メニュー
Enable LLMs to scale reasoning length from O(n²) to O(n) by structuring thinking into fixed-size chunks with learnable cross-chunk summaries. Trigger: train reasoning models with unbounded or expensive chain-of-thought sequences.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ADu2021/skillXiv --skill markovian-thinkerコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
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 職業分類に基づく
SKILL.md を表示中
| name | markovian-thinker |
| title | The Markovian Thinker: Streaming Language Models with Decoupled Thinking |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2510.06557 |
| keywords | ["reasoning","efficient-inference","streaming","rl-training","context-management"] |
| description | Enable LLMs to scale reasoning length from O(n²) to O(n) by structuring thinking into fixed-size chunks with learnable cross-chunk summaries. Trigger: train reasoning models with unbounded or expensive chain-of-thought sequences. |
Traditional RL-based reasoning in LLMs suffers from quadratic compute scaling as thinking length grows. The Markovian Thinker solves this by decomposing reasoning into fixed-size chunks where the model learns to write concise summaries at chunk boundaries. This enables the context to reset between chunks while maintaining reasoning continuity—achieving linear scaling instead of quadratic.
The key insight: Markovian structure allows "chunked thinking" where state transitions depend only on the previous chunk summary, not the entire prior history.
Define your chunk size based on available context window and thinking complexity. Typical configurations use 8K-token chunks for a 16K context window.
class ChunkThinkingConfig:
chunk_size = 8192 # tokens per thinking segment
context_window = 16384
summary_max_tokens = 256 # compressed carryover state
model_size = "1.5B" # parameter count
def validate(self):
assert self.summary_max_tokens < self.chunk_size // 32
assert self.context_window >= 2 * self.chunk_size
The environment enforces chunk boundaries and manages context resets. It provides reward signals for reasoning quality while penalizing inefficient summaries.
class DeleteinkEnvironment:
def __init__(self, model, chunk_config):
self.model = model
self.chunk_size = chunk_config.chunk_size
self.summary_tokens = chunk_config.summary_max_tokens
def step(self, thinking_tokens, summary_text):
# Measure reasoning quality on current chunk
chunk_reward = self.evaluate_chunk(thinking_tokens)
# Penalize verbose summaries
summary_penalty = -0.01 * len(summary_text.split())
# Reset context for next chunk
self.context = self.build_context(summary_text)
return chunk_reward + summary_penalty
def evaluate_chunk(self, tokens):
# Verify solution correctness from this chunk alone
solution = self.extract_solution(tokens)
return self.reward_for_correctness(solution)
Use policy gradient methods (PPO, GRPO) but apply them within the chunked structure. The key difference from standard RL is that you train the model to optimize both reasoning quality within chunks and summary quality at boundaries.
def train_chunked_reasoning(model, dataset, config):
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for epoch in range(config.num_epochs):
for problem in dataset:
# Start with fresh context
context = f"Problem: {problem}\n"
accumulated_summary = ""
total_reward = 0
# Process thinking in chunks
for chunk_idx in range(config.max_chunks):
# Generate thinking for this chunk (8K tokens)
thinking = model.generate(
context,
max_new_tokens=config.chunk_size
)
# Extract solution attempt
solution = extract_solution(thinking)
# Evaluate and get reward
chunk_reward = evaluate_solution(solution, problem)
# Generate summary (must fit in ~256 tokens)
summary = model.generate_summary(
thinking,
max_tokens=config.summary_tokens
)
# Compute policy loss
loss = -chunk_reward * log_prob(thinking, summary)
# Backprop and update
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_reward += chunk_reward
accumulated_summary += f"\n[Summary {chunk_idx}]: {summary}"
# Reset context for next chunk
context = f"{problem}\n{accumulated_summary}\n"
# Early stopping if solution found
if is_correct_solution(solution):
break
At inference time, the model streams thinking across chunks, regenerating context at boundaries without storing full history.
def stream_chunked_thinking(model, problem, config):
context = f"Problem: {problem}\nThinking:\n"
summaries = []
max_thinking_tokens = 24000 # Total thinking budget
token_count = 0
chunk_num = 0
while token_count < max_thinking_tokens:
# Generate next chunk of thinking
chunk_thinking, chunk_tokens = model.stream_generate(
context,
max_tokens=config.chunk_size,
return_token_count=True
)
token_count += chunk_tokens
chunk_num += 1
# Try to extract solution
candidate_solution = extract_solution(chunk_thinking)
if is_valid_solution(candidate_solution):
return candidate_solution
# Generate summary for carry-forward
summary = extract_key_insights(chunk_thinking)
summaries.append(summary)
# Reset context with accumulated summaries
context = f"{problem}\nPrior insights: {'; '.join(summaries[-3:])}\nContinuing...\n"
print(f"Chunk {chunk_num} complete. Total: {token_count} tokens")
return extract_final_solution(chunk_thinking)
Hyperparameters:
When to Use:
When NOT to Use: