FuguReport

One Reflection Is Not Enough: Self-Correcting Autonomous Research via Multi-Hypothesis Failure Attribution

Authors Jie Ma, Binfei Chu, Jie Gao, Jinlu Zhang, Yiwei Ma, Yi Tan, Jiayi Ji, Xiaoshuai Sun, Rongrong Ji
Affiliations Xiamen University / Xiaohongshu / Hello Inc.
Categories Method / Failure Diagnosis / Structural causal diagnosis for recovery, Application / Autonomous Research / Self-correcting grounded experiments, Evaluation / Benchmarking / Cross-topic and domain performance evaluation
License CC BY 4.0

Abstract Overview

This paper presents SAGE, a self-correcting autonomous research system designed to recover from failed experiments more reliably than standard single-reflection agents. Its core mechanism, Multi-Hypothesis Failure Attribution (MHFA), frames recovery as structured causal diagnosis: the system generates multiple evidence-grounded explanations for failure, ranks them with an independent critic, and deterministically routes intervention at the hypothesis, design, or implementation level. SAGE also includes a grounded reporting mechanism that restricts numeric tables to values traceable to measured results and redacts unsupported numbers. On a 12-topic benchmark spanning five domains, the study evaluates both recovery behavior and end-to-end artifact quality, while also emphasizing that scientific writing quality and method-provenance grounding remain unresolved bottlenecks.

Novelty

The main novelty is treating autonomous experiment recovery as multi-hypothesis causal attribution rather than a single free-form reflection. The work also combines hierarchy-aware deterministic routing with explicit numeric grounding, making recovery decisions and reported results more auditable.

Results

On the 12-topic, 5-domain benchmark, SAGE increased metrics-bearing outputs from 42% (5/12) under the reflection baseline to 92% (11/12). In blind artifact-level evaluation, it scored 52.0 overall versus 48.2 for AI-Scientist-v2 and 24.8 for the host reflection baseline, with gains concentrated in code development and execution. Under AR-Eval, the canonical SAGE deliverables averaged 6.75/10 versus 5.00/10 for the baseline, although both systems remained below a strict main-conference-style paper quality bar.

Key Points

  1. SAGE replaces monolithic reflection with MHFA, which generates, scores, and routes among multiple evidence-grounded failure explanations.
  2. The system adds grounded reporting that sanitizes numeric tables against measured result registries, aiming to reduce fabricated or unsupported quantitative claims.
  3. Empirically, SAGE substantially improves recovery to metrics-bearing experiments and overall artifact quality, but the paper identifies prose-level method-provenance grounding and scientific writing quality as continuing limitations.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.