evaluation protocols Section 6 discusses challenges and open problems. Section 7 outlines future research directions. Section 8 concludes the survey.

Complex logical reasoning in LVLMs differs fundamentally from pure textual reasoning and simple visual perception. The core distinction lies between perceptual-level and logical-reasoning-level understanding. Perceptual-level understanding involves recognizing objects, identifying attributes, and describing scenes, tasks that modern LVLMs perform with remarkable accuracy [5,40,52,53]. In contrast, complex logical reasoning requires manipulating abstract representations, applying inference rules, and deriving conclusions not explicitly given in the input [1,21,23]. Fig. 2 illustrates this transition from perceptual recognition to higher-level multimodal inference.

The distinction between perceptual and complex logical reasoning can also be understood through the lens of cognitive hierarchies. Perceptual cognition operates on sensory data to extract features, objects, and basic relations. Logical cognition operates on these extracted representations to reason about implications, dependencies, and abstract relationships. While perception is often relatively direct for both humans and AI systems, complex logical reasoning typically requires deliberate multi-step inference and may rely on explicit or implicit logical rules [11,25].

Complex logical reasoning in LVLM contexts exhibits several key characteristics that distinguish it from simpler forms of multimodal understanding: 1.

Multi-Step Nature: Complex logical reasoning typically requires chaining multiple inference steps, where each step builds upon previous conclusions. This property distinguishes logical reasoning from single-hop inference, where conclusions can be derived directly from premises without intermediate steps. Multi-step reasoning is essential for tasks such as solving complex puzzles, analyzing scenarios with multiple constraints, and answering questions that require integrating information from different sources [1,54–57].

Operationally, this survey treats a task as complex logical reasoning only when the final answer cannot be obtained through a single perceptual lookup or plain evidence retrieval, but instead requires at least one non-trivial inferential transformation over multimodal evidence. Such transformation may take the form of rule application, hypothesis selection, causal attribution, constraint satisfaction, or cross-source composition. Under this criterion, multi-step length alone is not sufficient: a retrieval pipeline that merely gathers several relevant facts without comparing, composing, or logically constraining them is not counted as complex logical reasoning in our taxonomy. By contrast, a multi-hop setting falls within scope when intermediate evidence must be integrated into a new conclusion whose validity depends on the consistency of the whole inference chain [54–56].

Drawing upon established frameworks from cognitive science and logic, we organize the literature into five recurrent reasoning families: deductive, inductive, abductive, multi-hop, and causal reasoning [11,25]. We do not treat these families as a strictly flat or mutually exclusive partition. Instead, we use a hierarchical reading of the taxonomy: deductive, inductive, abductive, and causal reasoning are operator-centered categories defined by the dominant inferential goal, whereas multi-hop reasoning is a structural category indicating that evidence or sub-goals must be chained across several steps. This framing reflects actual LVLM research more faithfully, because a model may be abductive and multi-hop at the same time, or causal and multi-hop, without creating a contradiction in the classification [24]. Fig. 3 summarizes this hierarchical organization of reasoning families.

To improve the comparability of related studies, Table 1 links each family to its operational definition and typical task settings. We assign representative works according to their dominant inference objective: formal entailment for deductive reasoning, rule abstraction for inductive reasoning, explanation generation for abductive reasoning, intervention or cause-effect analysis for causal reasoning, and cross-step evidence chaining for multi-hop reasoning. When a work clearly spans multiple families, we use its main evaluation target as the primary label and discuss the overlap explicitly in the surrounding text.⁹9

The representative works are categorized by their dominant inference objective rather than by an assumption of mutual exclusivity. In particular, multi-hop reasoning is treated as a cross-cutting structural pattern that may co-occur with deductive, abductive, or causal reasoning.

Deductive Reasoning. Deductive reasoning involves deriving specific conclusions from general premises through the application of formal logical rules [21,51]. In LVLM contexts, deductive reasoning typically appears in tasks requiring rule application, premise validation, or logical entailment determination [64]. For example, given a set of premises about relationships between objects and a query about a specific relationship, deductive reasoning enables the model to infer the answer by applying the relevant rule structure.

The formal nature of deductive reasoning makes it particularly amenable to neuro-symbolic approaches that combine neural representation learning with symbolic inference engines [41,51]. Works such as ReClor [64] have developed benchmarks specifically targeting deductive reasoning capabilities in language models. ReClor presents reading comprehension questions that require complex logical reasoning to determine entailment or contradiction relationships between premises and conclusions.

A key challenge in deductive reasoning for LVLMs lies in handling the semantic gap between natural language premises and formal logical representations [62]. Natural language is often ambiguous and may not map cleanly to formal logical forms. Furthermore, real-world premises may be incomplete or may contain implicit assumptions that must be identified for valid deduction.

Deductive reasoning can be further subdivided into propositional deduction (reasoning about the truth values of propositions) and predicate-logic deduction (reasoning about object properties and relations). In LVLM contexts, both forms are relevant, with predicate-level deduction being particularly important for reasoning about visual scenes that involve multiple objects and their interactions [74].

Inductive Reasoning. Inductive reasoning involves discovering general patterns or rules from specific instances or observations [11]. This form of reasoning is fundamental to scientific discovery and learning from experience. In LVLM contexts, inductive reasoning tasks include completing visual patterns, generalizing from examples, and identifying underlying rules.

The RAVEN benchmark [22] represents a significant contribution to evaluating inductive reasoning in visual contexts, presenting matrices in which models must identify underlying transformation rules. RAVEN includes visual matrices with varied structures, requiring models to understand relations between elements and apply transformation rules to complete the pattern.

Inductive reasoning in visual contexts presents unique challenges due to the complexity of visual representations and the need to handle variations in style, viewpoint, and context [46]. Unlike formal inductive reasoning in logic, visual induction often requires dealing with noisy, high-dimensional input and may involve probabilistic rather than deterministic generalization.

Inductive reasoning can be characterized by several key properties: the resulting generalization is uncertain (the conclusion may be false even if the observations are correct), the reasoning moves from specific observations to broader rules, and the strength of the generalization depends on the representativeness of the observed instances [11]. These properties make inductive reasoning fundamentally different from deductive reasoning and require different computational approaches.

Abductive Reasoning. Abductive reasoning, also known as inference to the best explanation, involves inferring the most plausible explanation for observed phenomena [9,23,75]. In LVLM contexts, abductive reasoning is particularly relevant for tasks such as visual explanation generation, where models must infer what events or conditions could have produced an observed scene.

Recent works such as KN-VLM [76] have focused on visual abductive reasoning, developing models that can infer hypotheses to explain visual contexts. This capability is essential for applications such as accident analysis and safety monitoring, where systems must infer what events could have led to an observed situation.

Abductive reasoning is particularly challenging because it requires both understanding observed phenomena and generating plausible explanations that account for them [23]. Unlike deductive reasoning, where conclusions follow necessarily from premises, abductive reasoning involves selecting the most plausible explanation among multiple candidates, often under incomplete or uncertain information.

