arXiv:2410.13185
10-Phase Pipeline
K=3 Chain Branches

CHAIN OF IDEAS
Research Idea Generator

Give it a research topic. Get back a novel, experiment-ready research idea — complete with paper chains, trend analysis, and next steps.

  • Builds 3 parallel paper-lineage chains from Recent Advances, Cross-Domain, and Foundational perspectives
  • 4-mode future direction prediction: Reflection, Analogy, Deep Dive, Imitate
  • ELO tournament across 5 axes: Novelty, Significance, Clarity, Feasibility, Effectiveness
  • Delivers a Research Brief with key entities, experiment plan, and concrete next steps

Pipeline at a Glance

Pipeline Phases

+ quality reviewer

10

Literature Branches

parallel chain builders

K=3

Future Prediction Modes

Reflection · Analogy · Deep Dive · Imitate

4

Idea Evaluation Axes

Novelty · Significance · Clarity · Feasibility · Effectiveness

5

Chain Depth

papers per branch

3–5

Paper Reference

DAMO-NLP-SG, 2024

Li et al.
10-Phase Pipeline

From Topic to Research Brief

Each phase builds on the last — chain construction informs trend analysis, which informs future prediction, which informs idea generation, which feeds the ELO selector.

1
1

Topic Scout

Map the research landscape, identify sub-fields and open problems, generate three distinct search perspectives (Recent Advances, Cross-Domain, Foundational)

2
2

Chain Builders (A/B/C)

Three parallel agents each build a 3–5 paper lineage chain from their assigned perspective using forward (citations) + backward (references) literature search

3
3

Trend Analyzer

Synthesize how each chain evolved paper-by-paper, identify the dominant research momentum, and surface cross-chain patterns and contradictions

4
4

Future Predictor

Forecast future directions using 4 modes: Reflection (limitations of latest work), Analogy (cross-domain transfer), Deep Dive (bottleneck attack), Imitate (replicate impact structure)

5
5

Idea Generator

Generate one research idea per chain branch — motivation, novelty statement, and method — then score each on 5 axes: Novelty, Significance, Clarity, Feasibility, Effectiveness

6
6

Novelty Checker

Verify each candidate idea against existing literature via search — confirm no published paper already covers the same motivation × method × improvement combination

7
7

Idea Selector (ELO)

Run a pairwise ELO tournament (A vs B, B vs C, A vs C) across all 5 Idea Arena axes to select the single best research idea from the three branches

8
8

Experiment Designer

Design a rigorous experimental plan: hypothesis, datasets/benchmarks, strong baselines, evaluation metrics, implementation steps, ablation studies, failure modes

9
9

Experiment Reviewer

Peer-review the experiment plan for clarity (unambiguous steps), feasibility (PhD-student executable), and technical quality (appropriate baselines and metrics)

10
10

Research Brief

Assemble the final Research Brief: chain narrative, trend analysis, winning idea, experiment outline, key technical entities, and concrete next steps for the next 1–3 months

Phase 2: Three Branch Perspectives — Parallel Chain Construction

🚀 Branch A — Recent Advances

Anchors to the most impactful work published in the last 2 years; traces the latest breakthrough directions forward and backward

🔀 Branch B — Cross-Domain Transfer

Searches adjacent fields for techniques applicable to this topic; builds a chain showing how borrowed methods evolved in their home domain

🏛️ Branch C — Foundational Depth

Traces seminal/theoretical papers; shows how the mathematical or conceptual core of the field was established and refined over time

K=3 branches · Chain length 3–5 papers · Forward + backward search · ELO tournament selects winning idea

Sample Output

What You Get

A complete Research Brief — not just a list of papers, but a narrative arc from the field's history to your novel contribution.

CHAIN OF IDEAS Research Brief
Sample Output

Topic: “Efficient fine-tuning of large language models with limited compute”

1. Chain Narrative

Chain A (Recent Advances): LoRA (2021) → QLoRA (2023) → GaLore (2024). Chain B (Cross-Domain): Low-rank matrix adaptation from signal processing → PEFT → Adapter layers. Chain C (Foundational): Full fine-tuning → Prompt tuning → Prefix tuning → LoRA.

