Add an intermediate RL stage between pretraining and post-training using dynamic token budgeting, curriculum sampling, and dual training. Trigger: reduce reasoning steps while maintaining or improving performance in post-training.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Add an intermediate RL stage between pretraining and post-training using dynamic token budgeting, curriculum sampling, and dual training. Trigger: reduce reasoning steps while maintaining or improving performance in post-training.
Reinforcement Mid-Training (RMT): Three-Stage LLM Development
Core Concept
Standard LLM development has two stages: pretraining (next-token prediction) and post-training (instruction-tuning + RL). RMT inserts a critical middle stage that applies reinforcement learning to discourage unnecessary reasoning steps while focusing training on high-value tokens. This achieves up to 64.91% performance improvement while using only 21% of reasoning tokens.
The key insight: Not all tokens contribute equally to performance. Mid-training learns to allocate compute efficiently before the model develops bad reasoning habits during post-training.
Architecture Overview
Dynamic Token Budgeting: Learns when to stop generating reasoning and produce answers
Curriculum-Based Sampling: Progressive training from simple to complex tokens
Dual Training Strategy: Combines RL with next-token prediction for token importance weighting
Create a mechanism to reward efficient reasoning and penalize excessive token generation.
classTokenBudgetingReward:
def__init__(self, max_tokens=4096, target_ratio=0.5):
self.max_tokens = max_tokens
self.target_ratio = target_ratio # Ideally use 50% of budgetdefcompute_reward(self, thinking_tokens, final_answer_tokens, is_correct):
"""
Reward correctness, but penalize verbose thinking.
Args:
thinking_tokens: Number of reasoning tokens generated
final_answer_tokens: Number of tokens in final answer
is_correct: Boolean whether answer is correct
Returns:
Scalar reward signal
"""# Base correctness reward
correctness_reward = 1.0if is_correct else -1.0# Efficiency bonus: reward using <50% of budget
efficiency_ratio = thinking_tokens / self.max_tokens
efficiency_bonus = 0.5if efficiency_ratio < self.target_ratio else -0.2# Penalize excessive reasoningif thinking_tokens > self.max_tokens:
excessive_penalty = -0.5else:
excessive_penalty = 0# Combined reward
total_reward = (
correctness_reward +
0.3 * efficiency_bonus +
excessive_penalty
)
return total_reward
defrecord_trajectory(self, problem, thinking, answer, is_correct):
"""Log a trajectory with token counts and correctness."""
thinking_tokens = len(thinking.split())
answer_tokens = len(answer.split())
reward = self.compute_reward(
thinking_tokens,
answer_tokens,
is_correct
)
return {
"problem": problem,
"thinking": thinking,
"answer": answer,
"thinking_tokens": thinking_tokens,
"answer_tokens": answer_tokens,
"is_correct": is_correct,
"reward": reward
}
3. Implement Curriculum-Based Sampling
Start with easy problems and progressively increase difficulty. This helps the model learn efficient reasoning on simpler tasks before tackling complex ones.
classCurriculumSampler:
def__init__(self, dataset, config):
self.dataset = dataset
self.config = config
self.current_stage = 0self.stage_step = 0defsample_batch(self, batch_size):
"""
Sample a batch weighted by curriculum.
"""# Determine current difficulty stage based on training progress
stage_config = self.config.difficulty_stages[self.current_stage]
# Sample from current stage with high probability# Sample from adjacent stages with lower probability
difficulties = ["easy", "medium", "hard"]
ifself.current_stage == 0:
difficulty_weights = [0.8, 0.2, 0.0]
elifself.current_stage == 1:
difficulty_weights = [0.2, 0.6, 0.2]
else:
difficulty_weights = [0.0, 0.3, 0.7]
# Weighted sampling
batch = []
for _ inrange(batch_size):
difficulty = np.random.choice(
difficulties,
p=difficulty_weights
)
example = self.dataset.sample_by_difficulty(difficulty)
batch.append(example)
# Progress through stagesself.stage_step += 1ifself.stage_step > stage_config.get("duration_steps", 30000):
self.current_stage = min(self.current_stage + 1, 2)
self.stage_step = 0return batch
defget_curriculum_weight(self):
"""Return curriculum weight for loss computation."""returnself.config.difficulty_stages[self.current_stage]["curriculum_weight"]
4. Implement Dual Training (RL + Supervised)
Combine RL losses (based on correctness) with supervised losses (predicting important tokens).
defcompute_mid_training_loss(
model,
batch,
token_budgeter,
curriculum_weight,
rl_weight=0.7,
supervised_weight=0.3):
"""
Compute combined RL and supervised loss for mid-training.
Args:
model: LLM to train
batch: List of training examples
token_budgeter: Reward computer
curriculum_weight: Importance weight from curriculum
Returns:
Scalar loss value
"""
total_loss = 0
batch_size = len(batch)
for example in batch:
# Generate thinking and answer
output = model.generate(
example["problem"],
max_thinking_tokens=4096,
return_intermediate=True
)
thinking = output["thinking"]
answer = output["answer"]
# Evaluate correctness
is_correct = evaluate_answer(answer, example["ground_truth"])
# Compute reward (RL component)
reward = token_budgeter.compute_reward(
len(thinking.split()),
len(answer.split()),
is_correct
)
log_prob = model.compute_log_prob(thinking + answer)
# RL loss: policy gradient
rl_loss = -reward * log_prob
# Supervised loss: predict important tokens# Identify "important" tokens (problem-solving steps, key insights)
important_tokens = identify_important_tokens(
thinking,
example["solution_tokens"]
)
supervised_loss = compute_token_importance_loss(
model,
thinking,
important_tokens
)
# Combine losses with curriculum weight
combined_loss = (
rl_weight * rl_loss +
supervised_weight * supervised_loss
) * curriculum_weight
total_loss += combined_loss
return total_loss / batch_size
