LogiCAM: Modular Multimodal Logical Reasoning

• LogiCAM is a modular framework that applies formal logical rules to multimodal inputs, integrating image and text evidence in vision-language settings.
• It decomposes reasoning into modules such as premise selection, logic type identification, and iterative inference, combining symbolic and commonsense methods.
• The framework significantly improves MuSLR benchmark accuracy, raising GPT-4.1 performance from 46.84% to 60.97% by structuring logical reasoning steps.

LogiCAM is a modular framework for multimodal symbolic logical reasoning introduced in the MuSLR study as a method for applying formal logical rules to multimodal inputs, specifically image evidence and text, within a vision-language setting. It is presented as a Chain-of-Thought-based system built on GPT-4.1 and designed for the MuSLR task, where a model must not merely describe multimodal content but carry out explicit logic-grounded inference. The framework is motivated by the paper’s empirical finding that contemporary vision-LLMs perform poorly on formal multimodal reasoning: on MuSLR-Bench, the best plain baseline, GPT-4.1, reaches 46.84% accuracy, and LogiCAM raises that figure to 60.97% (Xu et al., 30 Sep 2025).

1. Definition, scope, and naming

LogiCAM is defined as a “modular framework that applies formal logical rules to multimodal inputs” and as the paper’s proposed baseline for MuSLR, a benchmark for multimodal symbolic logical reasoning. The name is expanded in the paper as “Logical reasoning with Commonsense Augmentation with Multimodality”; the introduction also phrases it as “in Multimodalities.” In methodological terms, LogiCAM is not a classical theorem prover and not a newly trained model architecture. Rather, it is a modular neuro-symbolic prompting pipeline that uses a vision-LLM to approximate symbolic reasoning directly over image-text inputs (Xu et al., 30 Sep 2025).

The framework is positioned against two limitations of prior work. First, plain vision-LLMs are shown to be weak at formal multimodal logic, especially when reasoning chains are long or when correct inference requires precise alignment between textual symbols and visual referents. Second, text-only LLM-plus-solver pipelines are described as lossy in this setting because they require translating images into text before reasoning. LogiCAM keeps inference inside a vision-LLM that can inspect the image directly, while structuring the reasoning process into explicit modules (Xu et al., 30 Sep 2025).

Within the broader literature, LogiCAM occupies a distinct niche. LogicCLIP addresses logical sensitivity in CLIP-style joint embedding models through logic-aware training objectives and a benchmark spanning images, videos, anomaly detection, and medical diagnostics (Zhou et al., 15 Aug 2025). LogicAD targets industrial logical anomaly detection by combining autoregressive vision-LLMs, structured text representations, and theorem proving with Prover9 (Jin et al., 3 Jan 2025). CodeLogician, in turn, addresses software logic rather than multimodal reasoning, pairing LLMs with ImandraX for explicit formal modeling and exact reasoning over programs (Lin et al., 17 Jan 2026). Against these systems, LogiCAM is specifically an inference-time modular framework for multimodal symbolic reasoning in a VLM.

2. Task formulation and benchmark environment

LogiCAM is defined within the MuSLR task family. The paper specifies two task forms. In Truth Evaluation, the input is $(I, T, A)$, where $I$ is an image, $T$ is text, and $A$ is an argument; the required output is

$$\mathrm{Truth}(A) \in \{\mathrm{True}, \mathrm{False}, \mathrm{Unknown}\},$$

together with reasoning steps $R=\{R_1,\dots,R_n\}$. In Multiple Choice, the input is $(I, T, \{A_1,A_2,A_3,A_4\})$, and the model must output $\mathrm{BestArgument}(I,T)$, again with reasoning steps (Xu et al., 30 Sep 2025).

MuSLR is reported as comprising 1,093 instances, 35 atomic symbolic rules, 976 symbolic rule combinations, reasoning depths ranging from 2 to 9, and an average context length of 554.9. The paper describes the benchmark as spanning 7 domains, but its main results table uses eight domain columns: Healthcare, Traffic, Sports, Entertainment, Social, Science, Finance, and General. The coexistence of “7 domains” in the prose and eight domain columns in the table is an internal inconsistency in the paper rather than a settled property of the benchmark (Xu et al., 30 Sep 2025).

The logical families represented in MuSLR are propositional logic (PL), first-order logic (FOL), and non-monotonic logic (NM). Evaluation uses direct answer match for final-task correctness and ROUGE-L, BERTScore-F1, and ROSCOE for reasoning traces. The benchmark setting reported in the paper uses three-shot CoT prompting for all baselines with temperature 0.0 (Xu et al., 30 Sep 2025).

