frontier

name	frontier
description	Explore cutting-edge research in a domain. Use when user says "frontier [domain]", "latest research in [domain]", "cutting edge", "latest in [domain]", "what's new in [field]", "[domain] trends", or wants L4-level current research.
version	0.1.0

Frontier: Cutting-Edge Research Explorer

Discover the latest research frontiers in any domain. Focus on L4 (cutting-edge) level: specific researchers, recent papers, emerging concepts.

When to Use

• "Frontier [domain]"
• "Latest research in [domain]"
• "What's cutting edge?"
• "What's new in [field]?"
• "[domain] trends"
• "2024-2025 [domain] papers"
• Understanding current research direction

When NOT to Use

• Learning fundamentals → use domain-vocab
• Understanding concept history → use trace
• Broad domain survey → use deep-dive

Core Value

Standing at the research frontier, you can see what's coming next.

Frontier provides temporal depth (what's happening NOW) vs. deep-dive's conceptual breadth.

Workflow

Phase 1: Domain Frontline Identification

Input: Domain name

Actions:

1. Identify active research areas within domain
2. Find top venues (conferences, journals) for recent work
3. Identify leading research groups/labs
4. Map current "hot topics" from recent publications

Output:

domain: "Large Language Models"
timeframe: "2024-2025"

active_frontiers:
  - name: "Reasoning & Chain-of-Thought"
    heat: "🔥🔥🔥"
    key_venues: ["NeurIPS", "ICLR", "ACL"]

  - name: "Efficient Inference"
    heat: "🔥🔥🔥"
    key_venues: ["MLSys", "ICML"]

  - name: "Multimodal Integration"
    heat: "🔥🔥"
    key_venues: ["CVPR", "NeurIPS"]

  - name: "Alignment & Safety"
    heat: "🔥🔥🔥"
    key_venues: ["NeurIPS", "ICML", "specialized workshops"]

leading_labs:
  - "OpenAI"
  - "Anthropic"
  - "Google DeepMind"
  - "Meta FAIR"
  - "Stanford HAI"

Phase 2: Recent Paper Discovery

Objective: Find the latest impactful papers (last 12-18 months)

Actions:

1. Query arXiv for recent papers in domain
2. Filter by:
   • Recency (prioritize last 6 months)
   • Early citation velocity
   • Author reputation
   • Venue quality (if published)
3. Identify paper types:
   • Breakthrough: New capability or paradigm
   • Advancement: Significant improvement on existing work
   • Survey: Comprehensive field overview
   • Position: Future direction proposals

Heuristics for "Frontier Paper":

Signal	Score	Note
Posted in last 3 months	+3	Very fresh
Posted in last 6 months	+2	Fresh
Author has 10K+ citations	+2	Established researcher
From top lab	+2	Resource-backed
High early citation velocity	+3	Community interest
Addresses known limitation	+2	Practical impact
Novel architecture/method name	+2	Potential paradigm

Output:

recent_papers:
  breakthroughs:
    - title: "..."
      authors: ["..."]
      date: "2024-11"
      arxiv: "2411.xxxxx"
      contribution: "..."
      why_frontier: "First to demonstrate..."

  advancements:
    - title: "..."
      authors: ["..."]
      date: "2024-10"
      improves_on: "Previous SOTA by X%"

  surveys:
    - title: "..."
      date: "2024-09"
      covers: "Complete overview of..."

Phase 3: Researcher Spotlight

Objective: Identify who's driving the frontier

Actions:

1. Extract frequent authors from recent papers
2. Identify rising stars (high recent output, growing citations)
3. Map researcher → lab → focus area
4. Find their most recent work

Output:

frontier_researchers:
  established_leaders:
    - name: "Researcher Name"
      affiliation: "Lab/University"
      focus: "Specific area"
      recent_work: "Paper title (2024)"
      twitter/website: "..."

  rising_stars:
    - name: "Researcher Name"
      affiliation: "..."
      focus: "..."
      why_notable: "First author on breakthrough X"

Phase 4: Emerging Concept Extraction

Objective: Identify newly coined terms and concepts

Actions:

1. Extract terms that appear frequently in recent papers but not in older ones
2. Identify new architectural components, techniques, benchmarks
3. Track terminology evolution
4. Note concepts that don't yet have Wikipedia entries

Output:

emerging_concepts:
  new_terms:
    - term: "Constitutional AI"
      coined_by: "Anthropic (2022)"
      meaning: "Training AI with explicit principles"
      maturity: "Established in research, entering practice"

    - term: "Mixture of Experts at Scale"
      coined_by: "Multiple (2024)"
      meaning: "Sparse expert routing for efficiency"
      maturity: "Active research area"

  evolving_terms:
    - term: "Alignment"
      old_meaning: "Value alignment (philosophical)"
      new_meaning: "Practical training techniques (RLHF, DPO)"

  watch_list:
    - term: "..."
      first_seen: "2024-Q4"
      papers_using: 5
      potential: "May become standard term"

Phase 5: Trend Analysis

Objective: Identify research trajectory and predict near-future directions

Actions:

