메뉴
Bridge imitation learning and experience-driven RL by collecting state-based supervision from agents' own actions. Trigger: improve agent generalization when expert demonstrations are limited and environments lack dense rewards.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill agent-early-experience명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
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.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | agent-early-experience |
| title | Agent Learning via Early Experience: Implicit World Modeling and Self-Reflection |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2510.08558 |
| keywords | ["agent-learning","early-experience","world-modeling","self-reflection","imitation-learning"] |
| description | Bridge imitation learning and experience-driven RL by collecting state-based supervision from agents' own actions. Trigger: improve agent generalization when expert demonstrations are limited and environments lack dense rewards. |
Early Experience Learning addresses a fundamental challenge in agent training: expert demonstrations are limited and don't cover environment diversity, while many environments (websites, interactive systems) lack verifiable rewards. The approach collects supervision from the agent's own exploratory actions—using resulting states as training targets. This bridges imitation learning and RL by grounding policy in environmental dynamics through implicit world modeling and self-reflection.
The key insight: An agent's own experience (state transitions) is stronger supervision than limited expert demonstrations, and agent reflection on suboptimal actions drives improvement.
Agents explore environments while recording their actions and resulting states.
class EarlyExperienceCollector:
"""
Collect state-based supervision from agent's own exploratory actions.
"""
def __init__(self, agent_model, environment, max_steps=50):
self.agent = agent_model
self.env = environment
self.max_steps = max_steps
self.experiences = []
def collect_trajectory(self, task_description):
"""
Agent explores a task, recording (state, action, result_state).
Args:
task_description: Human-readable task prompt
Returns:
Trajectory with experience tuples
"""
trajectory = {
"task": task_description,
: [],
: [],
: [],
: []
}
state = .env.reset(task_description)
trajectory[].append(state)
step (.max_steps):
observation = .env.observe()
trajectory[].append(observation)
action_output = .agent.generate(
,
max_tokens=
)
action = parse_action(action_output)
trajectory[].append(action)
:
result_state = .env.step(action)
trajectory[].append(result_state)
.env.is_completed():
trajectory[] =
Exception e:
trajectory[].append({: (e)})
trajectory[] = (trajectory[])
trajectory
():
all_trajectories = []
task task_list:
_ (num_trajectories_per_task):
trajectory = .collect_trajectory(task)
all_trajectories.append(trajectory)
all_trajectories
Train agent to predict result states given actions, grounding it in environment dynamics.
class ImplicitWorldModelTrainer:
"""
Train world model implicitly: predict state transitions.
"""
def __init__(self, agent_model):
self.agent = agent_model
def create_world_modeling_examples(self, trajectories):
"""
Convert trajectories into (state, action, predicted_state) tuples.
"""
training_examples = []
for traj in trajectories:
for step_idx in range(len(traj["actions"])):
current_state = traj["states"][step_idx]
action = traj["actions"][step_idx]
result_state = traj["result_states"][step_idx]
example = {
"input": f"State: {format_state(current_state)}\n"
f"Action: {action}\n"
f"Resulting state: ",
"target": format_state(result_state)
}
training_examples.append(example)
return training_examples
def train_world_model(self, examples, num_epochs=5):
"""
Fine-tune agent to predict state transitions.
This implicitly teaches environmental dynamics.
"""
optimizer = torch.optim.Adam(self.agent.parameters(), lr=1e-5)
for epoch in range(num_epochs):
epoch_loss = 0
for example in examples:
# Generate prediction
output = self.agent.generate(
example["input"],
max_tokens=200
)
# Compute loss vs. ground truth state
loss = compute_state_matching_loss(
output,
example["target"]
)
epoch_loss += loss.item()
# Update
optimizer.zero_grad()
loss.backward()
optimizer.step()
avg_loss = epoch_loss / len(examples)
print(f"Epoch {epoch}: world model loss = {avg_loss:.4f}")
return self.agent
Agents evaluate their own actions and learn from mistakes without external feedback.
class SelfReflectionTrainer:
"""
Train agents to reflect on suboptimal actions and improve.
"""
def __init__(self, agent_model):
self.agent = agent_model
def create_reflection_examples(self, trajectories, success_threshold=0.8):
"""
Identify suboptimal actions and create reflection examples.
