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.
hub Canonical reference
LLaVA-CoT: Let Vision Language Models Reason Step-by-Step
Canonical reference. 90% of citing Pith papers cite this work as background.
48
Pith papers citing it
Background
90% of classified citations
abstract
Large language models have demonstrated substantial advancements in reasoning capabilities. However, current Vision-Language Models (VLMs) often struggle to perform systematic and structured reasoning, especially when handling complex visual question-answering tasks. In this work, we introduce LLaVA-CoT, a large VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-CoT independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation. This structured approach enables LLaVA-CoT to achieve marked improvements on reasoning-intensive tasks. To accomplish this, we construct the LLaVA-CoT-100k dataset, integrating samples from various visual question answering sources and providing structured reasoning annotations. Besides, we propose a test-time stage-wise retracing search method (SWIRES), which enables effective and efficient test-time scaling. Remarkably, with only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks, but also surpasses the performance of larger and even closed-source models, such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct. The code, dataset, and pre-trained weights are publicly available at https://github.com/PKU-YuanGroup/LLaVA-CoT.
hub tools
citation-role summary
background 9
baseline 1
citation-polarity summary
representative citing papers
Attentive-CoT is an attention-guided fine-tuning objective that improves chain-of-thought performance in multimodal LLMs by delaying answer commitment and increasing sustained visual-token access during rationale generation.
Test-Time Hinting trains a hint generator to prepend contextual guidance to VLM prompts, improving accuracy on natural-image VQA benchmarks with generalization to unseen tasks and models.
UIPress is the first encoder-side learned optical compression method for UI-to-Code that compresses visual tokens to 256, outperforming the uncompressed baseline by 7.5% CLIP score and the best inference-time baseline by 4.6% while delivering 9.1x TTFT speedup.
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than baselines on reasoning benchmarks.
Latent Visual Reasoning enables autoregressive generation of latent visual states that reconstruct critical image tokens, yielding gains on perception-heavy VQA benchmarks such as 71.67% on MMVP.
Audio Flamingo 3 introduces an open large audio-language model achieving new state-of-the-art results on over 20 audio understanding and reasoning benchmarks using a unified encoder and curriculum training on open data.
MGPO elicits grounding in LMMs via multi-turn RL with binary rewards, yielding 5.4% and 5.2% gains on MME-Realworld and V* Bench and surpassing GPT-4o on the latter after training on 21K samples.
VGR introduces a visual-grounded reasoning MLLM that detects and replays image regions during inference, achieving gains on visual benchmarks with 30% fewer image tokens than the LLaVA-NeXT-7B baseline.
Video-Holmes benchmark shows top MLLMs achieve at most 45% accuracy on tasks needing integration of multiple clues from suspense films, unlike existing perception-focused tests.
FakeReasoning is an MLLM-based framework for unified forgery detection and reasoning on AI-generated images, supported by the new MMFR-Dataset of 120K images and 378K annotations across 10 generators.
R1-VL uses StepGRPO with rule-based StepRAR and StepRVR rewards to let MLLMs learn step-by-step reasoning beyond imitation of positive paths.
RL on Qwen2-VL-2B with SAT dataset produces R1-like reasoning and 59.47% CVBench accuracy, outperforming base model by ~30% and SFT by ~2%.
CAVE is a GRPO-based process-reward method that improves VLMs on fragmented visual reasoning by crediting intermediate actions via belief update, evidence acquisition, and adaptive focus, shown on TRACER-Bench and public benchmarks.
RadThinking releases a large longitudinal CT VQA dataset stratified into foundation perception questions, single-rule reasoning questions, and compositional multi-step chains grounded in clinical reporting standards for cancer screening.
OceanPile is a new multimodal corpus with unified data collection, instruction tuning set, and benchmark to train foundation models for ocean science.
Video-ToC adds tree-guided cue localization, demand-based RL rewards, and automated datasets to video LLMs, reporting better results than prior methods on six understanding benchmarks plus a hallucination test.