Multi-Hop Reasoning. Multi-hop reasoning refers to the structural requirement that evidence or sub-goals be chained across multiple steps [54,56,57]. In this survey, it is treated as a cross-cutting reasoning pattern rather than a purely separate operator on par with deduction or abduction. A task is multi-hop when the model must maintain intermediate states and compose dispersed evidence before the final answer can be derived.

The CLEVR benchmark [39] represents a foundational contribution to multi-hop reasoning evaluation in visual question answering. CLEVR presents synthetic scenes with multiple objects and asks questions that require multi-step reasoning to answer. For example, a question might ask, “What is the color of the object that is to the left of the red sphere?”, which requires first identifying the red sphere, then finding objects to its left, and finally extracting the relevant color attribute.

This distinction is important for separating reasoning from simple retrieval. If a system merely fetches several relevant snippets or regions and one of them directly contains the answer, the process is multi-source retrieval but not necessarily multi-hop reasoning. By contrast, reasoning is required when retrieved pieces must be compared, composed, or constrained to infer a conclusion that is not explicitly stated in any single source. Multi-hop reasoning is therefore particularly relevant for real-world applications in which questions require integrating information across different parts of an image or across multiple images [36]. The ability to maintain intermediate conclusions and chain reasoning steps is essential for complex question answering, narrative understanding, and comprehensive scene analysis.

Causal Reasoning. Causal reasoning involves identifying causal relationships between events or variables, distinguishing correlation from causation, and predicting the effects of interventions [10,34,35]. This form of reasoning is essential for understanding how changes in one variable affect others and for predicting the outcomes of actions.

Recent works such as the Multimodal Causal Reasoning Benchmark [10] have developed benchmarks specifically targeting causal reasoning in multimodal contexts. These benchmarks challenge LVLMs to identify causal links across modalities, requiring the joint understanding of visual and textual information.

Causal reasoning is fundamental to many real-world applications, such as autonomous driving (predicting the effects of actions) [14]. The ability to distinguish causal relationships from simple correlations is a key capability that separates robust reasoning systems from superficial pattern-matching systems.

Causal reasoning in LVLMs faces unique challenges due to the need to handle both visual and textual information about causal relationships [10,34,35]. Visual information may show how events unfold over time, while textual information may describe causal mechanisms or provide explanatory context. Integrating these modalities to perform causal reasoning requires sophisticated representation learning and inference capabilities [10].

To avoid ambiguity in the subsequent taxonomy, we first clarify the conceptual boundaries between complex logical reasoning and related paradigms. Despite the increasing use of the term complex logical reasoning in LVLM research, its conceptual boundaries remain ambiguous and often overlap with related notions such as commonsense reasoning, mathematical reasoning, and visual perception [26,33,77]. This ambiguity may lead to inconsistent evaluation and weaken the validity of taxonomic frameworks. To address this issue, we provide a structured clarification of how complex logical reasoning differs from, and interacts with, these closely related paradigms.

We define complex logical reasoning as the ability to perform multi-step, structured inference over multimodal inputs, where intermediate reasoning steps are required to derive the final answer [1,25]. More specifically, a task falls into this scope when it requires at least one inferential transformation beyond direct lookup, such as rule application, evidence composition, contradiction resolution, or hypothesis comparison. This distinguishes it from perception, which primarily involves recognizing or extracting information from visual inputs (e.g., object detection or attribute identification), and from commonsense reasoning, which relies on implicit world knowledge and prior experience rather than explicit multi-step deduction [33]. It also clarifies the relation to compositional reasoning: compositional structure is often a useful ingredient, but it becomes complex logical reasoning only when composed parts must support a new inferential conclusion rather than a direct structural readout. In contrast, mathematical reasoning represents a more formalized and symbolic subset of logical reasoning, often involving precise calculations, algebraic manipulation, or rule-based derivations [15,26]. While mathematical reasoning is inherently logical, it typically operates in a more constrained and well-defined problem space compared to general multimodal reasoning [78,79].

Importantly, these paradigms are not mutually exclusive but instead exhibit hierarchical and compositional relationships. Complex logical reasoning often builds upon perceptual grounding and commonsense knowledge, while also incorporating mathematical reasoning in specialized tasks [11,13]. For example, solving a visual math problem may require perception (reading a diagram), commonsense (interpreting context), and symbolic reasoning (performing calculations) [80,81]. Therefore, rather than treating these categories as strictly separate, we view complex logical reasoning as an integrative capability that coordinates multiple cognitive components [12,77]. Fig. 4 further illustrates these conceptual boundaries by contrasting their input focus, core operation, evidence type, and typical failure mode.

As shown in Fig. 5, complex logical reasoning in LVLMs is closely related to several other reasoning paradigms, including commonsense, temporal, spatial, counterfactual, and strategic reasoning [24,25]. Although these paradigms emphasize different aspects of inference, they frequently interact in practical multimodal tasks. Complex logical reasoning provides a structured inference framework, while the other paradigms contribute complementary capabilities such as background knowledge, temporal ordering, spatial constraint modeling, hypothetical analysis, and decision-oriented planning. Understanding these relationships is important for developing comprehensive reasoning systems that can handle the full range of real-world applications.

Commonsense Reasoning. Commonsense reasoning involves applying everyday knowledge about the world to interpret situations and make reasonable inferences [33]. It frequently overlaps with logical reasoning, particularly abductive reasoning [9,23], when explaining observed phenomena. For example, if we see a wet sidewalk and dark clouds in the sky, commonsense reasoning allows us to infer that it likely rained.

Commonsense reasoning differs from formal logical reasoning in that it relies on implicit knowledge about the world rather than explicit logical rules [25]. This implicit knowledge is difficult to formalize and may vary across cultures and contexts. In LVLM contexts, commonsense reasoning is essential for interpreting ambiguous visual scenes and making reasonable assumptions about unobserved elements [82].

The relationship between commonsense reasoning and logical reasoning is complementary: logical reasoning provides formal frameworks for inference, while commonsense reasoning provides the background knowledge needed to apply those frameworks appropriately [24]. Integrating these reasoning types remains a key challenge for LVLM development.

Temporal Reasoning. Temporal reasoning focuses on understanding time-ordered sequences of events and reasoning about time-dependent relationships [83,84]. In multimodal contexts, temporal reasoning is essential for understanding video content, narrative descriptions, and cause-effect relationships that unfold over time.

Temporal reasoning in LVLMs requires representing time and reasoning about temporal relations such as before, after, simultaneous, and during [85]. This capability is essential for tasks such as video question answering, where questions may concern the order of events or the duration of activities. Beyond explicitly reasoning-oriented video benchmarks, large-scale video understanding resources such as YouTube-8M [86] have also played an important enabling role by providing millions of annotated videos for scalable representation learning.