3. Modular architecture and iterative workflow

LogiCAM is organized into three reasoning modules plus an iterative completion loop: Premise Selector, Reasoning Type Identifier, Reasoner, and Check for Completion / iterate. This decomposition is the framework’s defining architectural feature (Xu et al., 30 Sep 2025).

The Premise Selector receives the image $I$ and text $T$, where $I$0 contains context $I$1 and question $I$2. It first selects the most relevant symbolic rules $I$3, then identifies the visual evidence $I$4 relevant to those rules. The paper states that the symbolic rule $I$5 and visual details $I$6 are combined into $I$7. In the appendix, the same stage is formalized as selecting

$I$8

with the VLM instructed to prioritize a pair $I$9 for which a formal inference rule can apply (Xu et al., 30 Sep 2025).

The Reasoning Type Identifier decides whether the current inference step should use formal symbolic reasoning or commonsense reasoning. Its criterion is explicit: symbolic reasoning is chosen if the selected premises contain a pair such that a formal rule can apply, exemplified by

$T$0

Otherwise, the framework invokes heuristic or commonsense reasoning to bridge missing information (Xu et al., 30 Sep 2025).

The Reasoner then derives new knowledge. If symbolic reasoning is selected, it applies a formal inference pattern to the chosen premises. If commonsense reasoning is selected, it produces a plausible implication $T$1 grounded in real-world knowledge. The paper is explicit that this module uses a VLM to approximate formal logical rules rather than calling an external theorem prover (Xu et al., 30 Sep 2025).

After each step, LogiCAM checks whether the current conclusion is sufficient to answer the question. If it is, the process terminates. Otherwise, the inferred conclusion is appended to context,

$T$2

and the reasoning cycle repeats. This gives LogiCAM its iterative structure: premise extraction, reasoning-mode choice, inference, state update, and repetition (Xu et al., 30 Sep 2025).

4. Symbolic substrate and inference mechanics

The symbolic environment underlying LogiCAM is inherited from MuSLR’s benchmark construction. The paper defines a complete rule set

$T$3

a selected subset

$T$4

meaningful rule combinations

$T$5

and reasoning chains

$T$6

Grounding then produces visual features $T$7, retrieved text $T$8, a grounded rule set $T$9, and a hybrid reasoning chain

$A$0

This formalization shows that LogiCAM reasons over grounded symbolic and commonsense chains rather than over free-form captions alone (Xu et al., 30 Sep 2025).

MuSLR includes 35 atomic symbolic rules spanning propositional logic, first-order logic, and non-monotonic reasoning. The benchmark explicitly instantiates standard rules such as Modus Ponens

$A$1

Modus Tollens

$A$2

Hypothetical Syllogism

$A$3

Disjunctive Syllogism

$A$4

Constructive Dilemma, Destructive Dilemma, Biconditional Dilemma, De Morgan’s Transformation, Existential Generalization, and Universal Instantiation. It also includes extended multi-variable FOL rules such as

$A$5

as well as non-monotonic patterns including DRS, DRI, DRD, DRO, REI, REII, REIII, and RAP (Xu et al., 30 Sep 2025).

At the answer stage, the paper formalizes completion criteria precisely. For truth evaluation with hypothesis $A$6, if

$A$7

LogiCAM outputs True; if

$A$8

it outputs False; otherwise it continues. For multiple choice with hypotheses $A$9, it selects $$\mathrm{Truth}(A) \in \{\mathrm{True}, \mathrm{False}, \mathrm{Unknown}\},$$0 if exactly one satisfies

$$\mathrm{Truth}(A) \in \{\mathrm{True}, \mathrm{False}, \mathrm{Unknown}\},$$1

The iterative state update is

$$\mathrm{Truth}(A) \in \{\mathrm{True}, \mathrm{False}, \mathrm{Unknown}\},$$2

If the iteration bound is reached without a conclusive result, the system outputs Unknown for truth evaluation, or the multiple-choice result is counted as incorrect (Xu et al., 30 Sep 2025).