1. Analyze paper volume trends by sub-topic
2. Identify declining vs. rising areas
3. Note convergence patterns (multiple lines merging)
4. Predict likely next developments

Output:

## Trend Analysis: LLMs (2024-2025)

### Rising Trends 📈
- **Reasoning**: Chain-of-thought → Tree-of-thought → Graph-of-thought
- **Efficiency**: Quantization, distillation, speculative decoding
- **Evaluation**: Moving beyond benchmarks to capability assessment

### Stable Trends ➡️
- **Scale**: Continued scaling with efficiency focus
- **Multimodal**: Text + Vision standard, Audio emerging

### Declining Trends 📉
- **Pure RLHF**: Shifting to DPO and variants
- **Single-task fine-tuning**: Replaced by instruction tuning

### Predicted Next (6-12 months)
1. **Agent architectures**: LLMs as components in larger systems
2. **Inference-time compute**: Scaling test-time, not just training
3. **Synthetic data**: Self-improvement loops

Phase 6: Frontier Vocabulary Export

Objective: Provide L4-level tokens for prompting

Output:

frontier_tokens:
  must_know:
    - "constitutional AI"
    - "direct preference optimization (DPO)"
    - "speculative decoding"
    - "mixture of experts"

  emerging:
    - "inference-time scaling"
    - "chain-of-thought distillation"
    - "self-play fine-tuning"

  researcher_names:
    - "Yann LeCun"
    - "Ilya Sutskever"
    - "Percy Liang"

  lab_names:
    - "Anthropic"
    - "EleutherAI"
    - "Together AI"

prompt_example:
  before: "How do I make my LLM better at reasoning?"
  after: "Compare chain-of-thought prompting vs inference-time compute scaling (like OpenAI o1) for improving reasoning. When would I use speculative decoding?"

Output Formats

Format A: Frontier Report (Default)

# [Domain] Frontier Report
**Generated:** {date}
**Timeframe:** Last 12 months

## Executive Summary
{3-4 sentences on current state of frontier}

## Hot Topics
{ranked list with heat indicators}

## Recent Breakthrough Papers
{top 5-10 papers with summaries}

## Key Researchers
{leaders + rising stars}

## Emerging Vocabulary
{new terms with definitions}

## Trend Analysis
{rising/stable/declining + predictions}

## L4 Token Vocabulary
{export for prompting}

Format B: Paper List

Compact list of recent papers with arXiv links.

Format C: Researcher Network

Visual map of who's working on what, collaboration patterns.

Format D: Timeline

timeline
    title LLM Frontier Timeline 2024
    2024-Q1 : Gemini Ultra
            : Mixtral 8x7B (Open MoE)
    2024-Q2 : Llama 3
            : Claude 3 Opus
    2024-Q3 : GPT-4o (multimodal)
            : Open weights momentum
    2024-Q4 : o1 (reasoning)
            : Efficiency focus

Timeframe Options

Option	Papers From	Focus
--fresh	Last 3 months	Bleeding edge
--recent	Last 12 months	Standard (default)
--year	Full calendar year	Annual review

Example Session

Input: "Frontier reinforcement learning"

Output Summary:

# Reinforcement Learning Frontier (2024-2025)

## Hot Topics
1. 🔥🔥🔥 **Offline RL / Decision Transformers**
2. 🔥🔥🔥 **RL from Human Feedback (RLHF) variants**
3. 🔥🔥 **World Models & Model-Based RL**
4. 🔥🔥 **Multi-Agent RL**
5. 🔥 **Safe RL**

## Recent Breakthroughs
- "Decision Transformer" lineage → applied to robotics
- DPO surpassing PPO for LLM alignment
- Dreamer v3: World models that scale

## Key Researchers
- Sergey Levine (Berkeley): Offline RL
- Chelsea Finn (Stanford): Meta-RL, robotics
- Pieter Abbeel (Berkeley/Covariant): Applied RL

## Emerging Terms
- "Offline RL": Learning from fixed datasets
- "Conservative Q-Learning (CQL)": Pessimistic value estimation
- "Return-Conditioned Policies": Transformer-based RL
- "Direct Preference Optimization": RLHF without reward model

## Trend: RL → LLM Integration
RL techniques increasingly applied to LLM training (RLHF, DPO)
LLM techniques applied to RL (Decision Transformers)
→ Convergence happening

## L4 Tokens for Prompting
"Conservative Q-Learning", "implicit Q-learning", "decision transformer",
"return-conditioned", "DPO vs PPO", "Sergey Levine", "offline dataset"

Integration with Other Skills

From Frontier	To Skill	Use Case
Emerging concept found	trace	Trace its (short) lineage
Gap in understanding	domain-vocab	Learn foundational concepts
Deep interest in area	deep-dive	Full concept + paper treatment

Error Handling

Situation	Recovery
Domain too broad	Ask for sub-field focus
Too recent (no papers)	Expand timeframe, note as "pre-publication"
Mostly preprints	Note peer-review status, include anyway
Field moving fast	Recommend re-running monthly