Train LLMs with distribution-matching rewards instead of reward maximization to achieve 10% improvement on math reasoning while improving solution diversity by matching the full reward distribution via flow balancing, addressing mode collapse in long chain-of-thought tasks.
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.
Instruções da origem · Visualização somente leitura
name
flowrl-reward-distribution-matching
title
FlowRL: Matching Reward Distributions for LLM Reasoning
version
0.0.2
engine
skillxiv-v0.0.2-claude-opus-4.6
license
MIT
url
https://arxiv.org/abs/2509.15207
keywords
["LLM reasoning","reward distribution matching","flow-balanced optimization","policy learning","chain-of-thought"]
description
Train LLMs with distribution-matching rewards instead of reward maximization to achieve 10% improvement on math reasoning while improving solution diversity by matching the full reward distribution via flow balancing, addressing mode collapse in long chain-of-thought tasks.
Flow-Balanced Reward Distribution Matching: Improving LLM Reasoning Diversity
Traditional reinforcement learning approaches for large language models maximize expected rewards, which causes premature convergence to a single dominant solution mode. This severely limits diversity in reasoning paths, especially in long chain-of-thought (CoT) tasks where multiple valid solutions exist. FlowRL solves this by treating the problem as distributional learning—instead of maximizing rewards, it matches the full distribution of rewards across all generated trajectories using flow-based optimization.
This approach is grounded in GFlowNets theory, which proves that matching reward distributions via flow balance mathematically encourages exploration and diverse solution discovery. By normalizing rewards through a learnable partition function and minimizing KL divergence between the policy and target distribution, FlowRL recovers diverse reasoning strategies while maintaining or improving task performance.
Core Concept
FlowRL replaces the typical RL objective of maximizing expected reward with a distribution-matching objective: instead of pushing toward high-reward trajectories, it shapes the entire probability distribution to follow a target reward distribution. This subtle shift has profound effects:
Mode coverage: The policy explores multiple solution paths proportional to their rewards, preventing collapse to a single strategy
Length stability: By normalizing gradients by sequence length, long CoT reasoning doesn't dominate training signals
Trajectory balance: The flow-balanced formulation ensures information conservation across the generation process
The key insight: maximization creates peaks; distribution matching creates landscapes. An LLM that understands multiple reasoning strategies is more robust and generalizable than one trained solely on the highest-reward trajectory.
Architecture Overview
Partition function: A 3-layer MLP that normalizes scalar rewards into a probability distribution, taking hidden state features as input
Policy network: The base LLM, updated to match the target reward distribution instead of maximizing rewards
Loss function: Squared form of flow-balance objective, incorporating length normalization and importance sampling corrections
Reference policy: Maintained for stability during off-policy updates, preventing excessive deviation from the foundation model
Importance sampling: Applies PPO-style clipping to handle distribution mismatch between trajectory samples and current policy
Implementation
Step 1: Set Up Reward Normalization with Partition Function
The partition function Z_φ(x) converts scalar rewards into a proper probability distribution. This learnable component receives contextual features and outputs normalized reward weights. The function is crucial because raw rewards have different scales and ranges depending on the task.
import torch
import torch.nn as nn
classPartitionFunction(nn.Module):
"""Learnable partition function that normalizes rewards into a distribution."""def__init__(self, hidden_dim: int, output_dim: int = 1):
super().__init__()
self.net = nn.Sequential(
nn.Linear(hidden_dim, 256),
nn.ReLU(),
nn.Linear(256, 128),
nn.ReLU(),
nn.Linear(128, output_dim)
)
defforward(self, hidden_states):
"""
Args:
hidden_states: (batch_size, hidden_dim) contextual features
Returns:
log_Z: (batch_size, 1) log partition values
"""
log_Z = self.net(hidden_states)
return log_Z
The partition function takes the mean of the LLM's hidden states and produces a scalar normalization factor. During training, its parameters adjust to balance reward distributions across the batch.
Step 2: Compute Flow-Balanced Loss with Length Normalization
The core FlowRL loss matches distributions by minimizing KL divergence with length-normalized gradients. Length normalization prevents long sequences from dominating gradients—critical for stable training on variable-length CoT outputs.
defcompute_flowrl_loss(
log_Z: torch.Tensor, # (batch_size,)
log_probs: torch.Tensor, # (batch_size,) log π_θ(y|x)
rewards: torch.Tensor, # (batch_size,)
sequence_lengths: torch.Tensor, # (batch_size,)
log_ref_probs: torch.Tensor, # (batch_size,) log π_ref(y|x)
beta: float = 1.0,
eps: float = 1e-7) -> torch.Tensor:
"""
Compute the flow-balanced loss.
