acceptodds
Under review as a conference paper at ICLR 2027

Thinking with Comics: Efficient Multimodal Reasoning through Sequential Visual Narratives

Abstract

Chain-of-Thought reasoning has driven large language models to extend from thinking with text to thinking with images and videos. However, different modalities still have clear limitations: static images struggle to represent temporal structure and dynamic processes, while videos introduce substantial redundancy and computational cost. In this work, we propose Thinking with Comics (TwC), a visual reasoning paradigm that uses comics as a high information-density medium positioned between images and videos. Comics preserve temporal structure, embedded text, and narrative coherence while requiring significantly lower reasoning cost. To systematically evaluate reasoning over sequential visual narratives, we introduce GutterBench, a benchmark that measures whether models can represent, update, and apply intermediate states across comic panels rather than merely predict final answers. We further evaluate TwC across diverse reasoning tasks and visual understanding scenarios. Experimental results show that Thinking with Comics outperforms Thinking with Images on multi-step temporal and causal reasoning tasks, while remaining substantially more efficient than Thinking with Video. Further analysis reveals that comics provide an effective intermediate representation for modeling cross-panel state transitions, offering a promising direction for multimodal reasoning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Reasoning, Chain-of-Thought, Visual Reasoning, Comics, Thinking with Images, Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_4MbROpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes Thinking with Comics (TwC), a generate-then-reason pipeline in which an image generator converts a question (and, for visually grounded tasks, its visual premise) into a single composite image containing a sequence of comic panels. The original question and generated comic are then provided to a multimodal language model to produce the final answer. The authors position comics between single images and videos: compared with a single image, comics can encode multiple states and temporal or causal structure; compared with video, they are intended to be less redundant and less expensive. The paper also introduces GutterBench, a 300-item benchmark with 200 text-centric and 100 vision-centric items spanning paper folding, die tipping, egocentric navigation, tile transformation, and gossip propagation. Items cover reasoning depths 2, 4, 6, and 8 and include final-state, counterfactual, and process-tally questions. In addition to final-answer accuracy, the benchmark uses gold storyboards and process metrics intended to measure state matching, order, and textual anchors. The main implementation uses Gemini-3 Pro Image to generate comics and Gemini-3 Pro to answer from them. The paper reports results on GutterBench and a broad collection of mathematical, document, cultural, spatial, and visual-puzzle benchmarks, as well as analyses of narrative style, panel count and order, generator choice, cross-model transfer, global versus incremental generation, and generation cost. The reported results show sizable gains over the paper's single-image and video baselines on several tasks, although TwC remains weak on pixel-precise tasks such as ARC-AGI-2 and non-square mazes.

Critical strengths
  1. The paper identifies a novel and intuitively useful intermediate representation. Comics can combine discrete state transitions, spatial layout, and text in a compact artifact. This is a meaningful design space between a single generated image and a full video, and the paper explores it across an unusually broad set of tasks and analysis dimensions.

  2. GutterBench is potentially a useful process-oriented contribution. The benchmark covers five controlled state-transition families, two input tracks, four depths, and multiple query types, and it includes gold storyboards rather than evaluating only final answers. The reported human verification, inter-annotator agreement, human ceiling, and template/leakage controls are valuable. If the benchmark, generator, prompts, and scoring code are released as promised, this component could support future work independently of the current TwC claims.

Critical weaknesses
  1. The method presupposes the difficult reasoning capability that it claims to provide. Generating a correct comic requires selecting the right intermediate states, arranging them causally, maintaining their consistency, and often displaying the final answer. If the generator succeeds, the downstream VLM is effectively reading a compact worked solution; if it fails, the representation can propagate or amplify the error. The paper does not measure this upstream faithfulness or compare against the generator's direct answer. Thus, the experiments demonstrate the usefulness of an already-correct intermediate solution, not comics as a general method for improving VLM reasoning.

  2. The paper does not identify what is essential about the representation, so the claim of a general comic-based reasoning paradigm is too strong. Comics simultaneously add discrete state decomposition, ordering, textual anchors, visual grouping, narrative cues, and often the answer itself. The existing ablations do not hold semantic content fixed across media, so they cannot show that the comic format is responsible for the gain. The more fundamental hypothesis may be that a sparse, ordered, multimodal state representation helps certain tasks. Establishing that hypothesis would require content-matched comparisons with numbered text, tables or state graphs, diagrams, and image sequences, plus a characterization of which task structures admit faithful compression. The current theory assumes rather than explains this compression and contains invalid derivational steps.

Questions for authors
  1. Before invoking the downstream VLM, how often does the generated comic already contain a correct final answer and a faithful sequence of intermediate states? Please report comic-level state/answer correctness and downstream accuracy conditioned on comic correctness. How does the same generator perform when asked to return a direct answer or a concise structured textual state trace instead of a comic?

  2. Can the authors compare content-matched representations by first fixing one intermediate trace and then rendering exactly that trace as (i) numbered text, (ii) a table or state graph, (iii) an unstructured image sequence, and (iv) a comic? What remains of the gain when the final answer and answer-bearing text are masked in every condition?

  3. What is the intended task class for which generating a faithful comic is easier or more reliable than directly solving the task? Can the authors formulate the relevant structural property—for example, a small number of discrete state transitions—and report results stratified by whether tasks satisfy it? The current failures on pixel-precise tasks suggest an important limitation to the claimed generality.

  4. Can the authors reformulate the theory around the actual bottleneck, the construction of a faithful compressed state representation, rather than assuming that the comic generator selects the relevant states? Please also correct Equation (6), justify any conclusion drawn from the domain-adaptation bounds, and reconcile the efficiency factors reported in Figure 5 with the underlying TFLOPs, latency, and prices.