5. Empirical performance and benchmark behavior

The MuSLR evaluation reports that all tested state-of-the-art vision-LLMs struggle on the benchmark. The overall accuracies are GPT-4.1: 46.84%, InternVL: 45.20%, Qwen: 41.63%, GPT-4o: 38.93%, InstructBLIP: 35.59%, Llava: 35.13%, and Claude: 33.49%. LogiCAM, built on GPT-4.1, reaches 60.97%, yielding an absolute gain of 14.13 points over the GPT-4.1 Chain-of-Thought baseline (Xu et al., 30 Sep 2025).

The paper reports that gains are largest on more formal logic families. Relative improvements are 48.93% on FOL, 31.93% on PL, and 26.17% on NM. The baseline averages by logic type are 37.04% for FOL, 42.77% for PL, and 46.09% for NM, reinforcing the claim that quantified and relational reasoning is the most difficult regime and the one in which LogiCAM contributes most (Xu et al., 30 Sep 2025).

Depth analysis shows that accuracy declines as reasoning chains lengthen, but LogiCAM degrades less sharply than plain CoT baselines. GPT-4.1 shows a 16% drop from depths 2–3 to 8–9, while Claude declines by 20%. LogiCAM records 71.91% at depth 2–3 and 54.61% at depth 8–9, and the paper states that it exceeds GPT-4.1 by 13% on the hardest depth 8–9 subset (Xu et al., 30 Sep 2025).

Reasoning-trace evaluation shows LogiCAM with ROUGE-L = 0.170, BERTScore = 0.835, and overall mean 0.590. GPT-4.1 records ROUGE-L = 0.166, BERTScore = 0.833, and ROSCOE = 0.725. Claude attains the highest ROSCOE at 0.784 but a much lower ROUGE-L of 0.084. The paper interprets this as evidence that surface-level or semantic similarity does not reliably track rigorous logic (Xu et al., 30 Sep 2025).

The appendix also compares LogiCAM with a multimodalized Logic-LM adaptation. There, Logic-LM + VLM reaches 35.14 on PL and 32.65 on FOL, whereas LogiCAM reaches 60.44 on PL and 42.55 on FOL. This comparison is used to support the claim that naive image-to-text conversion before theorem proving discards too much relevant multimodal information (Xu et al., 30 Sep 2025).

6. Failure modes, limitations, and relation to adjacent approaches

The paper’s error analysis identifies six major failure categories: incorrect application of logical rules, failure to supplement with commonsense / rule misgeneralization, overlooking visual details, premise integration / alignment errors, heuristic shortcuts over formal logic, and visual perception / object recognition errors. The dominant issue is cross-modal alignment. In the abstract, the authors state that around 70% of failures stem from logical misalignment between modalities. The manual analysis gives model-specific alignment-error rates of 67% for LogiCAM, 74% for GPT-4.1, and 63% for InternVL (Xu et al., 30 Sep 2025).

Across logic types, the paper reports alignment errors of 79% for NM and 68% for PL. For FOL, the notable residual error types are overlooking errors: 16% and logical rule errors: 17%. LogiCAM also shows relatively high heuristic shortcut errors at 13%, which the paper attributes to the framework’s explicit attempt to combine symbolic and heuristic reasoning: the system sometimes chooses heuristics where formal logic would have been preferable (Xu et al., 30 Sep 2025).

These observations define the framework’s main limitations. LogiCAM remains an inference-time prompting framework rather than a learned symbolic solver with explicit formal guarantees. Its performance still declines at longer reasoning depths. Its central bottleneck remains the alignment of visual evidence with symbolic premises. The paper accordingly proposes tighter cross-modal architectures trained with logic-grounded objectives, more integrated multimodal LLM-plus-solver frameworks, and reasoning-focused training beyond surface-form CoT optimization (Xu et al., 30 Sep 2025).

In comparative perspective, LogiCAM is best understood as one member of a broader neurosymbolic trend rather than as a universal template. LogicAD moves from image descriptions to formal predicates and uses Prover9 for contradiction-based anomaly detection and explanation (Jin et al., 3 Jan 2025). CodeLogician uses LLMs to construct explicit formal models and delegates exact reasoning to ImandraX, with proofs and counterexamples over program logic (Lin et al., 17 Jan 2026). LogicCLIP, by contrast, emphasizes training-time improvement of logical sensitivity in joint embeddings rather than iterative symbolic prompting (Zhou et al., 15 Aug 2025). This suggests that LogiCAM’s distinctive contribution lies in its modular control of multimodal reasoning steps: selecting premises, deciding when formal rules are applicable, deriving new knowledge iteratively, and feeding inferred facts back into context without leaving the VLM-based reasoning loop.