Saliency-R1 uses a novel saliency map technique and GRPO with human bounding-box overlap as reward to improve VLM reasoning faithfulness and interpretability.
CoME-VL fuses contrastive and self-supervised vision encoders via entropy-guided multi-layer aggregation and RoPE cross-attention to improve vision-language model performance on benchmarks.
Chat-Scene++ improves 3D scene understanding in multimodal LLMs by representing scenes as context-rich object sequences with identifier tokens and grounded chain-of-thought reasoning, reaching state-of-the-art on five benchmarks using pre-trained encoders.
Fine-R1 uses chain-of-thought supervised fine-tuning on a structured FGVR reasoning dataset plus triplet augmented policy optimization to outperform general MLLMs and CLIP models on seen and unseen fine-grained categories with 4-shot training.
DRP decouples reasoning from perception in LMMs by using an LLM reasoner to query an LMM observer for visual details as needed, reducing visual grounding loss.
LaV-CoT introduces a multi-stage visual CoT pipeline and GRPO training with language-consistency rewards, delivering up to 9.5% accuracy gains on multilingual VQA benchmarks over similar-sized open models.
Survey that defines agentic RL for LLMs via POMDPs, introduces a taxonomy of planning/tool-use/memory/reasoning capabilities and domains, and compiles open environments from over 500 papers.
citing papers explorer
-
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.
-
Attention-guided Fine-tuning of Multimodal Large Language Models Improves Chain-of-Thought Reasoning
Attentive-CoT is an attention-guided fine-tuning objective that improves chain-of-thought performance in multimodal LLMs by delaying answer commitment and increasing sustained visual-token access during rationale generation.
-
Test-Time Hinting for Black-Box Vision-Language Models
Test-Time Hinting trains a hint generator to prepend contextual guidance to VLM prompts, improving accuracy on natural-image VQA benchmarks with generalization to unseen tasks and models.
-
UIPress: Bringing Optical Token Compression to UI-to-Code Generation
UIPress is the first encoder-side learned optical compression method for UI-to-Code that compresses visual tokens to 256, outperforming the uncompressed baseline by 7.5% CLIP score and the best inference-time baseline by 4.6% while delivering 9.1x TTFT speedup.
-
Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than baselines on reasoning benchmarks.
-
Latent Visual Reasoning
Latent Visual Reasoning enables autoregressive generation of latent visual states that reconstruct critical image tokens, yielding gains on perception-heavy VQA benchmarks such as 71.67% on MMVP.
-
Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models
Audio Flamingo 3 introduces an open large audio-language model achieving new state-of-the-art results on over 20 audio understanding and reasoning benchmarks using a unified encoder and curriculum training on open data.
-
High-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning
MGPO elicits grounding in LMMs via multi-turn RL with binary rewards, yielding 5.4% and 5.2% gains on MME-Realworld and V* Bench and surpassing GPT-4o on the latter after training on 21K samples.
-
VGR: Visual Grounded Reasoning
VGR introduces a visual-grounded reasoning MLLM that detects and replays image regions during inference, achieving gains on visual benchmarks with 30% fewer image tokens than the LLaVA-NeXT-7B baseline.
-
Video-Holmes: Can MLLM Think Like Holmes for Complex Video Reasoning?
Video-Holmes benchmark shows top MLLMs achieve at most 45% accuracy on tasks needing integration of multiple clues from suspense films, unlike existing perception-focused tests.
-
Toward Generalizable Forgery Detection and Reasoning
FakeReasoning is an MLLM-based framework for unified forgery detection and reasoning on AI-generated images, supported by the new MMFR-Dataset of 120K images and 378K annotations across 10 generators.
-
R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization
R1-VL uses StepGRPO with rule-based StepRAR and StepRVR rewards to let MLLMs learn step-by-step reasoning beyond imitation of positive paths.
-
R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Model
RL on Qwen2-VL-2B with SAT dataset produces R1-like reasoning and 59.47% CVBench accuracy, outperforming base model by ~30% and SFT by ~2%.
