Convert scaling and efficiency papers into practical resource planning guides. Extracts empirical scaling laws, compute-optimal allocation rules, and budget-performance trade-offs. Use this skill when extracting skills from Category 11 (Scaling and Efficiency) papers — Chinchilla-style scaling law papers, Flash Attention efficiency papers, or knowledge distillation studies.
Convert scaling and efficiency papers into practical resource planning guides. Extracts empirical scaling laws, compute-optimal allocation rules, and budget-performance trade-offs. Use this skill when extracting skills from Category 11 (Scaling and Efficiency) papers — Chinchilla-style scaling law papers, Flash Attention efficiency papers, or knowledge distillation studies.
When to Use
Apply this skill when you encounter arXiv papers that:
Establish scaling laws or empirical relationships (GPT-3 scaling, Chinchilla compute-optimal training, scaling laws for vision models)
Study compute efficiency, parameter efficiency, or memory efficiency (Flash Attention, model compression, knowledge distillation, quantization)
Compare resource-performance tradeoffs (training with fewer parameters, inference latency optimization, energy consumption analysis)
Provide specific empirical numbers from large-scale experiments (hours of compute, number of parameters, tokens seen, final accuracy)
Discuss budget constraints and how to optimize performance within them
Study diminishing returns, saturation points, or scaling frontiers
Examples: "Training Compute-Optimal Large Language Models", "Flash Attention: Fast and Memory-Efficient Exact Attention", "Scaling Laws for Neural Language Models", "Chinchilla Scaling Laws", "LoRA: Low-Rank Adaptation", papers on pruning, distillation, quantization.
When NOT to Use
Do not use this skill for:
Papers that propose a method without establishing resource-performance relationships
Papers with only anecdotal performance numbers (no systematic sweep or scaling study)
Purely algorithmic improvements without explicit compute/memory/latency analysis
Papers that discuss efficiency but don't quantify tradeoffs
Optimization papers that optimize a single component without resource-outcome mapping
Survey papers that discuss efficiency without contributing new empirical laws
Extraction Template
For Scaling Law Papers
Step 1: Resource Definition
Extract what resource(s) are being studied.
**Primary Resource:** What is being measured? (compute [FLOPs, GPU-hours, wall-clock time], data [tokens, examples], parameters [model size], or hardware [memory, batch size])
**Secondary Resources:** What other resources are tracked or discussed?
**Measurement Units:** How are resources quantified? (e.g., FLOPs, GPUs × hours, number of parameters in billions)
**Hardware Specification:** What hardware was used? (GPU type, number of GPUs, mixed precision settings)
Step 2: Empirical Laws
Document discovered relationships between resources and outcomes.
**Primary Law:** L = a * N^b or equivalent mathematical form (e.g., loss decreases as N^-0.07 where N is parameter count)
**Exponent & Coefficient:** Specific numerical values from the paper (e.g., b = -0.07, a = 0.45)
**Validity Range:** Over what resource range does this law hold? (e.g., from 1M to 100B parameters)
**Assumptions:** What conditions must be met for the law to apply? (e.g., compute-optimal scaling, no architectural changes)
**Functional Form:** Is it power law? Logarithmic? Sigmoidal? Saturation?
Step 3: Specific Numbers & Benchmarks
Extract concrete performance points from experiments.
**Training Configuration:**- N parameters: {value}
- D tokens (or training examples): {value}
- Compute budget: {value} GPU-hours (or FLOPs)
- Final loss (or benchmark score): {value}
- Time to convergence: {wall-clock time}
**Inference Configuration:**- Model size: {value}
- Latency per token: {milliseconds}
- Throughput: {tokens/second}
- Memory footprint: {GB}
Step 4: Compute-Optimal Allocation
Extract how to allocate compute optimally.
**Budget Constraint:** Given X GPU-hours, how should compute be split between:
- Training tokens vs. parameters (data-parameter tradeoff)?
- Number of training runs vs. model size?
- Batch size vs. learning rate?
**Chinchilla-like Trade-offs:** For a fixed compute budget:
- If you increase parameters by 2x, how should you scale training data?
- Typical optimal ratio: (training tokens) = k × (parameters)
**Sensitivity:** Which allocation decisions matter most? (parameter count > token count > batch size?)
For Efficiency Papers
Step 1: Efficiency Dimension
Extract what aspect of efficiency is being optimized.
**Target Efficiency:** What is being reduced? (memory, latency, FLOPs, energy, parameter count)
**Baseline:** What standard approach is being compared against?
**Improvement Metric:** How is improvement measured? (speedup factor, memory reduction %, FLOPs saved)
**Hardware Context:** What hardware is this optimized for? (A100 GPUs, TPUs, mobile devices, CPUs)
Step 2: Performance-Efficiency Tradeoff
Document what is gained and lost.
**Accuracy Preservation:** How much accuracy is retained compared to the baseline?
- At zero efficiency cost: {baseline accuracy}
- At 2x speedup: {accuracy}
- At 10x speedup: {accuracy}
- At maximum efficiency: {accuracy drop %}
**Latency Improvement:**- Baseline latency: {ms}
- With technique: {ms}
- Speedup factor: {x}
**Memory Savings:**- Baseline memory: {GB}
- With technique: {GB}
- Reduction: {%}
Step 3: Practical Integration
Extract how to implement the efficiency gain.
