Scale RL training for language models by automatically adapting problem difficulty to match model capabilities using verifiable reward environments—achieving 3.37% absolute improvement on reasoning benchmarks through environment diversity rather than compute scaling alone.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Scale RL training for language models by automatically adapting problem difficulty to match model capabilities using verifiable reward environments—achieving 3.37% absolute improvement on reasoning benchmarks through environment diversity rather than compute scaling alone.
Scale LLM Reasoning Through Adaptive Difficulty Environments
Training language models with reinforcement learning on fixed problem distributions fails to drive meaningful improvement—learning signals vanish as problems become either trivially easy or impossibly hard relative to the model's current capabilities. RLVE solves this through adaptive verifiable environments that algorithmically generate problems calibrated to the model's current performance level.
Rather than investing more compute in static training setups, RLVE multiplies training effectiveness by continuously adapting problem difficulty and diversifying the environment suite. A 1.5B parameter model trained on RLVE-Gym (400 engineered environments) achieves 3.37% absolute improvement, substantially outperforming 3x more compute on fixed problem sets (0.49% gain).
Core Concept
RLVE reframes RL training as an adaptive curriculum problem. The system maintains a portfolio of verifiable environments—problem generators with explicit reward functions—and dynamically adjusts problem difficulty to match each model's learning trajectory. Rather than one difficult benchmark, use 400 environments calibrated across difficulty ranges, allowing the model to graduate from easy to hard naturally.
Verifiable environments enable cost-free reward signals: problems have explicit solutions (math, code, logic) that can be automatically checked without oracle models or human labeling.
"""
Generate a problem with ground-truth solution.
Returns:
problem_text: Problem description as string
solution: Ground-truth solution (comparable to model output)
"""
raise
def
verify_solution
self, model_output, ground_truth
"""
Check if model output matches ground truth.
Args:
model_output: Model's response text
ground_truth: Expected solution
Returns:
is_correct: Boolean correctness
"""
raise
class
MathProofEnvironment
VerifiableEnvironment
"""Generates mathematical proof problems of variable difficulty."""
def
generate_problem
self
"""Generates problems like: 'Prove that sqrt(2) is irrational'."""
int
2
self
3
# Difficulty scales problem depth
f"""Prove the following with {num_concepts} logical steps:
[Problem statement parametrized by difficulty]"""
self
return
def
verify_solution
self, model_output, solution
"""Check proof structure: has required logical steps, no contradictions."""
'\n'
if
len
len
return
True
# Simplified: check step count
return
False
class
CodeGenerationEnvironment
VerifiableEnvironment
"""Generates coding problems with test cases."""
def
generate_problem
self
"""Generate problems like: 'Write a function that sorts and filters'."""
'sorting'
'search'
'graph'
'dp'
self
f"Implement {problem_class} with constraints: {constraints}"
self
return
def
verify_solution
self, model_output, test_cases
"""Execute model code against test cases."""
try
exec
for
in
if
return
False
return
True
except
return
False
Step 2: Implement Adaptive Curriculum Controller
Monitor model performance and route to environments at appropriate difficulty levels.
defadaptive_curriculum_controller(model, environment_pool, target_success=0.5):
"""
Adaptively select environments matching model capability.
Args:
model: Language model to train
environment_pool: List of VerifiableEnvironment instances
target_success: Target accuracy (keep ~50% to maintain learning signal)
Yields:
(problem, ground_truth, environment): Training samples
"""# Track per-environment performance
env_success_rates = {env.name: 0.5for env in environment_pool}
env_sample_counts = {env.name: 0for env in environment_pool}
for training_step inrange(num_training_steps):
# Select environment based on recent performance
selected_env = select_environment_by_performance(
env_success_rates, env_sample_counts, temperature=1.0
)
# Adjust difficulty toward target success rate
current_success = env_success_rates[selected_env.name]
if current_success > target_success:
selected_env.difficulty *= 1.1# Increase difficultyelif current_success < target_success - 0.1:
selected_env.difficulty *= 0.95# Decrease difficulty# Clamp difficulty in valid range
selected_env.difficulty = np.clip(selected_env.difficulty, 0.5, 3.0)
# Generate problem and get model response
problem, solution = selected_env.generate_problem()
model_response = model.generate(problem, max_tokens=2048)
# Verify correctness
is_correct = selected_env.verify_solution(model_response, solution)
# Update success rate (exponential moving average)
alpha = 0.1
old_rate = env_success_rates[selected_env.name]
env_success_rates[selected_env.name] = (
alpha * float(is_correct) + (1 - alpha) * old_rate
)
env_sample_counts[selected_env.name] += 1yield (problem, solution, selected_env), is_correct
Step 3: Verifiable Reward Computation
Define reward functions based on solution correctness (verifiable) rather than oracle models.