-
CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning
CAVE is a GRPO-based process-reward method that improves VLMs on fragmented visual reasoning by crediting intermediate actions via belief update, evidence acquisition, and adaptive focus, shown on TRACER-Bench and public benchmarks.
-
RadThinking: A Dataset for Longitudinal Clinical Reasoning in Radiology
RadThinking releases a large longitudinal CT VQA dataset stratified into foundation perception questions, single-rule reasoning questions, and compositional multi-step chains grounded in clinical reporting standards for cancer screening.
-
OceanPile: A Large-Scale Multimodal Ocean Corpus for Foundation Models
OceanPile is a new multimodal corpus with unified data collection, instruction tuning set, and benchmark to train foundation models for ocean science.
-
Video-ToC: Video Tree-of-Cue Reasoning
Video-ToC adds tree-guided cue localization, demand-based RL rewards, and automated datasets to video LLMs, reporting better results than prior methods on six understanding benchmarks plus a hallucination test.
-
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.
-
CoME-VL: Scaling Complementary Multi-Encoder Vision-Language Learning
CoME-VL fuses contrastive and self-supervised vision encoders via entropy-guided multi-layer aggregation and RoPE cross-attention to improve vision-language model performance on benchmarks.
-
Chat-Scene++: Exploiting Context-Rich Object Identification for 3D LLM
Chat-Scene++ improves 3D scene understanding in multimodal LLMs by representing scenes as context-rich object sequences with identifier tokens and grounded chain-of-thought reasoning, reaching state-of-the-art on five benchmarks using pre-trained encoders.
-
Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning
Fine-R1 uses chain-of-thought supervised fine-tuning on a structured FGVR reasoning dataset plus triplet augmented policy optimization to outperform general MLLMs and CLIP models on seen and unseen fine-grained categories with 4-shot training.
-
Mitigating Visual Context Degradation in Large Multimodal Models: A Training-Free Decoupled Agentic Framework
DRP decouples reasoning from perception in LMMs by using an LLM reasoner to query an LMM observer for visual details as needed, reducing visual grounding loss.
-
LaV-CoT: Language-Aware Visual CoT with Multi-Aspect Reward Optimization for Real-World Multilingual VQA
LaV-CoT introduces a multi-stage visual CoT pipeline and GRPO training with language-consistency rewards, delivering up to 9.5% accuracy gains on multilingual VQA benchmarks over similar-sized open models.
-
The Landscape of Agentic Reinforcement Learning for LLMs: A Survey
Survey that defines agentic RL for LLMs via POMDPs, introduces a taxonomy of planning/tool-use/memory/reasoning capabilities and domains, and compiles open environments from over 500 papers.
-
Listener-Rewarded Thinking in VLMs for Image Preferences
Listener-augmented GRPO uses an independent frozen VLM to provide dense confidence scores on reasoning traces, yielding 67.4% accuracy on ImageReward, up to +6% OOD gains on 1.2M-vote human data, and fewer reasoning contradictions.
-
Grounded Reinforcement Learning for Visual Reasoning
ViGoRL introduces visually grounded RL that anchors reasoning steps to image coordinates and uses multi-turn zooming to outperform standard RL and supervised baselines on spatial and GUI reasoning benchmarks.
-
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
Iterative SFT-RL cycles enable a 7B LVLM to develop sophisticated visual chain-of-thought reasoning and improve performance on math and general reasoning benchmarks.
-
MathFlow: Enhancing the Perceptual Flow of MLLMs for Visual Mathematical Problems
MathFlow decouples perception and inference stages in MLLMs for visual math, with a dedicated perception model delivering gains on the FlowVerse benchmark when paired with existing reasoners.
-
Search-o1: Agentic Search-Enhanced Large Reasoning Models
Search-o1 integrates agentic retrieval-augmented generation and a Reason-in-Documents module into large reasoning models to dynamically supply missing knowledge and improve performance on complex science, math, coding, and QA tasks.
