Lookahead Prediction-Guided Adaptive Visual Refocusing for Interleaved Multimodal Reasoning
Authors: Mengzhao Wang, Yanli Ji, Peng Ye
Organizations: Sun Yat-sen University · Shenzhen Loop Area Institute · Shanghai Artificial Intelligence Laboratory · The Chinese University of Hong Kong
Interleaved multimodal chain-of-thought reasoning improves the visual reasoning capabilities of multimodal large language models by dynamically incorporating visual evidence during reasoning. However, existing methods primarily select evidence from historical context. When the reasoning focus shifts, history-guided evidence may remain focused on prior regions and fail to cover the visual information required by the next step. We refer to this lag between visual evidence and subsequent reasoning demands as Temporal Evidence Myopia. To address this issue, we propose Lookahead Prediction-guided Adaptive Visual Refocusing (LPAVR), a training-free test-time framework. LPAVR transforms Multi-Token Prediction (MTP) from a decoding acceleration mechanism into a lookahead signal for anticipating upcoming reasoning intent. During MTP decoding, the draft head generates multiple future tokens, while the target model verifies them in parallel and retains the longest prefix consistent with its decoding trajectory. This target-verified prefix contains near-future semantics, which LPAVR uses to construct lookahead-guided visual evidence and assess whether history-guided evidence covers the indicated regions. If the coverage is sufficient, the history-guided evidence is retained; otherwise, LPAVR switches to the lookahead-guided evidence. This better aligns the visual context with upcoming reasoning and improves consistency between draft predictions and target-model decoding. Experiments on multiple multimodal reasoning benchmarks show that LPAVR improves both reasoning accuracy and the acceptance rates of MTP speculative tokens, jointly enhancing visual reasoning and speculative decoding efficiency.
Figures & tables
Figure 1: Motivation and overview of LPAVR. (a) Temporal Evidence Myopia. (b) History-guided and next-step visual evidence often have low overlap. (c) Lookahead-guided evidence improves alignment with next-step evidence. (d) General ICoT performs reactive, history-guided visual selection. (e) LPAVR performs lookahead-guided adaptive visual refocusing.
Figure 2: Overview of the LPAVR framework. (a) SVED performs provisional probing with history-guided visual evidence and estimates upcoming visual demands through target-verified MTP tokens, then a coverage-based gate selects either history-guided or lookahead-guided visual evidence. (b) SVCI keeps the target model and the MTP draft head synchronized during probing and finalized decoding.
Model
#Params
Method
M 3 CoT
MMStar
MathVista
Hallusion Bench
CS
MM
SC
ALL
Training-Based Methods
Qwen2-VL
7B
MINT-CoT ( Chen et al. 2026 )
80.07
33.61
58.38
60.09
58.13
75.80
64.67
Qwen2.5-VL
7B
R1-Onevision ( Yang et al. 2025 )
75.38
33.61
62.21
61.82
56.60
63.40
65.40
Llama-3.2
11B
LLaVA-CoT ( Xu et al. 2025 )
79.78
35.68
55.24
58.02
57.00
51.40
63.72
Qwen2.5-VL
7B
DeepEyesV2 ( Hong et al. 2026 )
78.68
67.88
54.36
68.59
62.00
70.90
49.32
Table 1: Multimodal reasoning accuracy across benchmarks. We compare training-based and training-free methods on M 3 CoT, MMStar, MathVista, and HallusionBench to evaluate their multimodal understanding and reasoning performance. The best and second-best results within each comparable model group are marked in bold and underlined , respectively. For M 3 CoT, CS, MM, and SC denote Common Sense, Mathematics, and Science.
Dataset
Method
P2
P3
P4
P5
P6
M 3 CoT
ICoT
74.35
43.40
32.02
25.20
21.20
LPAVR
74.90
48.02
36.24
28.45
24.22
Δ (vs ICoT)
+0.55
+4.62
+4.22
+3.25
+3.02
MMStar
ICoT
68.49
42.50
26.09
17.67
14.01
LPAVR
69.83
43.01
27.39
20.46
16.25
Δ (vs ICoT)
+1.34
+0.51
+1.30
+2.79
+2.24
Table 2: Position-wise acceptance rates of MTP lookahead predictions. We compare ICoT and LPAVR on Qwen3.5-9B at speculative positions P2–P6.
Method
M 3 CoT
MMStar
MathVista
No Refocusing
69.54
70.20
74.70
History-Guided Refocusing
72.69
71.40
75.80
Lookahead-Guided Refocusing
72.72
73.67
76.50
Adaptive Refocusing (LPAVR)
74.12
74.33
77.30
Table 3: Ablation study of visual refocusing strategies. We compare different visual refocusing variants on Qwen3.5-9B and report their overall accuracy on three multimodal reasoning benchmarks.