The integration of temporal reasoning with logical reasoning is particularly important for causal reasoning tasks, which require understanding how events unfold over time and identifying causal relationships between temporally ordered events [10,34,35].

Spatial Reasoning. Spatial reasoning deals with understanding and reasoning about physical space and spatial relationships between objects [87]. In multimodal contexts, spatial reasoning is often combined with logical deduction to answer questions about object relationships, spatial configurations, and layout understanding.

Spatial reasoning in LVLMs requires representing spatial relationships such as above, below, inside, outside, near, and far [88]. Chen et al. [89] highlights that multimodal reasoning also requires explicit referential grounding: by allowing spatial coordinates to be expressed directly in natural-language interaction, it extends MLLMs toward location-aware dialogue and grounded reasoning. GeoQA is a representative benchmark in this direction, as it explicitly evaluates geometric question answering that requires joint visual perception, diagram understanding, and numerical reasoning [81].

The integration of spatial reasoning with logical reasoning is particularly important for tasks that require reasoning about spatial constraints, such as path planning, object arrangement, and spatial inference tasks that combine visual evidence with textual premises about spatial relationships [58].

Counterfactual Reasoning. Counterfactual reasoning involves reasoning about hypothetical alternatives to past events. It requires considering what would have happened if circumstances had been different, a capability that is essential for planning, decision-making, and understanding causality [90].

Counterfactual reasoning is closely related to causal reasoning, as both involve understanding causal relationships [10,34,35]. However, counterfactual reasoning specifically focuses on hypothetical scenarios that did not actually occur, whereas causal reasoning may also concern actual causal relationships.

In LVLM contexts, counterfactual reasoning is important for tasks such as visual explanation generation, where models must consider alternative explanations for observed phenomena [23], and for decision support systems, where models must evaluate the outcomes of different actions.

These reasoning paradigms interact with logical reasoning in complex ways, and successful LVLM systems must integrate capabilities across multiple paradigms [12,24]. Realistic environments such as WebArena [91] highlight the importance of long-horizon reasoning, interaction planning, and functional task completion in complex environments. The complexity of real-world reasoning tasks often requires combining multiple reasoning types, making the development of comprehensive reasoning systems a significant challenge.

Another related paradigm is strategic reasoning, which concerns planning and decision-making under dynamic, uncertain, and often multi-agent environments. A recent survey [20] shows that strategic reasoning emphasizes anticipating others’ behavior and adaptive policy selection, making it related to the logical-inference focus of many LVLM studies.

We identify three primary architectural paradigms for logical reasoning in LVLMs: unified architectures, modular architectures, and tool-augmented architectures. Although the boundaries between these paradigms are not always strict, they reflect three distinct design philosophies for integrating visual perception, language understanding, and reasoning. Unified architectures emphasize end-to-end multimodal representation learning; modular architectures explicitly separate perception and reasoning; and tool-augmented architectures extend LVLMs with external computation or symbolic tools. This categorization provides a useful lens for comparing how different systems support logical inference under multimodal inputs. Table 2 summarizes representative architecture works under these paradigms and highlights their reasoning-oriented design choices and limitations.

Unified Architectures. Unified architectures process visual and textual inputs within a tightly integrated multimodal model and learn joint representations across modalities. In practice, this family includes models that rely on cross-attention connectors, alignment bridges, or unified multimodal token spaces, even when the vision encoder and language backbone are not literally a single encoder. Representative examples include Flamingo [93], BLIP-2 [40], MiniGPT-4 [96], LLaVA [5,53], and more recent systems such as Qwen2-VL [99] and Qwen2.5-VL [73].

Flamingo [93] is an early representative of this paradigm. By inserting cross-attention layers between frozen visual and language backbones, it enables multimodal few-shot in-context learning while preserving the capabilities of large pretrained components. BLIP-2 [40] further improves this design through the Q-Former, which acts as a compact interface between visual features and the language model. LLaVA [5,53] shows that visual instruction tuning can substantially improve multimodal dialogue and compositional reasoning, while MiniGPT-4 [96] suggests that part of multimodal reasoning ability can be elicited from strong language backbones through lightweight alignment alone. More recent unified models also strengthen document, chart, multi-image, and video reasoning by improving visual tokenization, long-context handling, and interleaved multimodal modeling [72,73,99,100].

The main strength of unified architectures lies in end-to-end multimodal optimization. Because perception, alignment, and generation are trained in a closely coupled manner, these models are often effective at compositional understanding and at tasks that can be acquired through large-scale multimodal instruction tuning [2,16]. However, this tight coupling also makes the reasoning process difficult to interpret or verify. Intermediate reasoning is usually implicit in hidden states rather than represented as explicit symbolic structure, which can make these models less reliable on tasks requiring strict rule application, formal verification, or faithful multi-step inference [1,44].

Modular Architectures. Modular architectures separate visual encoding, language understanding, and logical reasoning into distinct components. This separation allows each module to be optimized for its own role and makes it easier to incorporate specialized reasoning mechanisms when explicit inference is required.

A prominent instance of this paradigm is neuro-symbolic modeling, which combines neural perception with symbolic reasoning engines [41,63]. In such systems, visual and linguistic inputs are first parsed into structured intermediate representations, such as objects, relations, or logical forms. A symbolic module then performs explicit inference over these structures and produces conclusions that can be translated back into natural language. This approach is especially attractive for deductive reasoning and compositional visual reasoning because the reasoning procedure is more transparent and, in principle, verifiable [49,51].

The main advantage of modular architectures is interpretability. Because perception and reasoning are partly decoupled, one can inspect the intermediate representations and reason about where errors occur. This is particularly valuable for logical tasks in which correctness depends on precise relational structure rather than fluent answer generation. The main weakness, however, lies in the interface between continuous perceptual representations and discrete symbolic structures. Parsing errors, information loss, or misalignment between neural outputs and symbolic inputs can propagate through the pipeline and degrade reasoning quality [62,71]. As a result, modular systems may be more rigorous on structured reasoning tasks, but less flexible when dealing with ambiguous, noisy, or open-ended multimodal inputs.

Tool-Augmented Architectures. Tool-augmented architectures extend LVLMs with external tools such as code interpreters, symbolic solvers, search modules, or task-specific vision components [105,106]. Instead of requiring the core model to internalize all reasoning skills, this paradigm treats the LVLM as a coordinator that decides when to call tools, how to formulate the calls, and how to integrate the returned results.

This paradigm is particularly effective when reasoning requires forms of computation that neural models do not perform reliably on their own, such as exact calculation, formal logical checking, or structured program execution [42,43]. Representative examples include Visual Programming [107], which combines LLM-based planning with external vision modules and executable intermediate steps, and Chameleon [71], which integrates multiple external tools into compositional reasoning workflows.