**Prerequisites:** What is required to use this technique? (specific hardware, software framework, model architecture)
**Hyperparameter Sensitivity:** Which choices most affect the accuracy-efficiency tradeoff?
**Implementation Complexity:** How complex is this to implement? (simple replace, moderate refactoring, major rewrite)
**Reproducibility:** What do you need to reproduce the claimed efficiency gains?
Step 4: Diminishing Returns Analysis
Document where efficiency gains saturate.
**Efficiency Curve:** As the technique is pushed harder, what happens?
- At moderate application: {tradeoff point}
- At aggressive application: {tradeoff point}
- At maximum application: {tradeoff point}
**Saturation Point:** Beyond what parameter does further optimization yield little benefit?
**Bottleneck Shift:** What becomes the new bottleneck after this optimization?
For Compute Budget & Allocation Papers
Step 1: Budget Constraints
Extract how different resource budgets change the problem.
**Budget Scenarios:**- Ultra-low (< 1 GPU-day): What is achievable?
- Low (1-100 GPU-days): What methods become viable?
- Medium (100-10K GPU-days): What is state-of-the-art?
- High (10K+ GPU-days): What frontiers can be explored?
**Resource Dimensions:** How do tradeoffs change across:
- Wall-clock time constraint (days vs weeks vs months)
- Total compute constraint (GPU-hours)
- Parameter budget (model size)
- Data availability
Step 2: Allocation Strategy
Document how to spend a budget optimally.
**Allocation Rule:** Given a budget of X:
- How much goes to model size? (percentage or parameter count)
- How much goes to training compute? (tokens or GPU-hours)
- How much goes to data collection/preprocessing?
- How much goes to experimental iteration?
**Optimization Surface:** What allocation strategy is Pareto-optimal across multiple objectives? (speed, accuracy, cost)
Output Skill Format
Generate a new SKILL.md with the following structure:
Core Principle (1 paragraph): What resource-outcome relationship does this paper establish? Why does it matter?
Empirical Law or Scaling Relationship (1 paragraph + equation form): Specific mathematical relationship discovered. What are the exponents/coefficients from experiments?
Concrete Numbers Table (bullet points or table): Specific performance points: (parameter count, training tokens, GPU-hours) → (final loss/accuracy, inference latency, memory)
Budget Calculator (1 paragraph + decision tree): Given a compute budget, what should practitioners do? Parameter allocation rules.
Practical Guidance (3-5 bullet points): Implementation details, hyperparameter sensitivity, what decisions matter most
Diminishing Returns & Saturation (1 paragraph): Where does this technique stop working? What are the limits?
When to Use This Skill (1-2 sentences): What decision-making scenarios apply this resource-outcome relationship?
Length: 150-250 lines
Processing Instructions
Identify resource type: What resource is being studied? (compute, data, parameters, memory, latency, energy)
Obtain the paper: Fetch HTML from arxiv.org/html/{arxiv_id}, fallback to PDF
Extract empirical laws: What mathematical relationships are discovered? Specific exponents and coefficients.
Tabulate specific numbers: Create (resource) → (outcome) mappings from experiments
Identify allocation strategy: How should budget be split between model size, training data, and compute?
Find tradeoff curves: Where is the Pareto frontier? What is the accuracy-latency-memory space?
Determine saturation: At what point do further gains become diminishing?
Build budget calculator: Create a practical tool practitioners can use to plan their own allocations
Validate against real experiments: Confirm the extracted laws and numbers match the paper's reported results
Quality Checks
Paper establishes quantitative resource-outcome relationships (not just qualitative discussion)
Specific empirical numbers are extracted (parameter counts, token counts, GPU-hours, final metrics)
Scaling law or efficiency curve is expressed in mathematical form with numerical coefficients
Assumptions for the empirical law are explicitly stated (e.g., compute-optimal regime, specific hardware)
Hardware specifications are documented (GPU type, precision, batch size) so results can be contextualized
Tradeoff curves show accuracy loss at different efficiency levels
Budget allocation rules are specified (e.g., optimal parameter-to-token ratio)
Saturation points or limits are identified
Output skill provides actionable guidance for practitioners with specific budgets
Keywords (5-10) include "scaling", "efficiency", "compute-optimal", "resource", or domain-specific terms
Description is under 1024 characters
Engine tag matches skillxiv-v0.0.2-claude-opus-4.6
Common Pitfalls
Confusing this category with Component Innovation (Category 5): Scaling & Efficiency papers focus on resource-PERFORMANCE RELATIONSHIPS across many configurations. Component papers optimize one component (e.g., replacing one attention variant with another) without systematic resource study.
Extracting qualitative efficiency discussion instead of quantitative laws: Paper must provide specific numbers and curves. Skip if it's only "technique X is more efficient" without data.
Missing the specific hardware context: Efficiency gains are hardware-specific. Flash Attention has different speedups on different GPUs. Document this.
Assuming one scaling law applies everywhere: Laws often only hold in certain regimes (e.g., compute-optimal regime, or above a certain model size). Capture boundary conditions.
Overlooking data-parameter tradeoff: In scaling papers, both tokens and parameters scale. Capture both and their relationship.
Ignoring practical integration challenges: A technique may be theoretically efficient but hard to implement in practice. Extract both the theoretical efficiency and practical gotchas.
Missing diminishing returns: Many efficiency techniques have saturation points. Document where gains stop.
Extracting without wall-clock validation: FLOPs reduction doesn't always mean wall-clock speedup. Verify what the paper actually measures.