Method
FLOPs (TFLOPs)
Latency (s/sample)
Ours w/o SVCI
6.6
8.9
Ours (with SVCI)
5.2
7.6
Table 4: Ablation study of SVCI. We compare LPAVR with and without SVCI under identical settings. Without SVCI, full recomputation is performed after evidence switching.
Figure 3: End-to-end efficiency comparison. Latency and throughput of CoT, ICoT, ICoT+MTP, and LPAVR under identical hardware and decoding settings.
Figure 5: Analysis of lookahead configurations. (a) Accuracy and decoding speed under different lookahead horizons. (b) Accuracy and per-sample inference time under different numbers of iterative probing rounds.
Multimodal chain-of-thought reasoning studies how AI models combine visual, audio, and textual evidence through intermediate reasoning to solve complex tasks.
Processing long visual token sequences from high-resolution images makes multi-step reasoning computationally expensive for multimodal Large Language Models (MLLMs). Existing one-shot pruning and aggregation methods compress visual tokens into a fixed context before decoding. However, visual evidence needs can shift as reasoning unfolds, making it difficult for a fixed compressed context to retain all the details needed across stages. To address this challenge, we propose ViMoD, a lightweight framework that maintains a compact visual context while preserving access to original fine-grained evidence as reasoning needs evolve. Deformable Aggregation of Region-wise Tokens (DART) learns content-adaptive groups and aggregation capacities, constructing compact Coarse representations linked to recoverable original Fine tokens. Temporal Routing for Adaptive Contextual Evidence (TRACE) integrates decoding history to anticipate upcoming evidence needs and select, retain, or replace active Fine-token groups. Selected Fine tokens augment the persistent Coarse context in the frozen backbone, enabling stage-specific evidence access without continuously attending to all visual tokens. On Qwen3-VL-4B, ViMoD outperforms all evaluated baselines on all eight reasoning benchmarks at a 20% target visual token budget, improving the mean normalized score by 39.0% over the strongest evaluated one-shot baseline. These gains are achieved with only 0.0546% additional trainable parameters relative to the frozen backbone.
Yicheng Xue, Han Wu, Jufeng Yang +4
City University of Hong Kong · Zhejiang University · Peking University +2
With the rapid advancement of large multimodal models (LMMs), inference-time overhead has become a key bottleneck for real-world deployment. Existing methods typically prune visual tokens at prefill, assuming the required visual evidence remains static during reasoning. However, we empirically show that visual evidence is strongly step-dependent: only a sparse subset of visual tokens is critical at each decoding step, and the critical set evolves across reasoning. Furthermore, we identify a coupled bottleneck where redundant visual context can steer the model toward query-irrelevant regions, lengthening the reasoning trace. Guided by these insights, we propose VisionPulse, a step-wise visual token pruning framework during reasoning. VisionPulse computes a lightweight visual attention mass to estimate the step-wise retention budget by exploiting its strong positive correlation with LMMs' effective visual token usage and retain only the most critical tokens under this budget. By enforcing visual sparsity during reasoning, VisionPulse filters redundant visual context while preserving relevant visual evidence, shortening reasoning traces naturally. Extensive experiments show that VisionPulse only retains 5% of visual tokens per step with reasoning traces shortened by 11.2%, while keeping accuracy almost unchanged.
Hengbo Xu, Shengjie Jin, Yanbiao Ma +1
Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China.
Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions. Recent thinking with images approaches mitigate this limitation by allowing models to inspect images through operations such as zooming, cropping, or code execution. While effective, their multi-turn interactions with VLMs introduce redundant token computation and slow inference, and separate region inspection from the reasoning step that depends on it. To address these issues, we propose Look Before You Reason (LoRe), an implicit visual thinking framework for efficient visual reasoning within a single-turn VLM interaction. Given an image and a question, LoRe predicts the most relevant image region, aligns it with the corresponding image patch features, and strengthens the associated key-value (KV) cache entries, enabling subsequent decoding to directly exploit the selected visual cues without additional visual inputs or explicit tool calls. We further employ reinforcement learning to guide the VLM toward identifying informative image regions and leveraging them during reasoning. To support fine-grained evaluation of multimodal reasoning, we introduce COCO-ObjQA, a large-scale benchmark for object-centric multimodal reasoning. Experiments on three challenging multimodal benchmarks demonstrate that LoRe consistently improves VLM reasoning performance with low additional inference overhead.