Args:
log_Z: Log partition function values from the partition network
log_probs: Log probabilities of generated trajectories under current policy
rewards: Task rewards (0/1 for correctness or continuous scores)
sequence_lengths: Length of each generated sequence
log_ref_probs: Reference policy log probabilities
beta: Temperature parameter controlling reward emphasis
eps: Small epsilon for numerical stability
Returns:
loss: Scalar tensor representing flow-balanced objective
"""# Normalize by sequence length to prevent long sequences from dominating
length_factor = 1.0 / (sequence_lengths.float() + eps)
# Main flow-balance equation:# log Z_φ(x) + (1/|y|) log π_θ(y|x) = β r̂(x,y) + (1/|y|) log π_ref(y|x)
lhs = log_Z + length_factor * log_probs
rhs = beta * rewards + length_factor * log_ref_probs
# Squared loss for distribution matching
flow_loss = torch.mean((lhs - rhs) ** 2)
return flow_loss
The squared formulation of the flow-balance equation enables distribution matching. The length_factor term is critical: it divides the log-probability contribution by sequence length, ensuring that generating a 100-token vs. 10-token response doesn't create 10x gradient magnitude differences.
Step 3: Integrate Importance Sampling for Off-Policy Updates
Since trajectories come from the previous policy iteration, FlowRL applies PPO-style importance sampling to prevent off-policy divergence. This stabilizes training when the current policy drifts from the data collection policy.
defapply_importance_sampling_correction(
log_probs: torch.Tensor, # (batch_size,)
log_old_probs: torch.Tensor, # (batch_size,) from trajectory collection
flow_loss: torch.Tensor, # computed flow loss
clip_ratio: float = 0.2,
eps: float = 1e-7) -> torch.Tensor:
"""
Apply PPO-style importance sampling to correct for off-policy data.
Args:
log_probs: Current policy log probabilities
log_old_probs: Log probabilities when trajectories were sampled
flow_loss: Base flow-balanced loss
clip_ratio: Clipping threshold for importance weights
eps: Numerical stability epsilon
Returns:
corrected_loss: Flow loss adjusted for policy divergence
"""# Importance weight: π_θ / π_old
log_importance = log_probs - log_old_probs
importance_weights = torch.exp(torch.clamp(log_importance, min=-5.0, max=5.0))
# Clip importance weights to [1-clip, 1+clip]
clipped_weights = torch.clamp(
importance_weights,
min=1.0 - clip_ratio,
max=1.0 + clip_ratio
)
# Reweight loss by clipped importance weights
corrected_loss = torch.mean(clipped_weights * flow_loss)
return corrected_loss
Importance sampling correction prevents the loss from exploding when the current policy becomes too different from the data collection policy. The clipping mechanism bounds the maximum influence of any single example, similar to PPO's clip function.
Step 4: Training Loop Integration
Wire FlowRL into your LLM training pipeline. This shows how to orchestrate trajectory sampling, reward computation, and the three loss components.
deftrain_flowrl_step(
model: nn.Module, # LLM to train
partition_fn: PartitionFunction, # Reward normalizer
batch: dict, # Contains: input_ids, labels, sequence_lengths
reward_model: callable, # Function that scores (input, output) -> reward
reference_model: nn.Module, # Base model (frozen)
optimizer: torch.optim.Optimizer,
beta: float = 1.0,
clip_ratio: float = 0.2) -> float:
"""
Single FlowRL training step.
