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"

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..."

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"

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"

## 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

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?"

# [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}

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

# 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"