P2R decouples perception from reasoning in VLMs via a two-stage process and PRA-GRPO alternating RL training, reporting gains such as 93.2% on V-Star for the 4B model over its Qwen3-VL backbone.
hub Mixed citations
Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning
Mixed citation behavior. Most common role is background (50%).
67
Pith papers citing it
1
external citations · Pith
Background
50% of classified citations
abstract
Chain-of-thought reasoning has significantly improved the performance of Large Language Models (LLMs) across various domains. However, this reasoning process has been confined exclusively to textual space, limiting its effectiveness in visually intensive tasks. To address this limitation, we introduce the concept of reasoning in the pixel-space. Within this novel framework, Vision-Language Models (VLMs) are equipped with a suite of visual reasoning operations, such as zoom-in and select-frame. These operations enable VLMs to directly inspect, interrogate, and infer from visual evidences, thereby enhancing reasoning fidelity for visual tasks. Cultivating such pixel-space reasoning capabilities in VLMs presents notable challenges, including the model's initially imbalanced competence and its reluctance to adopt the newly introduced pixel-space operations. We address these challenges through a two-phase training approach. The first phase employs instruction tuning on synthesized reasoning traces to familiarize the model with the novel visual operations. Following this, a reinforcement learning (RL) phase leverages a curiosity-driven reward scheme to balance exploration between pixel-space reasoning and textual reasoning. With these visual operations, VLMs can interact with complex visual inputs, such as information-rich images or videos to proactively gather necessary information. We demonstrate that this approach significantly improves VLM performance across diverse visual reasoning benchmarks. Our 7B model, \model, achieves 84\% on V* bench, 74\% on TallyQA-Complex, and 84\% on InfographicsVQA, marking the highest accuracy achieved by any open-source model to date. These results highlight the importance of pixel-space reasoning and the effectiveness of our framework.
hub tools
citation-role summary
background 10
baseline 6
dataset 1
other 1
citation-polarity summary
representative citing papers
ZPPO improves distillation to small vision-language models by using binary and negative candidate prompts plus a replay buffer for hard questions, outperforming standard distillation and GRPO on a 31-benchmark suite with largest gains at the 0.8B scale.
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.
Astra couples an RL-trained VLM policy with a view-consistent Bagel-based world simulator to enable agentic imagination during spatial reasoning, yielding benchmark gains on MMSI-Bench and MindCube.
DeepLatent introduces a parallel latent visual reasoning framework with learnable 2D tokens and continuous RL, trained via distillation then RL, plus a new 180K dataset, claiming SOTA benchmark results.
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.
Proposes Spatial Narrative Score (SNS) evaluation for VLMs' camera motion understanding and introduces CaMo model achieving consistent performance on SNS and direct QA.
ATLAS uses a single functional token to unify agentic and latent visual reasoning without image generation or external execution.
UniVLR unifies textual and visual reasoning in multimodal LLMs by compressing reasoning traces and auxiliary images into visual latent tokens for direct inference without interleaved text CoT.
V-ABS is an action-observer beam search method with entropy-based adaptive weighting and an 80k-sample SFT dataset that delivers 19.7% average gains on visual reasoning tasks for MLLMs.
GazeVLM introduces internal gaze tokens that allow VLMs to dynamically suppress irrelevant visual features and simulate foveal attention for improved high-resolution multimodal reasoning.
Pest-Thinker is a reinforcement learning framework that improves MLLMs' expert-level reasoning on pest morphology via synthesized CoT trajectories, GRPO optimization, and an LLM-judged feature reward on new benchmarks QFSD and AgriInsect.
V-Reflection introduces a think-then-look mechanism where MLLM latent states actively interrogate visual features via two-stage distillation from a box-guided teacher to a dynamic autoregressive student, narrowing the fine-grained perception gap on benchmarks.
DLEBench is the first benchmark for small-scale object editing in instruction-based image editing models, using 1889 samples, seven instruction types, and a dual-mode evaluation protocol to reveal performance gaps in 10 tested models.
Q-Probe introduces the first agentic IQA framework that scales to high resolutions using context-aware probing, a new Vista-Bench benchmark, and three-stage training to reach state-of-the-art performance across scales.
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.
DeepEyes uses reinforcement learning to teach vision-language models active perception and image-based thinking, yielding gains on perception, reasoning, grounding, and hallucination benchmarks.
Modeling sparse visual state updates instead of full images cuts generated visual tokens ~55.6% and improves interleaved multimodal reasoning over full-image ULMMs.
A VLM planner that adaptively inserts latent visual thoughts of future states into its reasoning trace beats language-only and prior VLM planners on long-horizon kitchen tasks, especially under tight free space.
BUS trains VLMs for self-reflection without labels by verifying which sampled reasonings precede their own answers via backward prediction, improving multimodal reasoning.
SegAnswer trains an MLLM to generate segmentation masks instead of bounding boxes when zooming into image regions during visual reasoning, yielding consistent improvements across perception and hallucination benchmarks.