Reviewer_6ZiNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Thinking with Comics (TwC), a training-free pipeline: an image generator (Gemini-3 Pro Image) draws a multi-panel comic depicting the reasoning, and an MLLM (Gemini-3 Pro) answers using the question plus the comic. It also introduces GutterBench, 300 state-tracking items with gold storyboards and process-level metrics. TwC reports strong results on MathVista (85.8), DocVQA (99.4), and GutterBench at much lower cost than video-based reasoning, but underperforms direct prompting on saturated text tasks (MATH-500, GSM8K).

Critical strengths
  1. GutterBench is a useful contribution independent of the method. Its controlled state trajectories with gold storyboards enable process-level evaluation (Trace-F1, Order-τ, Anchor) across text, images, video, and comics, with careful distractor design and high annotator agreement (κ=0.87).
Critical weaknesses
  1. Limited novelty. TwC is essentially Thinking with Images with a multi-panel prompt: it still generates a single image followed by MLLM reasoning. The paper itself notes that N=1 equals TwI, and its efficiency argument relies on the single-image property. The method change is only a prompt.

  2. The claimed gains are not attributed to the comic format. (a) The TwI baseline (G-IMG) has the MLLM write a T2I prompt, while TwC gives the question directly to a reasoning-capable generator that typically writes the answer in the last panel. The gain over TwI may therefore come from generator-side solving at roughly twice the compute. (b) There is no fair control (same generator, question given directly, one panel vs. multiple panels), no non-comic step-by-step diagram, no compute-matched baseline, and no faithfulness test. (c) Numbers are inconsistent: MATH-500 is reported as 92.3, 90.0, and 95.0; cost as "~20x" vs. 86.6%; and Fig. 1 claims a single fixed instruction, while Appendix D shows task-specific prompts.

Questions for authors
  1. With the same generator receiving the question directly, how does a single panel compare with multiple panels?
  2. What are the generator's own text-only accuracy and the last-panel-only accuracy? How does TwC compare with a compute-matched two-call baseline?
  3. How often does the reader follow an injected wrong step, and what is the accuracy without the final answer panel?
  4. Which MATH-500 number is correct, and which prompts produced the main results?
Reviewer_TA3UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Thinking with Comics (TwC), which generates a comic showing a solution and feeds it, together with the question, to a multimodal model. It also introduces GutterBench, a 300-item benchmark with answer and intermediate-state evaluation. Experiments report gains on MathVista, DocVQA, and GutterBench, alongside comparisons of narrative styles, panel structure, generators, and generation costs.

Critical strengths
  1. The composite comic representation is simple and practical. It can organize several intermediate states within one image, making it a useful alternative to generating video.
  2. GutterBench evaluates intermediate states as well as final answers. Its controlled depths, counterfactual queries, and human verification could support more informative comparisons of visual reasoning methods.
Critical weaknesses
  1. The source of the gains remains unclear. Generation prompts explicitly request the final answer, which often appears in the comic. The downstream model may therefore be reading a supplied answer rather than reasoning across panels. This limits support for the paper’s central claim.
  2. The controls do not isolate comic structure. Global versus incremental generation also changes visual style and text requirements, while the medium comparison does not establish matched reasoning content or budgets. The theoretical appendix does not resolve this gap: Eq. (6) is incorrect, and Eqs. (11)–(13) infer an ordering of actual errors from an ordering of upper bounds.
Questions for authors
  1. Clarification: Which tables use Gemini-3 Pro for downstream reasoning, and which use GPT-5.2 for answer extraction? Please specify evaluation sample counts and generation attempts.
  2. Assessment-critical evidence: How does full-comic accuracy compare with final-panel-only input, answer-masked comics, and a text transcription of the same comic? These comparisons would help determine whether intermediate visual states contribute beyond answer reading.
  3. Clarification: In Table 6, are reasoning content, embedded text, and generation budgets matched across media? What exactly changes between conditions?
Reviewer_ddtiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces "Thinking with Comics" (TwC), a method that represents problems and their step-by-step solutions as multi-panel comics, which are then processed by a multimodal model to derive an answer. The authors also developed "GutterBench"—a dataset of 300 problems—to evaluate both the final answers and the intermediate reasoning steps. Experiments demonstrate significant performance gains on MathVista, DocVQA, and GutterBench, with lower generation costs compared to video-based approaches; however, the advantages are more limited in certain mathematical and fine-grained spatial tasks.

Critical strengths

Comics offer an intuitive and compact format for representing multi-step reasoning. Given the same generator and answering model, the comic-based approach outperforms both single-image and unstructured multi-image representations, providing empirical support for this direction. GutterBench goes beyond merely checking the correctness of the final answer by also verifying intermediate states—a process supported by human validation—which allows for a more nuanced understanding of a model's reasoning capabilities.

Critical weaknesses

Since the comics themselves often explicitly state the final answer, it is difficult to determine whether the performance gains stem from cross-panel reasoning or simply from the answering model reading the pre-calculated answer provided by the generator. The current experimental controls fail to fully rule out this possibility, which weakens the persuasiveness of the core conclusions. There are inconsistencies in the description of the experimental setup; for instance, while the main text mentions the use of a unified prompt, the appendix provides task-specific prompts. Furthermore, the reported scales of the open-source video models are inconsistent. Additionally, the cost analysis focuses primarily on the generation phase, failing to capture the efficiency of the entire pipeline.

Questions for authors

How much of the performance gain persists if the final answer within the comic is masked? Furthermore, how would the model perform if the textual steps from the comics were provided directly to the answering model? Such controlled experiments would help clarify the specific contribution of the comic structure itself.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.