The primary strength of tool-augmented architectures is flexibility. By outsourcing specialized operations to external modules, these systems can extend reasoning capabilities without redesigning the base LVLM. A deeper limitation is that tool-augmented systems shift the burden of reasoning from internal model representations to external orchestration. This introduces a new class of failure modes, including incorrect tool selection, invalid program generation, and misinterpretation of tool outputs [43,108]. In complex multi-step scenarios, such errors can accumulate across reasoning stages, making the overall system brittle despite the correctness of individual tools.

Comparative Synthesis across Paradigms. The current empirical evidence remains uneven across unified, modular, and tool-augmented architectures. Broad shared benchmark suites such as MMBench,bib31, MathVista, MuirBench, and Video-MME are currently dominated by unified or unified-style LVLMs, which means they provide the clearest evidence for general benchmark transfer under common reporting conventions (see Section 5.3). In our standardized evidence map in Section 5.3, the systems with directly comparable results across these shared benchmarks are all unified or unified-style models, such as GPT-4V, Qwen2.5-VL, LLaVA-OneVision, and InternVL2.5 [37,73,100,101]. By contrast, modular and neuro-symbolic systems are more often evaluated on structured settings such as CLEVR, GeoQA, or targeted reasoning-faithfulness studies, where their explicit intermediate representations improve inspectability and rule-level verification but reduce direct comparability with open-ended LVLM benchmarks [39,49,81]. Tool-augmented systems report concrete gains when exact computation, symbolic execution, or external retrieval is essential, especially in program-execution or math-oriented settings, yet these gains are inseparable from the quality of orchestration and from the assumptions imposed by external tools [42,105,107]. Accordingly, current literature supports three cautious conclusions: unified architectures offer the strongest evidence of broad benchmark coverage, modular architectures offer the strongest evidence of explicit intermediate-state verification on structured tasks, and tool-augmented architectures offer the strongest evidence of exactness when external computation is reliable. None of these observations should be overgeneralized into a universal ranking, because the supporting evaluations still differ substantially in task distribution, annotation format, and end-to-end system boundary.

The quality of visual feature encoding directly affects logical reasoning capability. To support multimodal inference, visual representations must preserve information that is relevant not only for recognition, but also for abstraction, relation modeling, and cross-modal grounding [52,109]. This subsection discusses three aspects that are particularly important for reasoning-oriented LVLMs: visual tokenization, cross-modal alignment, and spatial encoding. Fig. 7 presents these encoding and alignment components as a reasoning-oriented representation pipeline.

Visual Tokenization. Visual tokenization determines the granularity and structure of the visual information that is made available to the language model. Different tokenization strategies support different forms of reasoning.

Patch-based tokenization, widely used in ViT-style architectures and models such as CLIP, divides an image into fixed-size patches that are treated as visual tokens [52,110]. This design is simple and scalable, and it works well for general-purpose representation learning. However, fixed patch grids do not naturally align with semantically meaningful entities, which can make object-centric and relation-centric reasoning more difficult.

Region-based tokenization instead derives features from detected objects or salient regions, making the visual representation more compatible with object-level reasoning [111,112]. This can be advantageous for tasks involving counting, comparison, and spatial or relational inference. Its limitation is that it depends on upstream detection quality and may omit contextual information that is not well captured by explicit object proposals.

A related alternative is to rely on segmentation-aware representations, where pixel-level or region-level semantic labels provide dense structural cues [113]. Such representations can support fine-grained reasoning about attributes and scene layout, but the predefined semantic categories may still not align well with the more abstract or relational categories required for logical reasoning.

Cross-Modal Alignment. Cross-modal alignment is the mechanism that connects visual and textual representations so that reasoning can proceed across modalities. In reasoning-oriented LVLMs, alignment must do more than capture coarse semantic similarity; it must support the binding of entities, attributes, relations, and constraints between modalities.

One common strategy is contrastive alignment, as used in CLIP-style pretraining, where paired image-text examples are pulled together in representation space and unpaired examples are pushed apart [52]. This produces strong global semantic representations, but it does not necessarily capture the fine-grained alignments needed for logical inference.

A second strategy is cross-attention-based alignment, exemplified by Flamingo [93], where the language model dynamically attends to visual features during inference. This design is more flexible and can adapt the alignment process to the current question or context, which is beneficial for reasoning tasks that require selective use of visual evidence.

A third strategy is fusion-based integration, in which visual and textual representations are combined through learned transformation modules or joint multimodal spaces [114]. Fusion can enable richer multimodal interaction, but if it is too coarse, it may blur modality-specific structure that remains important for reasoning.

Spatial Encoding. Spatial information is often central to logical reasoning, particularly for tasks involving object relations, geometric structure, physical layout, and scene configuration [88,115]. As a result, the way spatial structure is encoded has a direct impact on reasoning performance.

Absolute position encoding represents where tokens occur in the visual field. This can help preserve layout information, but absolute coordinates alone are often insufficient for reasoning because they do not directly capture relations between entities.

Relative position encoding instead models pairwise spatial relations such as left of, right of, above, and below. These relational encodings are generally more robust across changes in scale, viewpoint, and composition, and they are often more relevant for logical reasoning tasks.

Spatial structure encoding aims to represent higher-order layout or scene organization, enabling reasoning about configurations rather than isolated pairwise relations [74]. This is especially important for tasks involving multi-object spatial constraints, embodied interaction, or scene-level planning.

Chain-of-thought (CoT) reasoning encourages models to generate explicit intermediate reasoning steps before producing a final answer [1,116]. Early studies demonstrate that providing intermediate reasoning steps can significantly improve performance across arithmetic, commonsense, and symbolic reasoning tasks [117,118]. In LVLMs, this idea has become one of the most influential mechanisms for improving multi-step inference, because it makes the reasoning trajectory more explicit and can partially reduce direct answer guessing based on superficial multimodal correlations.

For LVLMs, visual or multimodal CoT extends standard CoT by explicitly integrating visual observations into the reasoning chain [116,119]. A common pattern is to first identify salient visual evidence, then translate it into textual or semi-structured intermediate descriptions, and finally use these grounded observations to support subsequent inference. In this sense, multimodal CoT typically involves two coupled components.

Visual Grounding in Reasoning. The first component is the extraction or verbalization of visual evidence that is relevant to the reasoning process. This may take the form of textual descriptions of objects, attributes, spatial relations, charts, or diagram elements that can then be incorporated into the reasoning chain [5,94]. The quality of this grounding step is critical: if the visual evidence is misidentified or inaccurately verbalized, subsequent reasoning may remain coherent in form while being unsupported by the input.

Multimodal Reasoning Chains. The second component is the explicit incorporation of grounded visual evidence into intermediate inference steps [36,116,120]. Rather than simply appending an image caption before answering, effective multimodal CoT requires the model to connect observations to relational, compositional, or causal inferences. For example, a valid chain may first identify that one object lies to the left of another and then use this grounded relation to answer a spatial query.