Args:
model: LLM being trained
partition_fn: Partition function network
batch: Training data batch
reward_model: Evaluates correctness of completions
reference_model: Frozen reference policy
optimizer: Optimizer for both model and partition_fn
beta: Reward emphasis temperature
clip_ratio: Importance sampling clipping threshold
Returns:
loss_value: Scalar loss for monitoring
"""# 1. Forward pass through LLM
logits = model(
input_ids=batch['input_ids'],
attention_mask=batch['attention_mask']
).logits
# 2. Get log probabilities and hidden states
log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
batch_log_probs = extract_trajectory_log_probs(
log_probs,
batch['labels']
) # (batch_size,)
hidden_states = model.get_hidden_states() # Access final hidden layer
hidden_mean = hidden_states.mean(dim=1) # (batch_size, hidden_dim)# 3. Compute partition function
log_Z = partition_fn(hidden_mean).squeeze(-1) # (batch_size,)# 4. Compute rewards from reward model
rewards = batch.get('rewards', torch.zeros(batch['input_ids'].size(0)))
# 5. Get reference policy log probswith torch.no_grad():
ref_logits = reference_model(
input_ids=batch['input_ids'],
attention_mask=batch['attention_mask']
).logits
ref_log_probs = torch.nn.functional.log_softmax(ref_logits, dim=-1)
batch_log_ref_probs = extract_trajectory_log_probs(
ref_log_probs,
batch['labels']
) # (batch_size,)# 6. Compute flow-balanced loss with length normalization
flow_loss = compute_flowrl_loss(
log_Z=log_Z,
log_probs=batch_log_probs,
rewards=rewards,
sequence_lengths=batch['sequence_lengths'],
log_ref_probs=batch_log_ref_probs,
beta=beta
)
# 7. Apply importance sampling correctionwith torch.no_grad():
old_log_probs = batch.get('old_log_probs', batch_log_probs.detach())
corrected_loss = apply_importance_sampling_correction(
log_probs=batch_log_probs,
log_old_probs=old_log_probs,
flow_loss=flow_loss,
clip_ratio=clip_ratio
)
# 8. Backward pass and optimization
optimizer.zero_grad()
corrected_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
optimizer.step()
return corrected_loss.item()
defextract_trajectory_log_probs(log_probs, labels):
"""Helper: extract log probability of the chosen label at each position."""# Shift: predict token i+1 from token i
shift_logits = log_probs[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Gather log probs for selected labels
log_probs_gathered = torch.gather(shift_logits, -1, shift_labels.unsqueeze(-1))
# Sum across sequence dimension, normalize by length
traj_log_probs = log_probs_gathered.squeeze(-1).sum(dim=-1)
return traj_log_probs
The training loop orchestrates five key steps: forward pass through the LLM, partition function evaluation, reward computation, reference policy log-prob collection, and the combined FlowRL + importance sampling loss. The reference model is frozen to provide a stable baseline for KL divergence.
Practical Guidance
Hyperparameter Configuration
Hyperparameter
Range
Recommended
Notes
beta (reward scale)
0.5-2.0
1.0
Controls reward emphasis in distribution matching. Higher values make distribution sharper around high-reward solutions.
Learning rate (LLM)
1e-5 to 1e-4
5e-5
Smaller than standard supervised tuning; distribution matching is more sensitive to step size.
Learning rate (partition_fn)
1e-4 to 1e-3
5e-4
Can be higher than LLM since partition function has fewer parameters.
Batch size
32-256
64
Larger batches improve distribution estimate stability; smaller batches add regularization.
Sequence length (train)
256-2048
512-1024
Must accommodate CoT; longer sequences benefit more from length normalization.
Clip ratio (importance)
0.15-0.3
0.2
Standard PPO clipping; prevents outlier examples from dominating.
Gradient clip norm
0.5-2.0
1.0
Prevents gradient explosion from length normalization interactions.
When to Use FlowRL
You have multiple valid reasoning paths for a task (math problems, code generation)
Solution diversity is important for robustness and generalization
You're training on chain-of-thought reasoning where intermediate steps vary
Your reward model provides meaningful gradations (not just binary pass/fail)
Your baseline RL approach exhibits mode collapse or repetitive outputs
When NOT to Use FlowRL
Task has a single dominant optimal solution (e.g., translation with fixed answer)
Reward signal is extremely sparse or noisy (partition function won't find stable distributions)
Training stability is critical and you need maximum simplicity (basic PPO may be safer)
Your LLM is very large (>100B) without infrastructure for stable distributed training
You lack a reliable reward model; FlowRL amplifies reward signal quality issues
Common Pitfalls
Gradient explosion from length normalization: When sequences vary wildly in length (10 tokens to 1000+ tokens), the 1/length factor can create unstable gradient magnitudes. Solution: bucketing sequences by length or using length-conditional beta scheduling.
Partition function overfitting: If the partition function capacity is too high relative to batch size, it may memorize rewards rather than learning stable distributions. Solution: keep the MLP small (3 layers, 128-256 hidden dim) and monitor validation curves.
Off-policy divergence: If you use very old trajectory data with importance sampling, the clipped weights become mostly 1s and corrections fail. Solution: refresh trajectory buffer frequently (every 1-2 epochs) and monitor importance weight magnitudes.
Reference policy mismatch: If the reference model diverges from the training model too far, log-prob differences explode. Solution: periodically update reference model checkpoint and monitor KL divergence.
Reward scale sensitivity: Unlike reward maximization which naturally normalizes through expectation, distribution matching is sensitive to reward magnitude. Solution: normalize rewards to [0, 1] or use adaptive reward scaling based on running statistics.
