Easily extendable 100% local multi-agent system for generating novel research hypotheses, abstracts, and references.
All powered by local Ollama LLMs. No API keys. No cloud. Just you, your GPU/CPU, and public sources.
- Multi-agent pipeline: breakdown, critique, synthesize, innovate, and polish
- Pulls from public sources: arXiv, Semantic Scholar, EuropePMC, Crossref, DOAJ, bioRxiv, medRxiv, OpenAlex, PubMed
- Scores, ranks, and summarizes literature
- Uses Ollama for summarization and novelty checks
- Final output is a clean, human-readable panel with stats / insights
────────────────────────────────────────────── Pipeline 'Research Hypothesis Generation' Finished in 102.67s ───────────────────────────────────────────────
────────────────────────────────────────────────────────────────── Final Results Summary ───────────────────────────────────────────────────────────────────
╭────────────────────────────────────────────────────────────── Final Hypothesis Structured ───────────────────────────────────────────────────────────────╮
│ This research introduces a novel approach to Large Language Model (LLM) compression predicated on Neuro-Symbolic Contextual Compression. We propose a │
│ system that translates LLM attention maps into a discrete, graph-based representation, subsequently employing a learned graph pruning algorithm to │
│ remove irrelevant nodes while preserving critical semantic relationships. Unlike existing compression methods focused on direct neural manipulation, │
│ this approach leverages the established techniques of graph pruning, offering potentially significant gains in model size and efficiency. The │
│ integration of learned pruning, adapting to specific task and input characteristics, represents a fundamentally new paradigm for LLM compression, moving │
│ beyond purely neural optimizations. │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─────────────────────────────────────────────────────────────────── Novelty Assessment ───────────────────────────────────────────────────────────────────╮
│ │
│ │
│ **Novelty Score: 7/10** │
│ │
│ **Reasoning:** │
│ │
│ This hypothesis demonstrates a moderate level of novelty, primarily due to the specific combination of techniques and the integration of neuro-symbolic │
│ approaches. Let's break down the assessment: │
│ │
│ * **Elements of Novelty (Strengths):** │
│ * **Neuro-Symbolic Contextual Compression:** The core idea of translating LLM attention maps into a discrete, graph-based representation *is* a │
│ relatively new area of exploration. While graph pruning exists, applying it specifically to the output of LLM attention maps – and framing it within a │
│ neuro-symbolic context – is a distinctive aspect. │
│ * **Learned Graph Pruning:** The explicit mention of a *learned* graph pruning algorithm elevates the novelty. Many pruning methods are static, │
│ whereas learning the pruning criteria based on task and input characteristics is a significant step forward. │
│ * **Integration of Graph Pruning with LLMs:** While graph pruning is used in other domains, its application to LLMs, particularly in this way, is │
│ not widely established. │
│ │
│ * **Elements Limiting Novelty (Weaknesses):** │
│ * **Graph Pruning is Not Entirely New:** As highlighted in Paper 1, graph pruning techniques exist in general. The core concept of pruning nodes │
│ based on importance is well-established. │
│ * **Related Work Exists:** Several papers (Papers 2, 3, 4, 5, 6, 7) address aspects of model compression, including quantization, sparsity, and │
│ dynamic budgets. While the *combination* is novel, the individual components are not. Paper 7's "thinking step-by-step compression" is particularly │
│ relevant, even though it uses a different framing (dynamic compression of reasoning steps). │
│ * **Fine-grained vs. Coarse-grained:** The hypothesis positions itself against "coarse-grained" methods (Paper 1). However, many current compression │
│ techniques are moving towards finer-grained approaches. │
│ │
│ │
│ **Justification for the Score:** │
│ │
│ A score of 7 reflects that the hypothesis presents a novel *approach* rather than a completely new concept. The combination of learned graph pruning │
│ with attention maps represents a worthwhile exploration. However, it's not a revolutionary breakthrough because graph pruning itself isn’t entirely │
│ novel, and the field is already actively investigating various compression strategies. │
│ │
│ **Recommendations for Strengthening the Hypothesis:** │
│ │
│ * **Quantify the Expected Gains:** Adding specific claims about the expected reduction in model size and efficiency would strengthen the hypothesis. │
│ * **Elaborate on the "Neuro-Symbolic" Aspect:** Provide more detail on how the discrete graph representation represents the underlying semantic │
│ relationships within the LLM. │
│ * **Highlight the Advantage over Existing Methods:** Clearly articulate *why* this approach is expected to be superior to existing techniques (e.g., in │
│ terms of accuracy, speed, or ease of implementation). │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
- Clone this repo:
git clone https://github.com/tegridydev/abstract-agent cd abstract-agent - Install dependencies:
pip install -r requirements.txt
- Install Ollama and pull a model (e.g. gemma3:4b):
ollama pull gemma3:4b
- Run the agent:
python agent.py
- Agent A: Breaks down your topic into all the core pieces
- Agent B: Roasts the literature, finds gaps and trends
- Agent C: Synthesizes new directions
- Agent D: Goes wild, generates bold hypotheses
- Agent E: Polishes, references, and scores the final abstract
- Novelty Check: Checks if it's actually new or just recycled
- Final hypothesis, novelty score, references, and run stats (references searched/used, time taken)
- ollama
- rich
- arxiv
- requests
- xmltodict
- pydantic
- pyyaml
No API keys. All sources are public.
- Edit
agents_config.yamlto change the agent pipeline, prompts, or personas - Add new sources in
multi_source.py
MIT. Use it, fork it, break it, share it. Just give a shoutout to tegridydev if you want <3
Author: tegridydev
Repo: https://github.com/tegridydev/abstract-agent