Representative methods such as Visual-CoT [119] and Multimodal-CoT/MMCoT [116,120] show that explicitly structuring reasoning around grounded visual evidence can improve performance on complex multimodal reasoning tasks. A common design choice in these approaches is to separate rationale generation from answer prediction, so that the model first constructs a reasoning trajectory and only then produces the final response. This decomposition can improve transparency and sometimes accuracy, although it does not by itself guarantee reasoning faithfulness.

CoT has also been extended in more structured directions. For instance, compositional chain-of-thought prompting [36] incorporates scene-graph-style relational structure into intermediate reasoning, thereby encouraging the model to reason over object relations more explicitly. More broadly, multimodal CoT methods increasingly attempt to move from free-form natural language rationales toward more structured intermediate representations.

Despite its practical effectiveness, CoT has clear limitations. First, generated reasoning chains may be plausible but unfaithful, reflecting post-hoc rationalization rather than the true basis of the model’s prediction [59,60]. Notably, prior empirical studies have shown that even invalid or partially incorrect reasoning chains can still yield high task performance, suggesting that CoT may not faithfully reflect the underlying reasoning process [1,117,118]. Table 3 summarizes representative reasoning mechanisms in LVLMs, highlighting their core features and the main limitations observed in current studies.

Program-of-thought (PoT) approaches reformulate reasoning problems as executable programs, typically using code interpreters or symbolic runtimes to carry out intermediate computations [43,105]. This idea is closely related to interpretable reasoning datasets such as MathQA [123], in which solving a problem involves generating an explicit sequence of operations rather than only a final answer. In multimodal settings, PoT is especially appealing because it offers a way to move part of the reasoning process from opaque neural inference into verifiable procedural execution. Fig. 9 contrasts implicit neural reasoning with program-of-thought and neuro-symbolic reasoning pipelines.

The key strength of PoT is that it externalizes reasoning into a structured form. Instead of relying solely on free-form natural language rationales, the model generates code or executable procedures whose intermediate states can be inspected and verified. This is particularly valuable for tasks involving arithmetic computation, symbolic manipulation, geometric reasoning, or other forms of structured inference in which correctness can be checked mechanically [43,105].

From a logical-reasoning perspective, PoT offers three major benefits. First, it makes reasoning explicit: each computational step is represented in a formal procedure rather than being hidden in model activations. Second, it adds computational precision: once the problem has been translated into an executable program, the system can rely on the external runtime for exact calculation or rule application. Third, it improves debuggability: errors can often be localized to specific steps in the generated program instead of being diffused across a natural-language rationale.

Recent work on code-enhanced reasoning [38,108] further suggests that executable intermediate representations can improve reasoning beyond purely mathematical settings. However, the success of PoT depends critically on whether the model can correctly formalize the underlying problem. In multimodal contexts, this often requires translating visual inputs into structured symbolic variables, relations, or functions before execution. If this translation fails, the advantages of downstream executability are greatly reduced.

Neuro-symbolic approaches address a related problem from a complementary direction. Rather than merely converting reasoning into code, they explicitly combine neural perception modules with symbolic inference engines [41,63]. In these systems, neural components handle perception and language understanding, while symbolic components operate over structured intermediate representations to perform formal inference. This architecture is particularly well suited to tasks such as deductive reasoning, compositional visual reasoning, and logical entailment, where the reasoning process benefits from discrete structure and explicit rule application [62].

A representative example is Visual Programming (VisProg) [107], which decomposes complex visual tasks into modular executable procedures. More broadly, recent reviews of neuro-symbolic AI [51] and studies of robust reasoning in VLMs [49] show that neuro-symbolic methods remain one of the most promising routes toward stronger reasoning faithfulness and verifiability.

At the same time, both PoT and neuro-symbolic methods face a common bottleneck: the interface between neural representations and symbolic structure. To benefit from formal reasoning, the model must first produce accurate structured representations from ambiguous and noisy multimodal inputs. Errors in parsing, grounding, or formalization can easily propagate into the symbolic stage. For this reason, these approaches are often strongest on structured reasoning tasks, but less robust in open-ended scenarios where perceptual ambiguity and linguistic underspecification are substantial.

This point is crucial when comparing these methods with CoT-style reasoning. Program execution and symbolic inference can improve step verifiability once a correct formal representation has been obtained, but they do not automatically guarantee reasoning faithfulness in the broader multimodal sense. In practice, PoT often verifies the correctness of downstream computation rather than the correctness of upstream visual formalization, while neuro-symbolic pipelines often verify rule application rather than the reliability of perceptual grounding. As a result, these methods reduce some forms of structured hallucination and arithmetic error, yet they do not fully eliminate the gap between answer correctness and genuinely grounded reasoning in open-ended LVLM settings [49,59,60]. Their main contribution is therefore best understood as relocating the faithfulness bottleneck from free-form generation to representation quality, grounding fidelity, and interface robustness.

Self-correction mechanisms aim to improve reasoning reliability by enabling models to detect, revise, or filter erroneous reasoning trajectories [44,57]. These mechanisms are particularly important for LVLMs because multimodal logical reasoning is vulnerable to two intertwined sources of error: perceptual mistakes in interpreting the input and inferential mistakes in chaining the intermediate steps. Without some form of internal correction or verification, such errors can propagate directly to the final answer.

One widely used strategy is self-consistency [44]. Instead of relying on a single reasoning chain, the model samples multiple reasoning paths and selects the answer that is most consistent across them. The intuition is that correct reasoning is more likely to converge to the same conclusion across diverse trajectories, whereas incorrect reasoning may produce unstable or conflicting outputs. In multimodal reasoning, this can reduce sensitivity to local reasoning errors or brittle prompt effects, although its effectiveness depends on the diversity and quality of the sampled chains.

A second strategy is reflection or reflexive self-evaluation, in which the model explicitly reviews its own reasoning before committing to a final answer [38]. Here, the model may be prompted to critique intermediate steps, reconsider missing evidence, or justify why the conclusion follows from the input. Reflection can be useful for identifying obvious inconsistencies or unsupported leaps, but it does not guarantee genuine error detection: a model may also rationalize an incorrect answer more confidently rather than correct it.

A third strategy is iterative refinement, where reasoning is updated across multiple rounds rather than produced in a single forward pass [57,121]. In these approaches, intermediate reasoning states are progressively revised, expanded, or reorganized. This is especially helpful for tasks requiring long reasoning chains or hierarchical decomposition, since early-stage drafts can be corrected before the final output is generated.

Interpretability is crucial for understanding how LVLMs produce reasoning outputs and for diagnosing why they fail [59,60]. In the context of logical reasoning, interpretability is not merely a desirable auxiliary property; it is closely tied to the broader question of reasoning faithfulness. If a model produces a correct answer, we still need to understand whether the answer was derived through grounded inference, shortcut exploitation, or post-hoc explanation.

Several interpretability approaches are particularly relevant to multimodal logical reasoning.