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.
ViToS uses dual-stream RL with cross-feedback optimization to prune medical image tokens to 77% length while reporting 108.27% and 104.16% relative performance on two 7B VLMs across seven benchmarks.
citing papers explorer
-
Perceive-to-Reason: Decoupling Perception and Reasoning for Fine-Grained Visual Reasoning
P2R decouples perception from reasoning in VLMs via a two-stage process and PRA-GRPO alternating RL training, reporting gains such as 93.2% on V-Star for the 4B model over its Qwen3-VL backbone.
-
Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients
ZPPO improves distillation to small vision-language models by using binary and negative candidate prompts plus a replay buffer for hard questions, outperforming standard distillation and GRPO on a 31-benchmark suite with largest gains at the 0.8B scale.
-
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.
-
Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators
Astra couples an RL-trained VLM policy with a view-consistent Bagel-based world simulator to enable agentic imagination during spatial reasoning, yielding benchmark gains on MMSI-Bench and MindCube.
-
DeepLatent: Think with Images via Parallel Latent Visual Reasoning
DeepLatent introduces a parallel latent visual reasoning framework with learnable 2D tokens and continuous RL, trained via distillation then RL, plus a new 180K dataset, claiming SOTA benchmark results.
-
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.
-
CaMo: Camera Motion Grounded Evaluation and Training for Vision-Language Models
Proposes Spatial Narrative Score (SNS) evaluation for VLMs' camera motion understanding and introduces CaMo model achieving consistent performance on SNS and direct QA.
-
ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both
ATLAS uses a single functional token to unify agentic and latent visual reasoning without image generation or external execution.
-
UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs
UniVLR unifies textual and visual reasoning in multimodal LLMs by compressing reasoning traces and auxiliary images into visual latent tokens for direct inference without interleaved text CoT.
-
V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning
V-ABS is an action-observer beam search method with entropy-based adaptive weighting and an 80k-sample SFT dataset that delivers 19.7% average gains on visual reasoning tasks for MLLMs.
-
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.
-
Pest-Thinker: Learning to Think and Reason like Entomologists via Reinforcement Learning
Pest-Thinker is a reinforcement learning framework that improves MLLMs' expert-level reasoning on pest morphology via synthesized CoT trajectories, GRPO optimization, and an LLM-judged feature reward on new benchmarks QFSD and AgriInsect.
-
V-Reflection: Transforming MLLMs from Passive Observers to Active Interrogators
V-Reflection introduces a think-then-look mechanism where MLLM latent states actively interrogate visual features via two-stage distillation from a box-guided teacher to a dynamic autoregressive student, narrowing the fine-grained perception gap on benchmarks.
-
DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model
DLEBench is the first benchmark for small-scale object editing in instruction-based image editing models, using 1889 samples, seven instruction types, and a dual-mode evaluation protocol to reveal performance gaps in 10 tested models.
-
Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing
Q-Probe introduces the first agentic IQA framework that scales to high resolutions using context-aware probing, a new Vista-Bench benchmark, and three-stage training to reach state-of-the-art performance across scales.
-
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.
-
DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning
DeepEyes uses reinforcement learning to teach vision-language models active perception and image-based thinking, yielding gains on perception, reasoning, grounding, and hallucination benchmarks.
-
DeltaV: Thinking with Visual State Updates in Unified Large Multimodal Models
Modeling sparse visual state updates instead of full images cuts generated visual tokens ~55.6% and improves interleaved multimodal reasoning over full-image ULMMs.
-
APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts
A VLM planner that adaptively inserts latent visual thoughts of future states into its reasoning trace beats language-only and prior VLM planners on long-horizon kitchen tasks, especially under tight free space.
-
BUS: Brain-Inspired Unsupervised Self-Reflection via Backward Prediction for Multimodal Reasoning
BUS trains VLMs for self-reflection without labels by verifying which sampled reasonings precede their own answers via backward prediction, improving multimodal reasoning.
-
Segmentation before Answering: Pixel Grounding for MLLM Visual Reasoning
SegAnswer trains an MLLM to generate segmentation masks instead of bounding boxes when zooming into image regions during visual reasoning, yielding consistent improvements across perception and hallucination benchmarks.
-
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.
-
Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning
ViToS uses dual-stream RL with cross-feedback optimization to prune medical image tokens to 77% length while reporting 108.27% and 104.16% relative performance on two 7B VLMs across seven benchmarks.
-
VisReflect: Latent Visual Reflection for Fine-Grained Perception in Long Visual Context
VisReflect generates continuous latent visual reflections to emphasize relevant visual features and guide attention in LVLMs, yielding 4.1% gains on image benchmarks and 1.8% on video benchmarks with 44% less inference time than zooming methods.
-
TACO: Tool-Augmented Credit Optimization for Agentic Tool Use
TACO combines Differential Answer-Probe Reward (DAPR) and Outcome-Gated Advantage Routing (OGAR) to assign credit to tool calls in agentic visual reasoning, producing accuracy gains on multimodal benchmarks.
