adu2021 Component ID
Long-chain visual reasoning framework with spatial-temporal generalization.
Motivation
Image-centric visual reasoning models like Insight-V lack structured support for sequences, and scaling visual reasoning requires expensive human annotation. A unified framework supporting both images and temporal sequences with self-directed improvement would unlock scalable long-chain reasoning.
The Modification
Unified Multi-Agent Architecture
Replace single-model reasoning with a two-role decomposition: reasoning agent (generates long chains) and summary agent (integrates findings).
# Spatial-temporal reasoning architecture: reason over frames, then summarize
class InsightVPlusPlusPipeline:
"""
Extends image-centric Insight-V into generalized spatial-temporal architecture.
Multi-granularity assessment synthesizes structured reasoning trajectories autonomously.
"""
def __init__(self, image_encoder, reasoning_agent, summary_agent):
self.image_encoder = image_encoder
self.reasoning_agent = reasoning_agent # Handles temporal alignment
self.summary_agent = summary_agent # Integrative evaluation
def forward_long_chain(self, frames_or_image, reasoning_depth=10):
"""
Systematic evolution: image-centric → spatial-temporal sequence reasoning.
Progressive data generation pipeline with multi-granularity assessment.
"""
embeddings = [self.image_encoder(f) for f in frames_or_image]
# Reasoning chain with temporal alignment feedback
reasoning_trajectory = []
for step in range(reasoning_depth):
state = self.reasoning_agent(embeddings, reasoning_trajectory)
reasoning_trajectory.append(state)
# Summary integration via J-GRPO feedback
summary = self.summary_agent(reasoning_trajectory)
return summary, reasoning_trajectory
ST-GRPO and J-GRPO Reward Algorithms
ST-GRPO enforces temporal alignment and complex spatial-temporal logic; J-GRPO fortifies evaluative robustness.
# ST-GRPO: Reinforcement learning for spatial-temporal reasoning
def st_grpo_reward(reasoning_trajectory, ground_truth, frame_sequence):
"""
ST-GRPO forces reasoning agent to master temporal alignment and complex logic.
Rewards correct object tracking across frames and proper temporal causality.
"""
temporal_alignment_score = 0.0
logic_adherence_score = 0.0
for i, state in enumerate(reasoning_trajectory):
# Penalize misalignment with actual frame sequence timing
temporal_alignment_score += alignment_loss(state, frame_sequence[i])
# Reward logical consistency in reasoning chain
logic_adherence_score += consistency_check(state, ground_truth)
return temporal_alignment_score + logic_adherence_score
def j_grpo_reward(summary, reference_summary):
"""
J-GRPO significantly fortifies summary agent's evaluative robustness.
Rewards comprehensive and balanced integration of reasoning trajectory.
"""
completeness = measure_coverage(summary, reference_summary)
balance = measure_balance(summary) # Avoid over-weighting early steps
accuracy = compute_similarity(summary, reference_summary)
return completeness * balance * accuracy
Ablation Results
Performance across image and video reasoning tasks:
Image Reasoning:
- LLaVA-NeXT backbone: +8.1% average improvement on visual reasoning benchmarks
- Advanced reasoning task score: 53.9 on high-complexity image tasks
Video Reasoning:
- +6.9% average improvement across six video benchmarks
- Self-evolving mechanism enables continuous improvement without additional human annotation
Generalization:
- Scalable, progressive data generation pipeline synthesizes structured reasoning trajectories
- Multi-granularity assessment prevents reward collapse on easy examples
Conditions
- Input modality: Images or video frames (tokenizable or embedded)
- Sequence length: Long-horizon reasoning chains (10+ steps demonstrated)
- Backbone: Vision encoders supporting temporal sequence input (e.g., frame stacks)
- Scale: Applied to reasoning and summary agents in reasonable parameter ranges
- Training data: Autonomous synthesis via progressive generation (human annotation not required)
Drop-In Checklist
- Decompose your visual reasoning into reasoning agent + summary agent
- Implement ST-GRPO rewards focusing on temporal alignment and logical consistency
- Implement J-GRPO rewards for summary integration robustness
- Set up autonomous trajectory generation pipeline with multi-granularity assessment
- Validate performance gains on image tasks (+8% target) and video tasks (+7% target)
- Confirm self-evolving loop: no additional human annotations needed after first round
- Profile reasoning chain depth—10+ steps shown effective