Attention Visualization. One common approach is to inspect attention patterns to determine which visual elements and textual tokens influence the model during reasoning [88]. In multimodal settings, attention analysis can help reveal whether the model is attending to the relevant objects, regions, or relations when producing a reasoning trace or final answer. Although attention alone should not be overinterpreted as a full explanation of model behavior, it remains a useful diagnostic signal for detecting missing grounding, distractor sensitivity, or cross-modal misalignment.

Reasoning Step Tracking. Another important approach is to inspect intermediate reasoning states or generated step-by-step traces [120]. Tracking reasoning steps allows researchers to identify where a reasoning chain breaks down, whether the error originates from perception or inference, and how mistakes propagate across multiple steps. This is particularly valuable for tasks involving long-horizon reasoning, where final-answer evaluation alone provides little information about failure modes.

Explanation Analysis. A third approach is to examine generated natural-language explanations or rationales [1]. Such explanations can provide insight into the model’s apparent reasoning strategy and may reveal whether the conclusion is supported by grounded evidence or instead reflects superficial pattern matching. However, explanation generation also has a major limitation: plausible explanations can be post-hoc and unfaithful. As a result, explanations should be treated as evidence for analysis, not as definitive proof of the model’s internal reasoning process.

We categorize benchmarks according to the primary form of reasoning they evaluate. This organization is not absolute, since many benchmarks involve multiple reasoning skills simultaneously. Nonetheless, it provides a useful overview of the current evaluation landscape and helps clarify which aspects of logical reasoning are well covered and which remain underexplored. Table 4 provides a comparison of representative logical reasoning benchmarks, including their release year, primary reasoning type, modality, and core evaluation focus.

At a broad level, current benchmarks fall into several recurring groups: inductive or analogy-style benchmarks, compositional and multi-hop benchmarks, deductive or entailment-oriented benchmarks, and causal or counterfactual benchmarks. In addition, a number of broader multimodal evaluation suites include logical reasoning as one component, though often not as their exclusive focus.

Inductive reasoning benchmarks. RAVEN [22] is one of the most representative benchmarks for inductive reasoning in visual contexts. It evaluates rule induction through Raven-style matrix completion tasks, in which models must infer latent transformation rules from structured visual patterns. The benchmark includes multiple matrix configurations, such as relational and progressive structures, and therefore provides a controlled testbed for evaluating whether a model can abstract patterns rather than merely recognize objects. Its main strengths lie in its clear focus on rule induction and its systematic control of difficulty. Its main limitation, however, is that it is primarily visual and synthetic, which constrains its representativeness for broader multimodal logical reasoning.

Compositional and multi-hop reasoning benchmarks. CLEVR [39] remains a foundational benchmark for compositional and multi-step visual reasoning. It presents synthetic scenes containing multiple objects and asks questions that require combining attributes, relations, and counting operations across several reasoning steps. CLEVR is valuable because it was explicitly designed to reduce dataset bias and isolate reasoning structure. However, its synthetic nature also limits ecological validity: the benchmark evaluates controlled compositional reasoning well, but does not capture the ambiguity and noise of real-world multimodal inputs.

Beyond synthetic scenes, more recent multi-hop settings attempt to evaluate more realistic evidence integration. Such benchmarks are motivated by the observation that many real reasoning tasks require models to connect information scattered across different regions, images, or modalities rather than answer from a single local cue [54–56]. This broader class of benchmarks is particularly important for evaluating logical reasoning in document, chart, and multi-image settings.

Deductive reasoning benchmarks. ReClor [64] is a representative benchmark for deductive reasoning and logical entailment. It presents reading comprehension problems that require identifying whether a conclusion follows from a set of premises through valid logical inference. Its main value lies in its emphasis on formal reasoning structure rather than open-ended commonsense interpretation. However, because it is fundamentally text-based, its relevance to multimodal LVLM evaluation is indirect: it is highly useful for conceptual grounding, but does not by itself test whether a model can integrate visual evidence into deductive inference.

Causal reasoning benchmarks. Causal reasoning has only recently begun to receive dedicated multimodal benchmark support. MuCR (Multimodal Causal Reasoning Benchmark) [10] is a representative example. It evaluates whether models can identify causal relations across visual and textual evidence, distinguish correlation from causation, and reason about interventions or counterfactual changes. Compared with earlier general multimodal benchmarks, MuCR is notable because it explicitly targets a class of reasoning that is central to real-world decision making but difficult to capture with answer-only evaluation.

Related efforts extend causal evaluation to temporal settings. Video-based causal benchmarks assess whether models can infer causal relations from event sequences rather than static images alone [34]. Such tasks are especially important because causal reasoning often depends on temporal structure, yet they are also more difficult to evaluate reliably due to longer contexts and greater annotation complexity.

Broader observations on current benchmark coverage. Taken together, current benchmarks cover important but still fragmented slices of the reasoning problem. RAVEN captures abstract pattern induction, CLEVR captures synthetic compositional reasoning, ReClor captures text-based deductive reasoning, and MuCR and related datasets begin to probe multimodal causal inference. However, few benchmarks simultaneously combine real-world multimodal complexity, diverse reasoning types, and process-aware evaluation. This fragmentation remains one of the central challenges in assessing logical reasoning in LVLMs [25,29].

It is also worth noting that large-scale resources not specifically designed for logical reasoning have still played an enabling role in benchmark development and model training. For example, YouTube-8M [86] provided large-scale video infrastructure that later supported progress in video-language understanding and temporal reasoning. Although such resources do not directly evaluate logical inference, they contributed to the broader ecosystem in which reasoning-oriented multimodal systems emerged.

Evaluating complex logical reasoning in LVLMs remains fundamentally challenging because correct answers do not necessarily imply valid reasoning processes [125,126]. Existing evaluation protocols can be broadly divided into answer-level metrics and process-level metrics, each capturing a different aspect of model behavior. Fig. 11 summarizes this distinction between answer-level and process-level evaluation.

Answer-level metrics. Answer-level metrics evaluate whether the final prediction matches a reference answer, without explicitly considering how that answer was obtained. Common examples include accuracy, F1 score, and exact match. These metrics remain widely used because they are simple, scalable, and easy to compare across models and benchmarks.

However, their limitations are especially pronounced in logical reasoning tasks. First, answer-level scores cannot distinguish between genuine inference and shortcut-based prediction. A model may produce the correct answer by exploiting dataset regularities or shallow multimodal cues rather than by carrying out the intended reasoning process. Second, answer-only evaluation cannot represent the existence of multiple valid reasoning paths, which is common in open-ended or multi-step tasks. Third, strict surface-form matching can be brittle: two answers may be semantically equivalent while differing in wording or decomposition.

Process-level metrics. To address these weaknesses, process-level metrics attempt to evaluate the quality, validity, and structure of intermediate reasoning steps [124,126].

These dimensions can be organized as a structured evaluation protocol: 1.

