Awesome Auto Research Hub Papers · Datasets · Projects
← Back to papers

Negative Knowledge as Failure-aware Shared Memory for AutoResearch

arXiv 2026 73.9 method

TLDR

Proposes a negative knowledge memory layer for AI research agents that records failures, improving performance on scientific tasks while using fewer tokens.

Reasoning

Strengths: Novel concept of structured negative knowledge, demonstrated on multiple tasks with transfer learning. Weaknesses: Limited to two settings, may need broader validation.

Read-first score

Read-first score 73.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 80.

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 25%
81

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Methodology quality 25%
70

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,experiment,result

Topical relevance 42%
66.7

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Field roles

FrontierBridgeMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 15.

Keyword Scores

automated research
10
AI for scientific research
10
automated scientific discovery
9
autonomous research agent
9
research automation
9
scientific discovery agent
9
AI scientist
8
automated experimentation
8
experiment design agent
7
literature review agent
1
survey generation
0
paper writing agent
0

Deep Analysis

Innovations

  • Proposes a negative knowledge memory layer where a curator agent converts failed attempts into bounded, typed records in a shared bank, and a downstream research agent explicitly adopts or rejects those records before proposing its next experiment.

Methodology

A curator agent converts each failed attempt into a bounded, typed record in a shared negative knowledge bank. A downstream research agent then explicitly adopts or rejects those records before proposing its next experiment. The system is evaluated on same-task retry (ScienceAgentBench) and cross-task scientific research on two nonlinear math-physics PDE problems.

Key Results

The negative knowledge layer outperforms vanilla AutoResearch baselines while using fewer tokens, and agents with the negative knowledge bank solve new PDE tasks that all baselines fail to solve. The negative knowledge bank also transfers to enhance AutoResearch on different PDE problems.

Tags

AICYLGMA