| name | native-active-perception-reasoning |
| description | Native active perception methodology for omni-modal understanding using POMDP-based Observation-Thought-Action cycle, Agentic Supervised Fine-Tuning (ASFT), and TAURA turn-level credit assignment. |
| version | 1.0.0 |
| tags | ["active-perception","reasoning","POMDP","agentic-rl","video-understanding","multimodal"] |
| activation_keywords | ["active perception","POMDP","observation thought action","ASFT","TAURA","video understanding","turn-level entropy","test-time scaling","on-demand perception"] |
Native Active Perception as Reasoning
Overview
OmniAgent framework for native omni-modal understanding that formulates video understanding as a POMDP-based iterative Observation-Thought-Action cycle. Decouples reasoning complexity from raw video duration through on-demand perception.
Core Methodology
1. POMDP Formulation
- State: Persistent textual memory (distilled audio-visual cues)
- Action: On-demand observation actions (seek, query, focus)
- Observation: Selective audio-visual cue extraction
- Reward: Query-answering accuracy + efficiency bonus
2. Observation-Thought-Action Cycle
while not terminated:
# Thought: Analyze current memory state
thought = analyze_memory(current_state, query)
# Action: Select observation strategy
action = select_action(thought, uncertainty)
# Observation: Execute on-demand perception
new_cue = execute_observation(action)
# Memory Update: Distill into persistent text
memory = update_memory(memory, new_cue)
3. Agentic Supervised Fine-Tuning (ASFT)
- Best-of-N trajectory synthesis
- Dual-stage quality control:
- Outcome verification (answer correctness)
- Process verification (trajectory efficiency)
- Bootstrap native active perception from expert demonstrations
4. TAURA: Turn-aware Adaptive Uncertainty Rescaled Advantage
# Turn-level entropy for credit assignment
for each turn t:
entropy_t = compute_entropy(policy_t)
# Rescale advantage based on uncertainty
advantage_t = (reward_t - baseline) * rescale_factor(entropy_t)
# Pivotal discovery turns get higher credit
if is_pivotal_turn(entropy_t):
advantage_t *= discovery_boost
Key Insight: High-entropy turns indicate exploration → boost credit assignment
5. Test-Time Scaling
- Positive scaling: Performance improves with more reasoning turns
- Decoupling: Complexity independent of video duration
- On-demand: Only observe what's needed for the query
Implementation Pattern
class OmniAgent:
def __init__(self, vision_encoder, text_memory, policy_model):
self.memory = PersistentTextMemory()
self.policy = POMDPPolicy()
def understand(self, video, query, max_turns=10):
for turn in range(max_turns):
thought = self.policy.think(self.memory.state, query)
if thought.confidence > threshold:
return self.answer(query, self.memory)
action = self.policy.act(thought)
cue = self.observe(video, action)
self.memory.distill(cue)
Key Benefits
- Efficiency: Memory cost decoupled from video duration
- Test-Time Scaling: More turns → better performance
- Native: No pre-scanning required
- Adaptive: Entropy-guided credit assignment
Use Cases
- Long video understanding (>10min)
- Omni-modal QA (audio + visual + text)
- Active perception agents
- On-demand inference optimization
Performance
- 7B model beats 72B baseline (50.5% vs 47.3% on LVBench)
- VideoMME, LVBench state-of-the-art among open-source
Reference
- Paper: "Native Active Perception as Reasoning for Omni-Modal Understanding" (arXiv:2606.19341v1)
- Authors: Zhenghao Xing et al. (2026-06-17)