Step correctness: verifying whether each intermediate inference step is logically valid;

Despite their conceptual appeal, process-level metrics are difficult to deploy at scale. They often require detailed annotations of intermediate steps or access to internal model traces, both of which are expensive and difficult to standardize across tasks [125,126]. Moreover, ground-truth reasoning trajectories are often not unique: different but equally valid solution paths may exist for the same problem [124]. This makes process supervision and evaluation substantially harder than answer-only scoring.

More fundamentally, even current process-level metrics do not fully capture several key aspects of multimodal logical reasoning, including cross-modal grounding validity, abstraction consistency, and causal coherence [34,35]. For this reason, there remains a strong need for evaluation frameworks that jointly assess correctness, faithfulness, and generalization under realistic multimodal conditions [29,46].

To make the discussion of empirical findings more explicit, we further summarize the main patterns observed across existing evaluations. Current LVLMs generally show stronger performance on perception-oriented and short-answer multimodal tasks, but their performance becomes less stable when tasks require multi-step inference, cross-modal evidence integration, counterfactual reasoning, or faithful explanation generation. Therefore, we emphasize not only reported performance scores but also the evaluation conditions, reasoning requirements, and failure patterns behind these results.

Although existing studies report substantial progress in LVLM reasoning, direct cross-method comparison remains difficult because different works adopt different benchmarks, answer formats, decoding settings, and modality assumptions [2,3,50]. To make the comparison more systematic, we further conduct a standardized quantitative synthesis over representative LVLMs and reasoning-oriented benchmarks. Rather than treating heterogeneous benchmark results as fully interchangeable, we organize the comparison along three reasoning-relevant axes: (1) single-image multimodal reasoning, measured by MathVista [78], MMBench [45], and MMMU [31]; (2) multi-image reasoning, measured by MuirBench [132]; and (3) video temporal reasoning, measured by Video-MME [131]. These benchmarks are selected because they are widely used in recent LVLM evaluation and provide explicit quantitative results for representative models [4,30].

We adopt the following standardization criteria. First, we only include numerical results that are explicitly reported in the original papers or official benchmark pages [7,37,95]. Missing results are marked as “–” and are not imputed. Second, all reported values are accuracy scores in percentage form unless otherwise noted. Third, for the single-image reasoning axis, we compute a simple descriptive average over MathVista, MMBench, and MMMU only when all three values are available [19,26].(1)R-AvgSI(m)=13(AMathVista(m)+AMMBench(m)+AMMMU(m)),where Ab(m) denotes the accuracy of model m on benchmark b. This quantity is intended only as a coarse descriptive index, not as a statistically rigorous meta-analytic estimate, because MathVista, MMBench, and MMMU differ in domain mix, answer format, and evaluation protocol. Fourth, for Video-MME, we report the “without subtitles” setting when both subtitle-free and subtitle-enhanced results are available, because it better isolates visual-temporal reasoning from additional textual transcript information. Finally, we do not collapse single-image, multi-image, and video scores into a single overall score, since these axes measure different reasoning conditions and should not be interpreted as strictly equivalent. Open-weight and proprietary systems are also listed separately in Table 5, because their training data, system prompts, and evaluation interfaces are not fully standardized.

Table 5 and Fig. 12 summarize the resulting comparison. The results show several recurring trends. First, model scaling improves reasoning performance, as seen from the LLaVA-OneVision series: R-AvgSI increases from 39.4 for the 0.5B model to 70.1 for the 72B model, while MuirBench [132] and Video-MME [131] also improve consistently. Second, strong single-image reasoning does not automatically imply equally strong multi-image or video reasoning. For example, GPT-4V obtains a competitive single-image descriptive average [37,47], but LLaVA-OneVision-72B [100] reports stronger Video-MME performance in the subtitle-free setting. Third, the multi-image setting remains particularly challenging: even strong models show a noticeable drop on MuirBench compared with standard single-image benchmarks such as MMBench [45] and MMMU [31]. These observations should be interpreted within access-matched groups or as broad evidence trends rather than as a fully controlled ranking across all open and proprietary systems.

The quantitative comparison also clarifies the limitations of current evidence. Some reasoning mechanisms reviewed in Section 4, such as program-of-thought [43], neuro-symbolic reasoning [49,51], and self-correction [44], often report results on task-specific benchmarks rather than on a shared multimodal reasoning suite. For instance, MathVista [78] reports tool-augmented textual baselines such as PoT GPT-4 with caption and OCR inputs, but these results are not directly equivalent to end-to-end LVLM inference because the visual information has already been converted into text [26]. Therefore, instead of claiming a universal ranking across all reasoning mechanisms, we treat Table 5 as a conservative evidence map: it compares representative systems only where public and directly comparable benchmark scores are available. This analysis reinforces the need for future standardized evaluation protocols that jointly report final accuracy, reasoning-process validity, modality-specific grounding, and robustness under distribution shift [13,19,25].

Despite the growing number of multimodal benchmarks, existing evaluation protocols remain insufficient for rigorously assessing complex logical reasoning in LVLMs [25,29]. Fig. 13 summarizes the major limitations of current LVLM evaluation protocols. Existing benchmarks mainly emphasize final-answer accuracy, but this metric alone cannot reveal whether the model follows a valid reasoning process, grounds its answer in visual evidence, or remains robust under distribution shift and adversarial perturbations.

Benchmark contamination and data leakage. A fundamental concern is that benchmark instances may appear, directly or indirectly, in large pretraining corpora [37]. Because many popular benchmarks are publicly available, large-scale models may partially memorize patterns or examples from them, which can inflate apparent reasoning performance. This is especially problematic for logical reasoning, where benchmark success may otherwise be misinterpreted as evidence of genuine inferential capability [135].

Limited coverage of reasoning types. Most existing benchmarks probe only a narrow subset of reasoning skills, such as visual pattern induction [22] or synthetic compositional reasoning [39]. Yet real multimodal reasoning often requires the joint use of multiple forms of inference, including deductive, abductive, causal, and compositional reasoning [9,23,75]. The fragmented nature of current benchmark design therefore makes comprehensive evaluation difficult.

Outcome-oriented evaluation bias. A dominant limitation of current practice is the heavy reliance on final-answer correctness [44]. This biases evaluation toward outcomes rather than reasoning validity. As a result, models that exploit superficial multimodal regularities may appear competitive with models that genuinely perform structured inference. Existing benchmarks are therefore often insufficient to separate true reasoning from shortcut-based inference [1].

Synthetic-to-real gap. A substantial portion of existing reasoning benchmarks is based on synthetic data [22,39]. Synthetic settings are useful because they provide control and reduce annotation ambiguity, but they lack the noise, ambiguity, and long-tail variability of real-world multimodal inputs. Models that perform well in controlled synthetic environments may therefore fail to transfer their reasoning ability to realistic scenarios [31].

