Token-Efficient Multimodal Reasoning via Image Prompt Packaging

Multimodal large language models have advanced rapidly, but most remain optimized for single-image inputs. Multi-image reasoning is costly because each image adds many visual tokens, increasing encoding and attention demands. Existing techniques that address the above issue, including visual token pruning, token merging, visual token compression, learned resampling, hierarchical compression, retrieval-augmented visual reasoning, and memory-based architectures, have largely been designed for single-image scenarios, where redundancy is primarily spatial and local. Multi-image reasoning introduces additional challenges: redundancy may occur across images, task-relevant evidence may be sparse or distributed, and aggressive per-image compression can remove details needed for comparison, counting, temporal reasoning, retrieval, or evidence grounding. Moreover, existing token-reduction methods are often either (i) training-free but inference-inefficient, typically relying on accessing attention matrices to get token-importance that may be impractical to access with memory-efficient attention implementations such as FlashAttention, or (ii) training-dependent, requiring the base MLLM to be retrained or adapted to a new compressed-token paradigm. In this work, we study scalable multi-image reasoning and understanding without retraining the base model and without accessing attention matrices at inference time. To this end, we propose mmVB, a lightweight plug-and-play module that assigns query-conditioned per-image token budgets via cross-image context, adding only 1.6M trainable parameters while reducing visual input tokens by 9x. This reduction yields 5x faster time-to-first-token and 7x fewer prefill FLOPs, while maintaining 94.4% of full-resolution accuracy across eight multi-image reasoning benchmarks.

Multimodal large language models increasingly reason over screenshots and documents where the task itself may be written in pixels. Yet benchmarks usually place questions in text, leaving it unclear whether models use the same instruction equally well across channels. We introduce Visualized Task Semantics (VTS), a controlled intervention that moves the question into the image while keeping the source problem and answer fixed. Across six MLLMs and four benchmarks, accuracy drops in all 24 model-task pairs, by 17.8 points on average. Models often transcribe the visual question correctly yet fail to use it, exposing a semantic channel gap beyond OCR. To reduce this gap, we present prompt-region grounding, whose core design aligns the question region with typed semantics and recovers its clean representation from a masked view. At matched training cost, our method raises four-benchmark VTS accuracy from 58.0 to 66.3 while preserving accuracy on the original interface, and requires no OCR or region metadata at inference. Reading task-bearing text and grounding it as an instruction for reasoning are distinct capabilities.

Multimodal large reasoning models often rely on long Chain-of-Thought (CoT) traces in which a substantial fraction of tokens, such as repeated visual descriptions, self-reflection, and other visually-disengaged filler, inflate inference cost without contributing to the answer. Existing CoT compression methods optimize output length but never measure whether a reasoning token is actually grounded in the image. We propose VIG (Visual Information Gain), an information-theoretic GRPO reward that scores each reasoning token by how much the image reduces its predictive uncertainty. VIG is computed online from two forward passes of the same policy, one with and one without the image, so no reference chains, external annotations, or auxiliary reward models are needed. Across six main multimodal reasoning benchmarks and three Qwen3-VL-Thinking model sizes (2B/4B/8B), plus an additional R1-Onevision-Bench evaluation on 8B, VIG consistently improves the accuracy--efficiency trade-off, supporting our central claim: efficient multimodal reasoning emerges from raising visual information density, where every reasoning token earns its place by anchoring to the image, rather than from imposing a length budget. Our source code is available at https://github.com/chaser682/vig.