-
Latent Noise Mask for Reducing Visual Redundancy in Multimodal Large Language Models
Lens purifies visual evidence in MLLMs via question-conditioned latent noise masking with a LET token, yielding 2.4-6.4 point gains on VQA and grounding tasks.
-
Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
RA-RFT trains a retriever to rank contexts by expected reasoning benefit and uses the retrieved analogies inside reinforcement fine-tuning, yielding 7.1 and 2.8 point gains on AIME 2025 over GRPO for two Qwen3 models.
-
HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers
HYDRA-X presents the first unified multimodal model using a single ViT for holistic image-video tokenization, with ablations on attention and compression plus a latent-level editing improvement.
-
ROVER: Routing Object-Centric Visual Evidence for Grounded Multi-Image Reasoning
ROVER introduces a learnable routing plugin for object-centric visual evidence in MLLMs via token triplets and differential attention, reporting gains on MM-GCoT and VideoEspresso when integrated into Qwen2.5-VL-7B.
-
InterSketch: An Interleaved Reasoning Model with Self-correcting Visual Sketch and Stepwise Reward
InterSketch improves long-horizon visual-textual chain-of-thought in VLMs by dynamically generating and interleaving self-correcting visual sketches with text, using a synthesized dataset plus reflection in cold-start followed by stepwise-reward RL, and reports outperforming Gemini-3-Pro on benchmar
-
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.
-
Are Tools Always Beneficial? Learning to Invoke Tools Adaptively for Dual-Mode Multimodal LLM Reasoning
AutoTool uses dual-mode RL to let MLLMs adaptively choose tool use or text-only reasoning, reporting 21.8% accuracy gain on V* and 44.9% efficiency gain on POPE versus baselines.
-
Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation
Vision-OPD transfers an MLLM's privileged regional perception to its full-image policy through on-policy token-level self-distillation, yielding competitive results on fine-grained visual benchmarks.
-
Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth
Constraining visual token budgets during SFT and RL forces VLMs to learn functional active perception, yielding ~5% relative gains and strong transfer to unconstrained evaluation.
-
VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation
VideoSeeker integrates agentic reasoning and visual prompts into LVLMs via automated data synthesis, cold-start supervision, and RL training, yielding +13.7% gains on instance-level video tasks over baselines including GPT-4o.
-
PanoWorld: Towards Spatial Supersensing in 360$^\circ$ Panorama World
PanoWorld adds spherical spatial cross-attention and pano-native training data to MLLMs for improved spatial reasoning on ERP panoramas, outperforming baselines on new and existing benchmarks.
-
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.
-
When to Re-Commit: Temporal Abstraction Discovery for Long-Horizon Vision-Language Reasoning
Learns state-conditioned commitment depth in a 7B vision-language policy that jointly predicts actions and replan intervals, outperforming fixed-depth baselines and larger models on Sliding Puzzle and Sokoban while providing a theoretical dominance result.
-
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.
-
Affordance Agent Harness: Verification-Gated Skill Orchestration
Affordance Agent Harness is a verification-gated orchestration system that unifies skills via an evidence store, episodic memory priors, an adaptive router, and a self-consistency verifier to improve accuracy-cost tradeoffs in open-world affordance grounding.
-
See Further, Think Deeper: Advancing VLM's Reasoning Ability with Low-level Visual Cues and Reflection
ForeSight lets VLMs use low-level visual cues and mask-based visual feedback within an RL loop to reason more accurately, with the 7B model beating same-scale peers and some closed-source SOTA on a new benchmark.
-
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.
-
Test-time Scaling over Perception: Resolving the Grounding Paradox in Thinking with Images
TTSP samples and filters multiple zoom-in exploration traces and iteratively consolidates validated observations into an Evidence Ledger, improving fine-grained multimodal reasoning on V* Bench, HR-Bench, TreeBench, and MME-RealWorld-Lite.
-
AnomalyAgent: Agentic Industrial Anomaly Synthesis via Tool-Augmented Reinforcement Learning
AnomalyAgent uses tool-augmented reinforcement learning with self-reflection to generate realistic industrial anomalies, achieving better metrics than zero-shot methods on MVTec-AD.
-
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.
-
AdaTooler-V: Adaptive Tool-Use for Images and Videos
AdaTooler-V trains MLLMs to adaptively use vision tools via AT-GRPO reinforcement learning and new datasets, reaching 89.8% on V* and outperforming GPT-4o.
-
AgentIAD: Agentic Industrial Anomaly Detection via Adaptive Memory Augmentation
AgentIAD introduces an agentic VLM with Perceptive Zoomer, Web Searcher, and Comparative Retriever tools plus two-stage SFT-then-RL training, achieving 5.92% higher classification accuracy than prior SOTA on the MMAD benchmark.
-
LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling
LongVT adds native video-cropping tool calling to LMMs for interleaved multimodal chain-of-tool-thought reasoning on long videos and releases VideoSIAH data for training and evaluation.