Scale and statistical reliability. Many logical reasoning benchmarks are relatively small, limiting their power to support statistically stable model comparison [30,45]. Small test sets are more vulnerable to variance, narrow coverage, and benchmark-specific artifacts, making it difficult to draw reliable conclusions about general reasoning ability.

Distributional robustness and generalization. Performance on standard benchmarks often degrades substantially under distribution shift, including changes in visual style, wording, task format, or evidence layout [31,46]. This raises a broader concern: even when a model performs well on a benchmark, it may still lack the robustness required for practical reasoning deployment.

Annotation quality and ambiguity. Evaluation is further complicated by annotation noise, underspecification, and the fact that many logical reasoning tasks admit multiple plausible solution paths [126]. In such cases, defining a single gold answer or a single gold reasoning trace may be inherently reductive, which complicates fair comparison across systems.

A central challenge is ensuring that model outputs reflect genuine logical inference rather than superficial pattern matching [1,44]. Even when LVLMs produce correct answers, those answers may be driven by spurious correlations, memorized shortcuts, or dataset regularities rather than valid reasoning [135].

This faithfulness problem has several sources. First, training data may contain exploitable biases, allowing models to succeed without learning the intended reasoning patterns [33]. Second, current benchmarks often emphasize final-answer accuracy and therefore do not clearly distinguish between genuine inference and shallow answer prediction [29]. Third, many reasoning tasks involve implicit premises, background assumptions, or underspecified constraints, making it difficult to determine whether a model is reasoning correctly or simply generating plausible outputs [62].

Recent work on neuro-symbolic reasoning and reasoning robustness has further emphasized that answer correctness alone is insufficient evidence of valid reasoning [49]. Addressing this challenge requires training objectives and evaluation protocols that target faithfulness directly, rather than treating it as a by-product of answer accuracy.

Hallucination remains a major challenge in LVLM reasoning, but in reasoning settings it is especially problematic because errors can compound across inference steps [59,60]. A hallucinated intermediate conclusion may appear coherent and even support later steps, ultimately leading to a convincing but incorrect final answer.

This problem is more difficult than ordinary generation errors because reasoning hallucinations are often structurally plausible. Models may invent unsupported relations, attributes, or causal links, and such errors can remain hidden inside a multi-step chain [90]. The risk increases when reasoning chains become longer, when external grounding is weak, or when the model must infer latent structure from noisy visual evidence [44,57].

Reducing reasoning hallucination therefore requires more than improving fluent generation. It requires stronger grounding, intermediate verification, and mechanisms for checking whether each step is supported by the input or by trusted external knowledge [82].

Complex logical reasoning often requires maintaining coherence across long inference chains, yet long-range reasoning remains a persistent weakness of current models [57,121]. As the number of reasoning steps increases, models must preserve relevant intermediate states, retrieve earlier premises, and prevent local errors from cascading through later steps.

Current LVLMs still struggle with these requirements. Long reasoning chains place pressure on memory, attention allocation, and step-by-step consistency [32]. Earlier mistakes can easily propagate, and models may lose track of previously established constraints or fail to integrate them correctly into later stages [44,116]. These issues are particularly severe in multi-image, long-document, and video settings, where evidence is distributed across extended multimodal contexts.

Improving long-range reasoning will likely require better memory mechanisms, more reliable intermediate-state management, and architectures or training strategies that explicitly support long-horizon inference.

Robust logical reasoning in LVLMs depends on accurate alignment between visual and textual representations [5,52]. This remains difficult because the two modalities differ substantially in form: visual information is continuous, spatial, and often ambiguous, whereas textual information is discrete, sequential, and usually more explicit [7].

For reasoning tasks, alignment must go beyond coarse semantic matching. Models must determine which visual evidence corresponds to which textual premise, maintain consistency across modalities, and resolve conflicts when the two sources provide incomplete or partially inconsistent information [45,111]. Misalignment at this stage can undermine the entire reasoning process, even if the downstream inference mechanism is otherwise strong [95].

This challenge suggests that future work should not treat representation alignment as merely a perception problem. Instead, alignment mechanisms must be designed with reasoning in mind, especially for tasks involving compositional structure, abstraction, and multimodal constraint satisfaction.

The development of logical reasoning in LVLMs is constrained not only by limited data volume, but also by the quality of the reasoning data itself [12,26]. In contrast to general multimodal instruction data, reasoning-oriented corpora must often include carefully designed problem structures, valid intermediate reasoning steps, reliable annotations of latent relations, and enough diversity to prevent shortcut learning. These requirements make high-quality data expensive to construct and difficult to scale.

This challenge has several dimensions. First, annotating reasoning trajectories requires expertise and is often labor-intensive [64]. Second, synthetic data can help scale reasoning supervision, but may not transfer well to realistic multimodal settings [22,39]. Third, dataset quality is highly sensitive to hidden sampling bias, annotation regularity, and source imbalance. Seemingly broad datasets may still overrepresent particular domains, visual layouts, answer formats, or reasoning templates, which can allow models to exploit superficial regularities instead of learning transferable inference patterns [31,39,45]. In multimodal reasoning, such bias may appear as shortcut dependence on OCR artifacts, visual co-occurrence priors, or benchmark-specific answer distributions.

Data quality problems also interact with evaluation quality. If benchmark construction reuses narrow templates, weak negative cases, or partially contaminated sources, answer accuracy can overestimate genuine reasoning ability [29,48]. For this reason, better reasoning data should be judged not only by scale, but also by diversity, annotation reliability, source balance, and resistance to shortcut exploitation.

As a result, progress in reasoning often depends on relatively small or specialized datasets, which is unlikely to be sufficient for building broad, robust, and transferable multimodal reasoning capabilities. Better data construction pipelines, scalable synthetic-real mixtures, stronger annotation audits, and higher-quality process supervision remain important open problems.

Finally, current benchmarks and evaluation protocols still provide only a partial view of reasoning ability [29,134]. Benchmark performance may not reliably reflect whether a model can reason in a grounded, generalizable, and faithful manner [31,45].

Several issues contribute to this gap. Existing benchmarks often cover only a limited subset of reasoning types [35]; many are susceptible to contamination or data leakage [48]; and most still emphasize answer-level outcomes over reasoning processes [120]. As discussed in the previous section, this makes it difficult to distinguish true reasoning from benchmark-specific shortcuts.

For this reason, evaluation remains one of the most important unresolved problems in the field. Progress will depend on benchmarks that cover a broader range of reasoning skills, are more robust to contamination and distribution shift, and include process-aware protocols capable of assessing not only correctness, but also faithfulness and reasoning quality.

Taken together, these bottlenecks show that the central obstacle is no longer raw benchmark coverage alone, but the difficulty of maintaining grounded, verifiable, and transferable reasoning under realistic multimodal conditions. Table 6 summarizes the main challenge-to-evidence mapping discussed in this section and highlights the most actionable open directions [24,25].