| name | agpo-adaptive-group-policy-optimization |
| description | AGPO (Adaptive Group Policy Optimization) methodology — a critic-free refinement of GRPO that uses group-level statistics to adaptively control update magnitude and exploration. Uses a shared probe-derived statistical state to drive adaptive clipping (based on reward dispersion, skewness, probe entropy, policy entropy, KL drift) and bidirectional adaptive temperature sampling. Outperforms PPO/GRPO on 9 math/STEM benchmarks with Qwen2.5-14B. Use when: improving GRPO training stability, reducing hyperparameter tuning burden in LLM RL post-training, adaptive exploration for reasoning models. Activation: AGPO, adaptive group policy optimization, adaptive clipping GRPO, bidirectional adaptive temperature, critic-free RLVR, statistical feedback control. |
AGPO: Adaptive Group Policy Optimization with Dual Statistical Feedback
Paper: arXiv:2605.20722 | Submitted: 20 May 2026
Authors: Miaobo Hu, Shuhao Hu, Bokun Wang, Ruohan Wang, Xin Wang, Xiaobo Guo, Daren Zha, Jun Xiao
Core Problem
PPO/GRPO in LLM reasoning training uses fixed clipping thresholds and fixed decoding temperature, making training brittle and requiring extensive hyperparameter tuning. When reward distributions shift during training (e.g., as the model improves, response quality distribution changes), fixed parameters become suboptimal.
Key Innovations
1. Shared Probe-Driven Statistical State
AGPO uses a shared probe (a lightweight model head) that provides group-level statistical signals:
- Reward dispersion and skewness: How spread out and asymmetric rewards are within a group
- Probe vote entropy: Uncertainty of the probe's evaluations
- Policy entropy: How diverse the model's token probabilities are
- Step-wise KL drift: How much the policy changes per step
These statistics drive two adaptive controllers.
2. Adaptive Clipping
Instead of a fixed ε clip parameter (standard in PPO/GRPO), AGPO sets the trust-region size dynamically:
epsilon_t = g(reward_dispersion, skewness, probe_entropy, policy_entropy, KL_drift)
When uncertainty is high → wider clipping (allow more exploration)
When confidence is high → tighter clipping (conservative learning)
3. Bidirectional Adaptive Temperature Sampling
Instead of a fixed decoding temperature, AGPO uses bidirectional adjustment:
temperature_t = base_temperature + delta * centered_uncertainty
- Heats (increases temperature) when uncertainty is above a running baseline → more exploration
- Cools (decreases temperature) when uncertainty is below baseline → more exploitation
- Centered relative to a running baseline for stability
Experimental Results
| Model | Benchmark | AGPO | PPO | GRPO |
|---|
| Qwen2.5-14B | GSM8K | 67.3% | - | - |
| Qwen2.5-14B | MATH | 40.5% | - | - |
| Llama-3-8B | Math avg | ✓ | - | - |
| Gemma-2-9B | Math avg | ✓ | - | - |
- Gains transfer across multiple backbone architectures
- Ablations confirm both adaptive clipping and adaptive temperature are complementary
- Public implementation available
Relationship to Existing Skills
- [[advantage-collapse-grpo-avspo]] - Addresses GRPO advantage collapse via different mechanism (virtual samples vs statistical adaptation)
- [[gcpo-cooperative-policy-optimization]] - Cooperative GRPO variant addressing exploration collapse
- [[d2evo-dual-difficulty-self-evolution]] - Difficulty-aware sample selection (complementary to AGPO)
- [[delta-discriminative-token-credit-assignment]] - Token-level credit (complementary: AGPO controls group-level dynamics)
- [[gaussian-grpo]] - GRPO improvement via Gaussian modeling
Implementation Notes
- Critic-free: No need for a separate critic network (unlike PPO)
- Shared probe: Lightweight head, minimal overhead
- Running statistics: Maintain running baseline for temperature centering
- Complementary to other GRPO improvements: Can be combined with DelTA, D²Evo, etc.
- Public implementation: see paper for repository link
Use Cases
- LLM Reasoning RL Post-Training: Direct drop-in for GRPO/PPO
- Hyperparameter-Sensitive Training: Reduces tuning burden for clipping and temperature
- Training With Shifting Distributions: Adapts to changing reward landscape as model improves
Activation Keywords
AGPO, adaptive group policy optimization, adaptive clipping GRPO, bidirectional adaptive temperature, critic-free RLVR, statistical feedback control, probe-derived statistics, trust-region adaptation, dual statistical feedback, adaptive exploration LLM RL, Qwen2.5 RL training, group-level statistics, reward dispersion clipping, policy entropy adaptation
