A paired-image benchmark reveals that many MLLMs fail to update predictions when task-critical visual evidence changes, even when they answer individual images correctly.
hub Canonical reference
Adaptive Chain-of-Focus Reasoning via Dynamic Visual Search and Zooming for Efficient VLMs
Canonical reference. 100% of citing Pith papers cite this work as background.
35
Pith papers citing it
Background
100% of classified citations
abstract
Vision language models (VLMs) have achieved impressive performance across a variety of computer vision tasks. However, the multimodal reasoning capability has not been fully explored in existing models. In this paper, we propose a Chain-of-Focus (CoF) method that allows VLMs to perform adaptive focusing and zooming in on key image regions based on obtained visual cues and the given questions, achieving efficient multimodal reasoning. To enable this CoF capability, we present a two-stage training pipeline, including supervised fine-tuning (SFT) and reinforcement learning (RL). In the SFT stage, we construct the MM-CoF dataset, comprising 3K samples derived from a visual agent designed to adaptively identify key regions to solve visual tasks with different image resolutions and questions. We use MM-CoF to fine-tune the Qwen2.5-VL model for cold start. In the RL stage, we leverage the outcome accuracies and formats as rewards to update the Qwen2.5-VL model, enabling further refining the search and reasoning strategy of models without human priors. Our model achieves significant improvements on multiple benchmarks. On the V* benchmark that requires strong visual reasoning capability, our model outperforms existing VLMs by 5% among 8 image resolutions ranging from 224 to 4K, demonstrating the effectiveness of the proposed CoF method and facilitating the more efficient deployment of VLMs in practical applications.
hub tools
citation-role summary
background 8
citation-polarity summary
verdicts
UNVERDICTED 35roles
background 8polarities
background 8representative citing papers
Visual CoT agents exhibit tool-use collapse where tool usage declines but task accuracy rises, and adding entropy regularization for rollout diversity produces the strongest performance.
CaST-Bench creates a benchmark with causal-chain annotations and novel metrics showing that current VLMs struggle to construct precise grounded causal chains in video QA.
UniPath adaptively models coordination-path diversity in unified multimodal models by training a path-conditioned executor and using a lightweight planner for input-dependent selection, improving performance over fixed strategies.
GazeVLM introduces internal gaze tokens that allow VLMs to dynamically suppress irrelevant visual features and simulate foveal attention for improved high-resolution multimodal reasoning.
LAGO achieves state-of-the-art zero-shot performance with fewer image regions by using class-agnostic object discovery followed by confidence-controlled language-guided refinement and dual-channel aggregation.
DMLR performs dynamic visual-textual interleaving in latent space using confidence-guided latent policy gradient optimization and a dynamic visual injection strategy, yielding improved multimodal reasoning on benchmarks.
IMAgent trains a multi-image vision agent via pure end-to-end RL with visual reflection tools and a two-layer motion trajectory masking strategy, reaching SOTA on single- and multi-image benchmarks while revealing tool-use effects on attention.
ReVisIT achieves near-SOTA performance on open multimodal tasks by retrieving and reasoning over labeled images as visual exemplars in a train-free scaffold, closing the open-vs-closed gap for models like Qwen3-VL-30B.
PixelEyes decouples reasoning and perception via mask-guided search and semantic BFS, introduces PixelEyes-6K dataset and Pinpoint-Bench benchmark, and open-sources code and models.
Maestro uses outcome-based RL to train a lightweight policy that orchestrates ensembles of frozen expert models and skills, reporting 70.1% average accuracy across ten multimodal benchmarks and outperforming GPT-5 and Gemini-2.5-Pro while generalizing to unseen components.
Introduces Zoom-then-Diagnose paradigm and uncertainty-aware reward in GRPO for confidence-aware ultrasound VQA, reporting 39.3% improvement in lesion localization across liver, breast, and thyroid datasets.
SCOLAR fixes information gain collapse in latent visual reasoning by generating independent auxiliary visual tokens via a detransformer, extending acceptable CoT length over 30x and delivering +14.12% gains on reasoning benchmarks.
Imagining in 360° decouples visual search into a single-step probabilistic semantic layout predictor and an actor, removing the need for multi-turn CoT reasoning and trajectory annotations while improving efficiency in 360° environments.
In LVLMs, attention can be replaced by random Gaussian weights with little or no performance loss, indicating that current models get lost in attention rather than efficiently using visual context.
Foveated Reasoner integrates foveation as stateful actions inside the autoregressive decoding loop of vision-language models, trained via cold-start supervision then reinforcement learning to achieve higher accuracy at low token budgets.
ToolsRL trains MLLMs via a tool-specific then accuracy-focused RL curriculum to master visual tools for complex reasoning tasks.
Zoom consistency provides a geometric, cross-model confidence signal in zoom-in grounding pipelines that correlates with prediction correctness and enables modest gains in specialist-generalist routing.
