メニュー
Group Filtered Policy Optimization (GFPO) reduces inference-time computation by sampling larger groups during training and filtering responses based on length and token efficiency to teach models efficient reasoning.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SOC 職業分類に基づく
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ADu2021/skillXiv --skill sample-more-think-less-gfpoコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
| name | sample-more-think-less-gfpo |
| title | Sample More Think Less: Group Filtered Policy Optimization |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2508.09726 |
| keywords | ["policy-optimization","rl-training","token-efficiency","length-penalization","inference-optimization"] |
| description | Group Filtered Policy Optimization (GFPO) reduces inference-time computation by sampling larger groups during training and filtering responses based on length and token efficiency to teach models efficient reasoning. |
Group Filtered Policy Optimization (GFPO) addresses a critical problem in RL-trained language models: length inflation, where models generate unnecessarily verbose outputs during inference. Rather than optimizing purely for correctness, GFPO trades training-time computation for inference efficiency by sampling multiple candidate responses and filtering them based on efficiency metrics.
The key insight is that larger sample groups at training time enable models to learn more efficient reasoning patterns, reducing computation needed at test time.
Start with a base language model and define your reward function (e.g., correctness on reasoning tasks). Set up PPO or similar RL algorithm infrastructure.
# Hugging Face Transformers + TRL style setup
from trl import PPOTrainer
from transformers import AutoTokenizer, AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b")
def compute_reward(response, reference_answer):
"""Reward function: 1.0 for correct, 0.0 for incorrect"""
is_correct = check_correctness(response, reference_answer)
return 1.0 if is_correct else 0.0
For each training problem, generate multiple candidate responses. Compute both correctness reward and efficiency metrics for each.
# Group sampling during training
def sample_group(prompt, problem_id, group_size=8):
"""Sample multiple responses per problem"""
responses = []
metrics = []
for _ in range(group_size):
response = base_model.generate(
prompt,
max_length=2048,
temperature=0.7,
do_sample=True
)
response_text = tokenizer.decode(response)
# Compute metrics
correctness_reward = compute_reward(response_text, gold_answers[problem_id])
num_tokens = len(response)
efficiency = correctness_reward / (num_tokens + 1) # Reward-per-token
responses.append(response_text)
metrics.append({
'correctness': correctness_reward,
'length': num_tokens,
'efficiency': efficiency,
'response': response_text
})
return responses, metrics
Filter the group using both length and efficiency criteria. Select responses that are both correct and relatively concise.
def filter_responses(metrics, length_threshold=None, efficiency_threshold=0.5):
"""
Filter responses based on two metrics:
- Must exceed efficiency threshold (reward per token)
- Preferentially selects shorter responses among valid ones
"""
# First pass: filter by efficiency (correctness-weighted by length)
efficient_responses = [
m for m in metrics
if m['efficiency'] > efficiency_threshold
]
if not efficient_responses:
# Fallback: take highest efficiency even if below threshold
efficient_responses = sorted(metrics, key=lambda x: x['efficiency'],
reverse=True)[:1]
# Second pass: among efficient ones, prefer shorter ones
if length_threshold is not None:
filtered = [m for m in efficient_responses if m['length'] <= length_threshold]
if filtered:
return filtered
# Return top efficiency responses
return sorted(efficient_responses, key=lambda x: x['efficiency'],
reverse=True)[:2]
Integrate length awareness into the RL loss. Add a penalty term that discourages verbose responses.
def compute_training_loss(response, reward, num_tokens, length_weight=0.1):
"""
Combined RL objective with efficiency component:
L = -reward + length_weight * log(num_tokens)
"""
base_reward = reward
length_penalty = length_weight * torch.log(torch.tensor(num_tokens + 1))
# Negative because we're optimizing (policy gradient uses -log_prob * advantage)
advantage = base_reward - length_penalty
return advantage
Use the filtered, efficient responses to update the policy. Only train on responses that pass the efficiency filter.
# PPO training loop with GFPO filtering
for batch_idx, problem_batch in enumerate(training_problems):
group_responses = []
group_metrics = []
for problem_id, prompt in problem_batch:
responses, metrics = sample_group(prompt, problem_id, group_size=8)
filtered = filter_responses(metrics, efficiency_threshold=0.6)
group_responses.extend(filtered)
group_metrics.extend([m['response'] for m in filtered])
# Train PPO on filtered responses
ppo_trainer.train({
'prompts': [p for p, _ in problem_batch],
'responses': group_responses,
'rewards': [m['correctness'] for m in group_metrics]
})
Dynamically increase sampling for harder problems where the efficiency filter rejects more candidates.
def adaptive_group_size(problem_id, difficulty_scores):
"""
Allocate more samples to harder problems
difficulty_scores[problem_id] should be between 0 and 1
"""
base_group_size = 8
difficulty = difficulty_scores.get(problem_id, 0.5)
# Harder problems get larger groups
adaptive_size = int(base_group_size * (1 + difficulty))
return min(adaptive_size, 32) # Cap at 32
# In training loop:
for problem_id, prompt in problem_batch:
group_size = adaptive_group_size(problem_id, problem_difficulties)
responses, metrics = sample_group(prompt, problem_id, group_size=group_size)
filtered = filter_responses(metrics)
Sample More Think Less (2508.09726): https://arxiv.org/abs/2508.09726
GFPO reduces inference-time token consumption by 46-71% on reasoning tasks while maintaining accuracy by training on length-filtered, efficiency-optimized responses sampled in larger groups.