defcompute_verifiable_reward(model_response, environment, ground_truth):
"""
Compute reward based on automatic solution verification.
Args:
model_response: Model's generated output
environment: VerifiableEnvironment instance
ground_truth: Expected solution
Returns:
reward: Float in [-1, 1]
"""# Primary signal: correctness
is_correct = environment.verify_solution(model_response, ground_truth)
if is_correct:
# Bonus for efficiency/conciseness (e.g., shorter code)
efficiency_bonus = compute_efficiency_metric(model_response)
return1.0 + 0.1 * efficiency_bonus
else:
# Penalty: how far from correctness?
partial_credit = compute_partial_credit(model_response, ground_truth)
return -1.0 + partial_credit
defcompute_partial_credit(model_output, solution):
"""
Award partial credit for partially correct solutions.
Examples: right approach but wrong constant, correct structure with bugs.
"""# Check if output has correct logical structureif has_correct_structure(model_output, solution):
return0.5# Check if output attempts the right approachif uses_correct_approach(model_output, solution):
return0.2return0.0
Step 4: Train with Adaptive RL Loop
Integrate adaptive curriculum with RL optimization (e.g., PPO or GRPO).
deftrain_with_adaptive_rl(model, environment_pool, num_steps=100000):
"""
Main training loop combining adaptive curriculum with on-policy RL.
Args:
model: LLM to optimize
environment_pool: Verifiable environments
num_steps: Total training iterations
"""from transformers import AutoModelForCausalLM
import torch.optim as optim
optimizer = optim.AdamW(model.parameters(), lr=2e-5)
curriculum = adaptive_curriculum_controller(model, environment_pool)
for step inrange(num_steps):
# Get problem from adaptive curriculum
(problem, solution, env), is_correct = next(curriculum)
# Generate response with temperature for explorationwith torch.no_grad():
response = model.generate(
problem, max_tokens=2048, temperature=0.8
)
# Compute verifiable reward
reward = compute_verifiable_reward(response, env, solution)
# On-policy RL update (simplified; use GRPO in practice)
logprobs = model.forward(response).log_probs
loss = -reward * logprobs.mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (step + 1) % 1000 == 0:
# Evaluate on benchmark
benchmark_score = evaluate_on_benchmarks(model)
print(f"Step {step}: Reward {reward:.2f}, Benchmark {benchmark_score:.2f}")
return model
Practical Guidance
When to Use RLVE:
Large-scale RL training for reasoning tasks (math, coding, logic)
Scenarios where compute scaling alone provides diminishing returns
Domains with automatic verification (not human evaluation)
When NOT to Use:
Real-world tasks without ground-truth verification (subjective quality)
Limited environment diversity available
Preference learning scenarios requiring human feedback
Hyperparameters and Configuration:
Target success rate: 0.5 (maintains learning signal; adjust ±0.1 based on domain)
Difficulty bounds: [0.5, 3.0] (start conservative; expand if needed)
Environment pool size: 100+ for diverse coverage; 400+ for production
Curriculum temperature: 1.0 (uniform selection); increase to explore easier tasks more
Pitfalls to Avoid:
Static environment pools - Diversity drives improvement; single benchmark provides weak signal
Ignoring partial credit - Binary right/wrong loses information; award gradual credit for partial solutions
Over-aggressive difficulty scaling - Adapt smoothly (×1.05-1.1); jumping difficulty too fast destabilizes learning
Verification gaps - Ensure verifiers are fast and accurate; bugs in verification corrupt reward signal