Visual replay module and adaptive depth scaling improve multimodal latent reasoning, reaching SOTA benchmarks with faster inference than explicit chain-of-thought methods.
ReAlign improves visual document retrieval by training retrievers to match query-induced rankings with rankings derived from VLM-generated, region-focused descriptions of relevant page content.
MAPO improves multimodal chain-of-thought reasoning by requiring explicit textual descriptions of visual tool results and using a novel advantage estimator that combines semantic alignment with task rewards.
Saliency-R1 uses a novel saliency map technique and GRPO with human bounding-box overlap as reward to improve VLM reasoning faithfulness and interpretability.
Skyra is an MLLM that detects AI-generated videos by identifying and reasoning over grounded visual artifacts, supported by a new annotated dataset and benchmark.
CropVLM uses reinforcement learning to learn image zooming policies that boost fine-grained perception in VLMs on out-of-domain high-resolution tasks without labeled boxes, synthetic data, or VLM changes.
citing papers explorer
-
VisualFLIP: Do Predictions Depend on Task-Critical Visual Evidence in Multimodal Reasoning?
A paired-image benchmark reveals that many MLLMs fail to update predictions when task-critical visual evidence changes, even when they answer individual images correctly.
-
Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents
Visual CoT agents exhibit tool-use collapse where tool usage declines but task accuracy rises, and adding entropy regularization for rollout diversity produces the strongest performance.
-
CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering
CaST-Bench creates a benchmark with causal-chain annotations and novel metrics showing that current VLMs struggle to construct precise grounded causal chains in video QA.
-
UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning
UniPath adaptively models coordination-path diversity in unified multimodal models by training a path-conditioned executor and using a lightweight planner for input-dependent selection, improving performance over fixed strategies.
-
GazeVLM: Active Vision via Internal Attention Control for Multimodal Reasoning
GazeVLM introduces internal gaze tokens that allow VLMs to dynamically suppress irrelevant visual features and simulate foveal attention for improved high-resolution multimodal reasoning.
-
LAGO: Language-Guided Adaptive Object-Region Focus for Zero-Shot Visual-Text Alignment
LAGO achieves state-of-the-art zero-shot performance with fewer image regions by using class-agnostic object discovery followed by confidence-controlled language-guided refinement and dual-channel aggregation.
-
Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space
DMLR performs dynamic visual-textual interleaving in latent space using confidence-guided latent policy gradient optimization and a dynamic visual injection strategy, yielding improved multimodal reasoning on benchmarks.
-
Training Multi-Image Vision Agents via End2End Reinforcement Learning
IMAgent trains a multi-image vision agent via pure end-to-end RL with visual reflection tools and a two-layer motion trajectory masking strategy, reaching SOTA on single- and multi-image benchmarks while revealing tool-use effects on attention.
-
Retrieved Images as Visual Thought: Training-Free Multimodal In-Context Learning for the Open-vs-Closed Gap
ReVisIT achieves near-SOTA performance on open multimodal tasks by retrieving and reasoning over labeled images as visual exemplars in a train-free scaffold, closing the open-vs-closed gap for models like Qwen3-VL-30B.
-
PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking
PixelEyes decouples reasoning and perception via mask-guided search and semantic BFS, introduces PixelEyes-6K dataset and Pinpoint-Bench benchmark, and open-sources code and models.
-
Maestro: Reinforcement Learning to Orchestrate Hierarchical Model-Skill Ensembles
Maestro uses outcome-based RL to train a lightweight policy that orchestrates ensembles of frozen expert models and skills, reporting 70.1% average accuracy across ten multimodal benchmarks and outperforming GPT-5 and Gemini-2.5-Pro while generalizing to unseen components.
-
Look-Closer-Then-Diagnose: Confidence-Aware Ultrasound VQA via Active Zooming
Introduces Zoom-then-Diagnose paradigm and uncertainty-aware reward in GRPO for confidence-aware ultrasound VQA, reporting 39.3% improvement in lesion localization across liver, breast, and thyroid datasets.
-
Self-Consistent Latent Reasoning: Long Latent Sequence Reasoning for Vision-Language Model
SCOLAR fixes information gain collapse in latent visual reasoning by generating independent auxiliary visual tokens via a detransformer, extending acceptable CoT length over 30x and delivering +14.12% gains on reasoning benchmarks.
-
Beyond Thinking: Imagining in 360$^\circ$ for Humanoid Visual Search
Imagining in 360° decouples visual search into a single-step probabilistic semantic layout predictor and an actor, removing the need for multi-turn CoT reasoning and trajectory annotations while improving efficiency in 360° environments.
-
Large Vision-Language Models Get Lost in Attention
In LVLMs, attention can be replaced by random Gaussian weights with little or no performance loss, indicating that current models get lost in attention rather than efficiently using visual context.
-
Foveated Reasoning: Stateful, Action-based Visual Focusing for Vision-Language Models
Foveated Reasoner integrates foveation as stateful actions inside the autoregressive decoding loop of vision-language models, trained via cold-start supervision then reinforcement learning to achieve higher accuracy at low token budgets.
