Nova Idea and Hypotheses Generation Assistant
Faithful implementation of the nova-idea-gen SKILL — multi-stage iterative planning for novel research ideas. Model: Claude Opus 4.6 (PoE API)
NOVA Framework — Conference slides
From: NOVA_conference_slides_EN.pptx
Slide 1
NOVA: Iterative Planning for Scientific Discovery
A framework for generating novel research ideas with Large Language Models
Slide 2
The Paradox of LLM Creativity
LLMs can generate research ideas but often produce repetitive suggestions. True innovation requires diverse knowledge sources. External knowledge acquisition is key to breakthrough ideas.
Slide 3
NOVA Approach
Strategic planning and search for knowledge. Iterative refinement of research ideas. Integration of insights from multiple disciplines.
Slide 4
NOVA Pipeline
Seed Idea Generation Iterative Planning & Search Final Proposal Generation
Slide 5
Idea Growth Through Iteration
Start with 15 seed ideas. After iteration: 45 refined ideas. After deeper iteration: 135 enriched ideas.
Slide 6
Results: 3.4× More Unique Ideas
Slide 7
Results: 2.5× More Top-Rated Ideas
Slide 8
NOVA Architecture
LLM Core (GPT-style or other models) Agent-based orchestration layer Academic search engine with embeddings Planning, search, and ideation agents
Slide 9
NOVA Innovation Flywheel
Seed → Plan → Search → Iterate → Propose A continuous cycle expanding the idea space. Transforms LLMs into active research assistants.
Slide 10
Impact
More novel research directions. Better exploration of interdisciplinary knowledge. Structured pipeline for scientific idea generation.
🧠 NOVA 2.0 — User Manual
AI System for Scientific Discovery & Research Idea Generation
This manual explains how to use the NOVA app to generate novel scientific research ideas, hypotheses, and full research proposals from papers or research topics.
The system implements an enhanced NOVA pipeline, which generates ideas through iterative search, refinement, and synthesis across multiple scientific discovery strategies.
1. What NOVA Does
NOVA is designed to help researchers:
- discover new research directions
- generate testable hypotheses
- create full research proposals
- identify paradigm-shifting ideas
- map multi-paper research programs
The system is especially useful for:
- ML / AI research
- computational science
- interdisciplinary research
- research brainstorming
- proposal writing
- literature exploration
2. System Overview
The app runs a multi-stage scientific discovery pipeline.
NOVA 2.0 pipeline
Input Topic / Paper
↓
Stage 0 — Deep Paper Analysis
↓
Stage 1 — Seed Idea Generation
↓
Stage 2 — Iterative Knowledge Search
↓
Stage 2.5 — Paradigm Synthesis
↓
Stage 3 — Red-Team Scientific Stress Testing
↓
Stage 4 — Ranking & Selection
↓
Stage 5 — Proposal Generation
↓
Stage 6 — Research Program Mapping
Each stage progressively improves ideas using scientific discovery methods, literature search, and reasoning.
3. Getting Started
Step 1 — Open the NOVA app
Navigate to your NOVA app. You will see the Input panel.
Step 2 — Enter API Keys
Two API keys are required.
1️⃣ POE API KEY
Used for:
- Claude-Opus LLM reasoning
- idea generation
- proposal writing
- scientific reasoning
Enter your POE API key in the field POE API Key.
2️⃣ BRAVE SEARCH API KEY
Used for:
- real web search
- finding relevant papers
- retrieving research knowledge
Enter your Brave Search API key in Brave Search API Key.
4. Input Options
You can start NOVA in three different modes.
Mode 1 — Paper Analysis Mode (Recommended)
Provide:
- Title
- Abstract
- Optional references
Example:
Title: Isotropic Gaussian Representation Learning
Abstract: This paper proposes...
NOVA will:
- analyze the paper
- identify limitations
- generate new research ideas building on it
This is the best mode for scientific discovery.
Mode 2 — Research Topic Mode
Provide a topic instead of a paper.
Example: Topic: World models for robotics or Topic: Self-supervised learning for multimodal embeddings
NOVA will synthesize a virtual seed paper and generate ideas.
Mode 3 — Idea Expansion Mode
Provide an existing idea. Example: Idea: Use topology-aware embeddings to detect collapse in self-supervised learning.
NOVA will expand, refine, generate experiments, and produce proposals.
5. Advanced Settings
Optional configuration parameters:
Max output tokens — Controls length of generated output. Default: 24576.
Seed ideas per method (Stage 1) — Number of ideas per discovery method. Default: 5.
Stage 2 iterations — Controls idea refinement depth. Recommended: 1 (fast) to 3 (full pipeline).
Number of full proposals (Stage 3) — Top ideas expanded into proposals. Recommended: 3–5.
Use live web search in Stage 2 — Enable Brave Search for idea-specific retrieval. Recommended: ON.
Use live web search in Stage 3 — Enable search for datasets/baselines in proposals. Recommended: ON.
Target venues — Optional. Example: NeurIPS, ICLR, CVPR — biases ideas toward those communities.
Temperature — 0 = repeatable, 1 = more creative.
6. Running NOVA
Click Run Nova Pipeline.
The pipeline will begin executing. Progress appears in the live log.
Example stages:
- Running Stage 1 only (paper analysis + seed ideas)
- Running Brave Search per idea
- Running Stage 2 + Stage 3
Total runtime: typically 2–6 minutes (depending on LLM latency and web search).
7. Understanding the Output Tabs
NOVA Explanation — Framework overview; slide content if NOVA_conference_slides_EN.pptx is provided.
Raw output — Full Nova results text.
Formatted preview — Markdown rendering of results.
Full prompt used — System and user prompt sent to the API (for reproducibility).
Stage outputs — Separate windows for Stage 1 (paper analysis + seed ideas), Stage 2 (iterative planning & search + visited URLs), Stage 3 (proposals & scorecard).
Download — Markdown and PDF export of the full report.
8. Best Practices for Scientific Discovery
- Start with strong papers — Best results with recent top-conference papers, emerging topics, or controversial results.
- Run multiple times — Different runs produce different ideas. Try 3–5 runs to explore the idea space.
- Use live web search — Enable Stage 2 and Stage 3 web search for more grounded, diverse ideas.
- Combine results manually — Sometimes breakthroughs appear by combining ideas from different runs.
9. Example Workflow
Researcher studying representation learning
Input paper: Title: Isotropic Gaussian Representation Learning
NOVA might generate ideas such as:
- topology-aware collapse detection
- geometry-adaptive embeddings
- distribution-aware latent spaces
These ideas could then become research projects.
10. Troubleshooting
Problem: No output — Check API keys and internet connection.
Problem: Slow execution — Reduce Stage 2 iterations or max tokens.
Problem: Ideas look similar — Increase temperature slightly or ensure web search is ON for diversity.
Problem: Parse/API errors — Check PoE and Brave API keys; ensure input is not empty.
11. Security Advice
Never share your API keys publicly. Do not commit keys to version control. For shared deployments, use environment variables or platform secrets (e.g. Hugging Face Space Secrets) instead of entering keys manually in the UI.
12. Intended Users
NOVA is designed for:
- academic researchers
- PhD students
- AI researchers
- R&D teams
- research labs
- innovation groups
13. Summary
NOVA 2.0 is an AI system for structured scientific discovery. It combines literature search, scientific reasoning, idea synthesis, and proposal generation to help researchers discover novel research directions and paradigms.
Run the pipeline to see formatted results.