Grounded autonomous research - a fault-tolerant LLM pipeline from corpus to manuscript in frontier computational physics

Sep 21, 2026

Uploaded by Ivy

Autonomous-research agents have shown end-to-end LLM automation in calibration-friendly machine-learning sandboxes, but frontier physical science breaks the execution-as-calibration assumption. The work presents a fault-tolerant, grounded pipeline from a corpus of 11,083 condensed-matter physics arXiv papers to a publication-grade manuscript, using literature anchors and numerical confrontation during calibration. Across 47 fresh-context sessions and 2,162 literature-consultation events, redundancy, grounding isolation, and adversarial review catch failures and prevent unverified internal-prior outputs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages39

Reading Time1 h 38 min

AccessFree to download

Abstract

Introduction

Calibration breakdown in frontier physical science

Grounding defined via numerical confrontation

Methods and pipeline overview

Corpus-to-manuscript workflow

Fault tolerance via redundancy and review

Experiments and failure-mode ablations

Pre-architecture baseline and no-pilot ablation

Intervention and reproduction failures

Grounded autonomous research: a fault-tolerant LLM pipeline from corpus to manuscript in frontier computational physics Haonan Huang 1 arXiv :2607 .02329v 1 [ cs .AI] 2 Jul 2026 Abstract Autonomous-research agents have demonstrated end-to-end LLM automation in machine-learning sandboxes where execution provides calibration. Frontier physical science differs categorically: physical reasoning underlies every methodology choice, toolchains are often underdocumented, and calibration must come from external literature anchors—which unscaffolded agents cite but do not confront, hallucinating plausible, unverifiable results from internal priors. We present a pipeline that runs end-to-end from a corpus of 11,083 recent condensed-matter physics arXiv papers toa publication-grade manuscript with three substantive physics findings (here on altermagnetic piezomagnetism): the agent autonomously conceives a research direction by mapping the corpus, calibrates methodology by reproducing published references, conducts novel first-principles computations, and writes the manuscript—grounded in literature throughout, across 47 fresh-context sessions in six phases sharing only on-disk state, with 2,162 literature-consultation events. Fault tolerance emerges from redundancy: fresh-context isolation, distributed grounding, and adversarial review catch what any single session misses; preand post-pilot stages are fully autonomous, and pilot requires bounded human intervention only at reproduction failures—operational knowledge curation, not scientific direction. Two paired failure modes—a pre-architecture baseline and a nopilot ablation—isolate structurally enforced numerical confrontation at calibration checkpointsas the operative grounding mechanism. The primitives, characterized failure modes, and quantified intervention pattern lay a foundation for autonomous research in high-stakes scientific domains beyond computational physics. 1Department of Physics, Princeton University, Princeton, NJ, USA. Correspondence to: Haonan Huang . Accepted at ICML 2026 AI for Science Workshop. 1. Introduction AI-for-research has progressed from tool-assistance —search, summarization, code completion (Taylor et al., 2022 ; Chen et al., 2021 ; Bran et al., 2024)—to agentic automation that conducts research with diminishing human oversight (Lu et al., 2024 ; Yamada et al., 2025 ; Schmidgallet al., 2025 ; Miao et al., 2025) . End-to-end automated research has been demonstrated in domains where execution itself provides intrinsic calibration: machine-learning sandboxes where training-loss curves verify methodology (Lu et al., 2024 ; Yamada et al., 2025), defined computational tasks with established reference values (Miao et al., 2025), and pre-formulated benchmarks where ground truth is given. Within these calibration-friendly domains, agent reliability has nonetheless been critiqued (Si et al., 2024 ; Kambhampati et al., 2024 ; Beel et al., 2024): agents over-confidently produce plausible outputs, hallucinate from internal priors, and propagate methodology errors that internal-consistency checks do not catch. As the field pushes toward genuine frontier research —open direction selection from large literature corpora, novel parameter spaces, underdocumented toolchains where standard recipes do not exist—execution-as-calibration breaks down structurally. Frontier physical science is the canonical case: physical reasoning is required at every methodology choice, toolchains are often poorly documented, and calibration must come from external literature anchors—anchors that are, at the frontier, themselves scarce and provisional (§5)—because internal execution provides no ground truth. Without external anchor enforcement, agents default to internal-prior generation, producing confident plausible outputs that cannot be verified by execution alone. The same critiques leveled at simpler automated systems become d

FAQs

What problem does the pipeline address in frontier physical science research?

Execution-based calibration fails because physical methodology lacks intrinsic ground truth, so agents can generate plausible but unverified results from internal priors. This blocks reliable manuscript-level conclusions without stronger literature grounding.

How does the pipeline operationalize “grounding in literature”?

It enforces numerical comparison at calibration checkpoints, requiring agent-computed values to be confronted with published reference values rather than relying on citation availability alone.

What role does the pilot stage play in fault-tolerant grounding?

Paired failure modes isolate the pilot stage as the architectural element that operationalizes grounded-throughout. A no-pilot ablation skips the numerical confrontation step, weakening calibration.

Sep 21, 2026

Uploaded by Ivy