-
Visual Reasoning through Tool-supervised Reinforcement Learning
ToolsRL trains MLLMs via a tool-specific then accuracy-focused RL curriculum to master visual tools for complex reasoning tasks.
-
Zoom Consistency: A Free Confidence Signal in Multi-Step Visual Grounding Pipelines
Zoom consistency provides a geometric, cross-model confidence signal in zoom-in grounding pipelines that correlates with prediction correctness and enables modest gains in specialist-generalist routing.
-
Visual Enhanced Depth Scaling for Multimodal Latent Reasoning
Visual replay module and adaptive depth scaling improve multimodal latent reasoning, reaching SOTA benchmarks with faster inference than explicit chain-of-thought methods.
-
ReAlign: Optimizing the Visual Document Retriever with Reasoning-Guided Fine-Grained Alignment
ReAlign improves visual document retrieval by training retrievers to match query-induced rankings with rankings derived from VLM-generated, region-focused descriptions of relevant page content.
-
Walk the Talk: Bridging the Reasoning-Action Gap for Thinking with Images via Multimodal Agentic Policy Optimization
MAPO improves multimodal chain-of-thought reasoning by requiring explicit textual descriptions of visual tool results and using a novel advantage estimator that combines semantic alignment with task rewards.
-
Saliency-R1: Enforcing Interpretable and Faithful Vision-language Reasoning via Saliency-map Alignment Reward
Saliency-R1 uses a novel saliency map technique and GRPO with human bounding-box overlap as reward to improve VLM reasoning faithfulness and interpretability.
-
Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning
Skyra is an MLLM that detects AI-generated videos by identifying and reasoning over grounded visual artifacts, supported by a new annotated dataset and benchmark.
-
CropVLM: Learning to Zoom for Fine-Grained Vision-Language Perception
CropVLM uses reinforcement learning to learn image zooming policies that boost fine-grained perception in VLMs on out-of-domain high-resolution tasks without labeled boxes, synthetic data, or VLM changes.
-
DeepEyesV2: Toward Agentic Multimodal Model
DeepEyesV2 uses a two-stage cold-start plus reinforcement learning pipeline to produce an agentic multimodal model that adaptively invokes tools and outperforms direct RL on real-world reasoning benchmarks.
-
RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning
RSICCLLM introduces a post-training framework with RSICI dataset, difference-aware supervised fine-tuning, and dual-negative preference optimization that claims to outperform much larger models on remote sensing image change captioning.
-
From Structure to Synergy: A Survey of Vision-Language Perception Paradigm Evolution in Multimodal Large Language Models
The survey formalizes MLLM perception as a unified vision-language capability and traces its evolution via a new five-stage taxonomy while outlining future challenges.
-
Latent Visual States for Efficient Multimodal Reasoning
EVA generates adaptive Latent_slot tokens as internal visual thoughts, trained end-to-end with text tokens via D-GSPO on the EVA-230K dataset, claiming performance gains and better inference efficiency.
-
CaptchaMind: Training CAPTCHA Solvers via Reinforcement Learning with Explicit Reasoning Supervision
Presents CaptchaBench benchmark and CaptchaMind RL solver achieving 82.9% success on benchmark tasks and 71% on real-world CAPTCHAs via explicit reasoning process supervision.
-
Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth
Constraining visual token budget per observation during VLM training forces genuine active perception and delivers 5% average relative improvement without auxiliary losses or architecture changes.
-
Reversing the Flow: Generation-to-Understanding Synergy in Large Multimodal Models
Generation-to-Understanding synergy lets multimodal models create self-generated visual edits as intermediate steps, improving performance on twelve benchmarks while revealing limits in task-aligned self-reflection.
-
MedLVR: Latent Visual Reasoning for Reliable Medical Visual Question Answering
MedLVR interleaves latent visual reasoning segments in autoregressive decoding and uses two-stage training to raise average medical VQA accuracy from 48.3% to 53.4% over a Qwen2.5-VL-7B backbone on OmniMedVQA and five other benchmarks.
-
Mixture-of-Visual-Thoughts: Exploring Context-Adaptive Reasoning Mode Selection for General Visual Reasoning
MoVT unifies different visual reasoning modes in a single model and uses the AdaVaR two-stage framework with supervised cold-start and RL via AdaGRPO to enable context-adaptive mode selection, yielding consistent gains on visual reasoning tasks.
-
Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search
Mini-o3 scales visual search reasoning to tens of interaction turns via a new probe dataset, iterative trajectory collection, and over-turn masking in RL, claiming SOTA performance while training only up to six turns.
-
UniReason-Med: A Shared Grounded Reasoning Interface for 2D-to-3D Transfer in Medical VQA
UniReason-Med introduces a unified framework for 2D and 3D medical VQA with shared grounded reasoning, trained on a 220K dataset, claiming that joint 2D+3D supervision improves 3D performance over 3D-only training.