2. Research Trend

All three chains converge on a single momentum: reducing the trainable parameter count while maintaining task-specific expressiveness. The bottleneck has shifted from memory to gradient quality — most recent work still suffers from suboptimal gradient signal in the low-rank subspace.

3. Future Directions

Mode 1 (Reflection): QLoRA acknowledges rank selection is heuristic. Mode 2 (Analogy): Compressed sensing theory uses adaptive basis selection — could dynamically choose LoRA rank per layer. Mode 3 (Deep Dive): Gradient subspace alignment through training is poorly understood. Mode 4 (Imitate): GaLore's full-rank gradient projection idea applied earlier in training.

4. Winning Idea (ELO)

Title: "Adaptive Rank LoRA via Compressed Sensing". Motivation: Fixed LoRA rank wastes capacity on saturated layers and under-fits dynamic layers. Novelty: Apply compressed sensing signal recovery to adaptively assign rank per layer per training step. Method: Monitor gradient energy distribution per layer; increase rank where energy spills outside subspace, decrease where it doesn't. Score: 7.8/10.

5. Experiment Outline

Baselines: LoRA (r=8,16,32), QLoRA, AdaLoRA. Datasets: GLUE, MMLU, GSM8K. Metrics: Task accuracy, trainable params (M), peak GPU memory (GB), convergence steps. Ablations: (1) fixed vs adaptive rank, (2) per-layer vs global threshold, (3) gradient energy window size.

6. Key Entities & Next Steps

Entities: LoRA, QLoRA, GaLore, PEFT, AdaLoRA, GLUE, MMLU, LLaMA-3. Next Steps: (1) Implement rank monitor on LLaMA-3-8B fine-tuning, (2) Run AdaLoRA baseline for comparison, (3) Submit to arXiv by end of month.

3 chains · 12 papers · 3 candidate ideas · 1 winner · ELO score 7.8/10
NOVEL + FEASIBLE
The Science

Why Chain of Ideas Works

Most AI research tools retrieve a flat list of related papers. CoI organises literature into a progressive chain — so the agent reads the field's developmental trajectory, not just its current snapshot.

Chain Structure

  • Papers ordered chronologically (earliest → latest)
  • Each paper explicitly builds on the previous one
  • Extracts: idea, technical contribution, entities, gap
  • Ablation shows chains outperform flat RAG by 18%

Trend Analysis

  • Paper-to-paper transition narratives
  • Cross-chain synthesis finds consistent momentum
  • Surfaces hidden technical bottlenecks
  • Enables "field velocity" forecasting

ELO Tournament

  • Pairwise comparison: A vs B, B vs C, A vs C
  • 5 Idea Arena axes (CoI paper's evaluation rubric)
  • 70.8% alignment with human expert judges
  • Eliminates subjective single-judge bias

Based on: Li et al. (2024), “Chain of Ideas: Revolutionizing Research via Novel Idea Development with LLM Agents” — arXiv:2410.13185. Original implementation: DAMO-NLP-SG/CoI-Agent. Adapted for Google ADK with GoogleSearchTool (no Semantic Scholar API required).

Comparison

How It Compares

Chain of Ideas vs. other research-ideation agents on the platform.

DimensionCHAIN OF IDEASSCIENCEVIRSCIFACET
Primary FocusResearch idea generationLiterature survey + hypothesisCollaborative ideation panelCross-domain facet recombination
InputResearch topic (text)Research question + domainResearch topic + team sizeResearch problem
Core MechanicProgressive paper chains + ELOLiterature scan + peer reviewMulti-agent team discussionFacet extraction + recombination
OutputResearch brief + experiment planResearch paper draftIdea with team consensusNovel cross-domain idea
Chain/StructureK=3 branches, 3-5 papers eachFlat literature listDiscussion roundsFacet matrix
PaperLi et al., arXiv:2410.13185ResearchAgent / SciAgentVirSci (AAAI 2025)Scideator (CHI 2024)
Ready to explore?

Start your research chain

Enter any research topic and get a complete Research Brief — paper chains, trend analysis, novel idea, experiment design, and next steps — in minutes.