메뉴
Guide LLM exploration through rubric-based scaffolding that gradually diminishes, enabling models to internalize reasoning patterns while maintaining exploration quality for robust RL training.
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 ruscarl-rubric-scaffolded-rl명령은 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 | ruscarl-rubric-scaffolded-rl |
| title | RuscaRL: Rubric-Scaffolded RL for General LLM Reasoning |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2508.16949 |
| keywords | ["reinforcement-learning","rubric-guidance","exploration-guidance","reasoning-scaffolding","curriculum-learning"] |
| description | Guide LLM exploration through rubric-based scaffolding that gradually diminishes, enabling models to internalize reasoning patterns while maintaining exploration quality for robust RL training. |
The core challenge in LLM RL is exploration: what cannot be explored cannot be learned. RuscaRL (Rubric-Scaffolded RL) addresses this by using checklist-style rubrics in two phases: first as explicit guidance during exploration (gradually fading), then as reference for scoring. This approach breaks the constraint that LLM limitations restrict sample quality, enabling models to learn general reasoning capabilities through guided exploration.
Create structured evaluation checklist templates:
from dataclasses import dataclass
from typing import List, Dict, Any
from enum import Enum
class RubricLevel(Enum):
EXPLICIT = "explicit" # Full rubric in prompt
IMPLICIT = "implicit" # Rubric referenced but not detailed
ABSENT = "absent" # No rubric mentioned
@dataclass
class RubricCriterion:
"""Single rubric evaluation criterion."""
name: str
description: str
exemplar_good: str
exemplar_bad: str
weight: float =
:
task_name:
criteria: [RubricCriterion]
instructions:
() -> :
guidance =
guidance +=
criterion .criteria:
guidance +=
guidance +=
guidance +=
guidance
:
():
.rubrics: [, TaskRubric] = {}
():
.rubrics[task_type] = rubric
() -> TaskRubric:
.rubrics.get(task_type)
() -> TaskRubric:
TaskRubric(
task_name=,
criteria=[
RubricCriterion(
name=,
description=,
exemplar_good=,
exemplar_bad=,
weight=
),
RubricCriterion(
name=,
description=,
exemplar_good=,
exemplar_bad=,
weight=
),
RubricCriterion(
name=,
description=,
exemplar_good=,
exemplar_bad=,
weight=
),
RubricCriterion(
name=,
description=,
exemplar_good=,
exemplar_bad=,
weight=
)
],
instructions=
)
Use rubrics to enhance exploration quality:
class RubricGuidedExploration:
"""Guide exploration with rubric scaffolding."""
def __init__(
self,
model: "LLM",
rubric_library: RubricLibrary
):
self.model = model
self.rubric_library = rubric_library
def generate_with_rubric_guidance(
self,
task: str,
task_type: str,
guidance_level: RubricLevel = RubricLevel.EXPLICIT,
temperature: float = 0.9
) -> str:
"""
Generate response with rubric guidance.
Guidance level controls how much structure to provide.
"""
rubric = self.rubric_library.get_rubric(task_type)
if guidance_level == RubricLevel.EXPLICIT:
# Full rubric in prompt
prompt = f"{rubric.render_as_guidance()}\n\nProblem: {task}\n\nSolution:"
elif guidance_level == RubricLevel.IMPLICIT:
# Rubric referenced but not detailed
criterion_names = ", ".join(c.name for c in rubric.criteria)
prompt = f"Solve this task considering: {criterion_names}.\n\nProblem: {task}\n\nSolution:"
elif guidance_level == RubricLevel.ABSENT:
# No rubric
prompt = f"Solve this problem:\n\n{task}\n\nSolution:"
# Generate with higher temperature for exploration
response = self.model.generate(
prompt,
temperature=temperature,
max_tokens=500
)
return response
def collect_exploration_trajectories(
self,
tasks: List[Dict],
task_type: str,
num_trajectories_per_task: int = 5,
guidance_schedule: List[RubricLevel] = None
) -> List[Dict]:
"""
Collect exploration trajectories with curriculum guidance schedule.
"""
if guidance_schedule is None:
# Default: explicit -> implicit -> absent
guidance_schedule = [
RubricLevel.EXPLICIT,
RubricLevel.EXPLICIT,
RubricLevel.IMPLICIT,
RubricLevel.IMPLICIT,
RubricLevel.ABSENT
]
trajectories = []
for task_idx, task in enumerate(tasks):
for traj_idx in range(num_trajectories_per_task):
# Select guidance level from schedule
schedule_idx = min(traj_idx, len(guidance_schedule) - 1)
guidance_level = guidance_schedule[schedule_idx]
# Generate with guidance
response = self.generate_with_rubric_guidance(
task["prompt"],
task_type,
guidance_level
)
trajectories.append({
"task": task["prompt"],
"response": response,
"guidance_level": guidance_level,
"task_id": task.get("id"),
"expected_answer": task.get("answer")
})
return trajectories
Create LLM judge using rubrics:
class RubricBasedJudge:
"""Evaluate responses using rubric criteria."""
def __init__(
self,
judge_model: "LLM",
rubric_library: RubricLibrary
):
self.judge_model = judge_model
self.rubric_library = rubric_library
def score_response(
self,
task: str,
response: str,
task_type: str
) -> Dict[str, float]:
"""
Score response against rubric criteria.