-
HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs
HuatuoGPT-o1 achieves superior medical complex reasoning by using a verifier to curate reasoning trajectories for fine-tuning and then applying RL with verifier-based rewards.
-
LASER: A Corrective Lens for LVLMs via Visual Attention Preservation and Sink Suppression
LASER uses Visual Grounding Reward and Sink Suppression Reward to preserve visual attention trajectories and suppress sink tokens, reducing visual forgetting in LVLMs.
-
Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety
Yuvion VL is a multimodal LLM family using adversarial-aware data construction, three-stage training, and contrastive fine-tuning that claims industry-leading safety performance on new benchmarks while retaining general capabilities.
-
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.
-
RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning
POPO uses recency-based prioritized group replay and decoupled off-policy optimization to avoid zero-variance ineffective samples in RLVR, accelerating LLM reasoning finetuning with fewer rollouts.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
D$^2$Evo: Dual Difficulty-Aware Self-Evolution for Data-Efficient Reinforcement Learning
D²Evo mines medium-difficulty anchors from the current model, trains a Questioner to generate matching questions, and jointly optimizes Solver and Questioner for progressive gains, outperforming baselines on math reasoning with under 2K real samples.
-
APCD: Adaptive Path-Contrastive Decoding for Reliable Large Language Model Generation
APCD adaptively branches LLM decoding paths based on token entropy and contrasts divergent paths to improve factual accuracy while preserving efficiency.
-
ARGUS: Policy-Adaptive Ad Governance via Evolving Reinforcement with Adversarial Umpiring
ARGUS uses a Prosecutor-Defender-Umpire multi-agent setup plus RAG and chain-of-thought rewards to adapt ad policy enforcement to new regulations using minimal fresh labels.
-
Grounding Everything in Tokens for Multimodal Large Language Models
GETok partitions images with grid tokens and refines locations via offset tokens to enable better native 2D spatial reasoning in MLLMs.
-
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.
-
Efficient Reasoning with Hidden Thinking
Heima compresses verbose CoT into hidden thinking tokens via information-theoretic analysis and an adaptive interpreter, claiming maintained or improved zero-shot accuracy on reasoning benchmarks.
-
ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP
ReasonCLIP-58M applies continual pretraining with visually grounded reasoning captions on 58M examples to improve CLIP-style models on commonsense and compositional reasoning tasks.
-
One Image is All You Need: Agentic One-Shot Image Generation via Text-Based World Models for Long-Tail Spatial Perception
WMGen-v1 generates diverse long-tail spatial images from one reference image via LVLM scene parsing, LLM-guided expansion, and diffusion synthesis, with detectors trained only on the synthetic data approaching real-data performance on benchmarks.
-
Improving the Reasoning of Multi-Image Grounding in MLLMs via Reinforcement Learning
A pipeline of chain-of-thought data synthesis, LoRA-based supervised fine-tuning, rejection sampling, and rule-based reinforcement learning raises multi-image grounding accuracy by 9.04% on MIG-Bench and 4.41% on average across seven other benchmarks.
-
Vision-EKIPL: External Knowledge-Infused Policy Learning for Visual Reasoning
Vision-EKIPL injects high-quality actions from external models into RL training to expand exploration and raise the reasoning ceiling of MLLMs, reporting up to 5% gains on the Reason-RFT-CoT benchmark.
-
R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization
R1-Onevision turns images into structured text for multimodal reasoning, trains on a custom dataset with RL, and claims SOTA results on an educational benchmark.
-
From System 1 to System 2: A Survey of Reasoning Large Language Models
The survey organizes the shift of LLMs toward deliberate System 2 reasoning, covering model construction techniques, performance on math and coding benchmarks, and future research directions.
-
Multimodal Chain-of-Thought Reasoning: A Comprehensive Survey
The paper provides the first comprehensive survey of multimodal chain-of-thought reasoning, including foundational concepts, a taxonomy of methodologies, application analyses, challenges, and future directions.