"""
reflection_examples = []
for traj in trajectories:
# Assess if trajectory was successful
task_success = traj.get("completed", False)
num_steps = traj["num_steps"]
efficiency = 1.0 / (num_steps + 1) # Penalize long trajectories
success_score = 1.0 if task_success else 0.0
overall_score = 0.7 * success_score + 0.3 * efficiency
# If not perfect, create reflection example
if overall_score < success_threshold:
# Identify which actions were problematic
for step_idx in range(len(traj["actions"])):
action = traj["actions"][step_idx]
state = traj["states"][step_idx]
result = traj["result_states"][step_idx]
# Check if action led toward goal
progress = self.evaluate_action_progress(
action,
state,
result,
traj["task"]
)
if progress < 0.5: # Action was suboptimal
# Create reflection prompt
reflection_example = {
"input": f"Task: {traj['task']}\n"
f"Current state: {format_state(state)}\n"
f"Action taken: {action}\n"
f"Result: {format_state(result)}\n"
f"Reflection (what went wrong): ",
"target": generate_reflection(action, result)
}
reflection_examples.append(reflection_example)
return reflection_examples
def train_self_reflection(self, examples, num_epochs=5):
"""
Train agent to articulate what went wrong with its actions.
"""
optimizer = torch.optim.Adam(self.agent.parameters(), lr=1e-5)
for epoch in range(num_epochs):
epoch_loss = 0
for example in examples:
# Generate reflection
output = self.agent.generate(
example["input"],
max_tokens=100
)
# Loss: how well does reflection match ground truth
loss = compute_reflection_loss(
output,
example["target"]
)
epoch_loss += loss.item()
# Update
optimizer.zero_grad()
loss.backward()
optimizer.step()
avg_loss = epoch_loss / len(examples)
print(f"Epoch {epoch}: reflection loss = {avg_loss:.4f}")
return self.agent
def evaluate_action_progress(self, action, state, result, task):
"""
Measure if action moved agent closer to task goal.
"""
state_embedding = encode_state(state)
result_embedding = encode_state(result)
task_embedding = encode_task(task)
# Cosine similarity to goal
similarity_improvement = (
cosine_similarity(result_embedding, task_embedding) -
cosine_similarity(state_embedding, task_embedding)
)
return max(0.0, similarity_improvement)
Integrate world modeling and self-reflection into coherent training pipeline.
def train_with_early_experience(agent, environment, task_list, config):
"""
Full early experience training pipeline.
"""
# Phase 1: Collect early experiences
print("Phase 1: Collecting early experiences")
collector = EarlyExperienceCollector(agent, environment)
trajectories = collector.collect_batch(
task_list,
num_trajectories_per_task=3
)
print(f"Collected {len(trajectories)} trajectories")
# Phase 2: World modeling training
print("\nPhase 2: Implicit world modeling")
world_trainer = ImplicitWorldModelTrainer(agent)
wm_examples = world_trainer.create_world_modeling_examples(trajectories)
agent = world_trainer.train_world_model(wm_examples, num_epochs=5)
# Phase 3: Self-reflection training
print("\nPhase 3: Self-reflection learning")
reflection_trainer = SelfReflectionTrainer(agent)
ref_examples = reflection_trainer.create_reflection_examples(trajectories)
agent = reflection_trainer.train_self_reflection(
ref_examples,
num_epochs=5
)
return agent
Assess early experience training and prepare for downstream RL.
def evaluate_early_experience_agent(agent, benchmark_tasks):
"""
Test agent on diverse tasks post-early-experience training.
"""
success_rates = {}
generalization_scores = []
for task in benchmark_tasks:
successes = 0
trials = 5
for _ in range(trials):
trajectory = collect_single_trajectory(agent, task)
if trajectory.get("completed"):
successes += 1
task_success_rate = successes / trials * 100
success_rates[task] = task_success_rate
# Measure generalization: novel tasks similar to training
generalization = measure_transfer_learning(agent, task)
generalization_scores.append(generalization)
print(f"Average success rate: {np.mean(list(success_rates.values())):.1f}%")
print(f"Generalization: {np.mean(generalization_scores):.2f}")
return {
"success_rates": success_rates,
"generalization": np.mean(generalization_scores)
}
Hyperparameters:
When to Use:
When NOT to Use:
Agent Learning via Early Experience — arXiv:2510.08558