Returns: {criterion_name: score, "overall": score}
"""
rubric = self.rubric_library.get_rubric(task_type)
scores = {}
# Score each criterion
for criterion in rubric.criteria:
score = self._score_criterion(
task, response, criterion
)
scores[criterion.name] = score
# Weighted overall score
overall = sum(
scores[c.name] * c.weight
for c in rubric.criteria
) / sum(c.weight for c in rubric.criteria)
scores["overall"] = overall
return scores
def _score_criterion(
self,
task: str,
response: str,
criterion: RubricCriterion
) -> float:
"""Score single criterion using LLM judge."""
prompt = f"""Evaluate the following response against this criterion:
Criterion: {criterion.name}
Description: {criterion.description}
Good example: {criterion.exemplar_good}
Bad example: {criterion.exemplar_bad}
Task: {task}
Response: {response}
Rate this response on criterion '{criterion.name}' on a scale 0-1:
- 0: Does not satisfy criterion at all
- 0.5: Partially satisfies criterion
- 1: Fully satisfies criterion
Score (just the number, 0-1):"""
score_text = self.judge_model.generate(prompt, max_tokens=10)
# Extract numerical score
try:
score = float(score_text.strip())
return min(max(score, 0.0), 1.0) # Clamp to [0,1]
except:
return 0.5 # Default to middle
def score_batch(
self,
trajectories: List[Dict],
task_type: str
) -> List[Dict]:
"""Score multiple trajectories."""
for trajectory in trajectories:
scores = self.score_response(
trajectory["task"],
trajectory["response"],
task_type
)
trajectory["scores"] = scores
trajectory["reward"] = scores["overall"]
return trajectories
Train with rubric-guided RL:
class RuscaRLTrainer:
"""Training with rubric-scaffolded RL."""
def __init__(
self,
model: "LLM",
rubric_library: RubricLibrary,
learning_rate: float = 1e-5
):
self.model = model
self.rubric_library = rubric_library
self.explorer = RubricGuidedExploration(model, rubric_library)
self.judge = RubricBasedJudge(model, rubric_library)
self.optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
def train_with_rubric_curriculum(
self,
tasks: List[Dict],
task_type: str,
num_iterations: int = 5,
guidance_schedule: List[RubricLevel] = None
) -> Dict[str, float]:
"""
Train model with rubric curriculum.
Gradually remove rubric guidance over iterations.
"""
metrics = {
"iteration": [],
"avg_reward": [],
"without_guidance_reward": []
}
for iteration in range(num_iterations):
print(f"RuscaRL Iteration {iteration + 1}/{num_iterations}")
# Phase 1: Generate trajectories with rubric guidance
trajectories = self.explorer.collect_exploration_trajectories(
tasks,
task_type,
num_trajectories_per_task=3,
guidance_schedule=guidance_schedule
)
# Phase 2: Score trajectories using rubric
trajectories = self.judge.score_batch(trajectories, task_type)
# Compute reward statistics
rewards = [t["reward"] for t in trajectories]
avg_reward = sum(rewards) / len(rewards)
metrics["iteration"].append(iteration)
metrics["avg_reward"].append(avg_reward)
print(f" Avg Reward: {avg_reward:.4f}")
# Phase 3: Train on scored trajectories
for trajectory in trajectories:
# Get log prob of response
response = trajectory["response"]
reward = trajectory["reward"]
# Policy gradient loss
loss = -reward * self.model.get_log_prob(response)
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
# Phase 4: Evaluate without guidance
print(f" Evaluating without guidance...")
guidance_free_trajectories = self.explorer.collect_exploration_trajectories(
tasks,
task_type,
num_trajectories_per_task=1,
guidance_schedule=[RubricLevel.ABSENT] * 5
)
guidance_free_trajectories = self.judge.score_batch(
guidance_free_trajectories, task_type
)
guidance_free_rewards = [t["reward"] for t in guidance_free_trajectories]
avg_guidance_free = sum(guidance_free_rewards) / len(guidance_free_rewards)
metrics["without_guidance_reward"].append(avg_guidance_free)
print(f" Without Guidance Reward: {avg_guidance_free:.4f}")
return metrics
Researchers. (2024). Breaking the Exploration Bottleneck: Rubric-Scaffolded RL for General LLM Reasoning. arXiv preprint arXiv:2508.16949.