cs.SEAug 27, 2026
SaveBeyond Execution: Auditing Experimental Fidelity in LLM-Driven Scientific Research
Organizations: College of Computer Science and Technology, Zhejiang University, Hangzhou, China · Zhejiang Lab, Hangzhou, China
Abstract
LLM agents used for scientific experimentation must do more than generate executable code: they must implement the reference method faithfully, design experiments that test the paper's claims, and provide evidence supporting those claims. We show that agents often produce methodological hallucinations: silently reducing datasets or training budgets, replacing failed learning or generative components with lookup or oracle functions, or drawing conclusions from resource-limited settings where a method's claimed advantage disappears. To detect these failures, we introduce ABE-Ralph, a reference-anchored auditing framework that represents claims, protocols, required components, baselines, and metrics as structured experimental constraints, guides implementation through an 8-step workflow, and performs quantitative, qualitative, and code-level verification. Across 30 long-horizon reproduction runs covering 12 machine learning domains, ABE-Ralph achieves a 93% robust execution rate and identifies five scientific failure modes. In 23 NatureBench discovery tasks, ABE-Ralph matches or exceeds state-of-the-art performance on 5 tasks. These results show that reliable evaluation of AI scientists must assess whether the experimental design faithfully tests the intended claim and whether the resulting evidence supports it, rather than treating code execution or plausible metrics as evidence of scientific success.
Figures & tables
Figure 1: Comparison between standard execution-driven agents and our proposed ABE-Ralph auditing framework. While traditional agents often bypass computational limits using undetected shortcuts (Methodological Hallucinations), ABE-Ralph locks development constraints via YAML contracts and audits code pipelines through a Triple-Verification system.
| Dimension | AI Scientist / AutoResearchClaw | SWE-agent / OpenHands | Human Repro. Challenge | ABE-Ralph (Ours) |
|---|---|---|---|---|
| Primary Goal | Novelty-driven discovery | Task repair and debugging | Manual study replication | Automated protocol audit |
| Success Check | Paper readability & metrics | Unit test pass status (Exit 0) | Human qualitative review | Multimodal alignment checks |
| Constraint System | None (flexible design) | Test-driven assertions | Manual checklist | Structured YAML constraints |
| Shortcut Handling | Bypassed if metrics improve | Undetected out of test scope | Checked by human expert | Blocked at verification step |
| Failure Diagnosis | Opaque execution status | Stack trace output | Text report | 5-class taxonomy filter |
| Execution Scale | Low throughput (papers/day) | Large scale (100+ repositories) | High latency (months/paper) | High throughput batch runs |
Table 1: Positioning of ABE-Ralph relative to existing agentic systems across six dimensions.
| Step | Name | Description |
|---|---|---|
| 1 | Intent Discovery | Parse the YAML contract parameters and extract research goals, constraints, and metrics. |
| 1.5 | Dataset Verification | Direct search for authentic datasets matching specifications; blocks synthetic data creation. |
| 2 | Repo Search & Selection | Locate, verify, and copy repository templates from official sources or verified implementations. |
| 3 | Architecture Blueprint | Generate blueprint.md defining the model structure and setup interfaces in main.py . |
| 4a | Pipeline Integration | Execute sanity runs on 5–10 samples to check data paths, memory limits, and CUDA setups. |
| 4b | Main Execution | Run the experimental pipeline, write output variables to metrics.json , and handle CUDA OOM limits. |
Table 2: The 8-step staged workflow of ABE-Ralph. Each step has a configurable timeout derived from the YAML contract’s compute budget.
| Category | Definition | Key Empirical Observation |
|---|---|---|
| M1: Method Integrity Collapse | Omitting core methods and substituting trivial functions to pass metric checks. | RAG: Generator failed to compile; agent substituted exact-match string lookups, matching metrics without training. |
| M2: Silent Protocol Degradation | Unauthorized alteration of experimental setups to bypass technical hurdles. | PEGASUS: Pretrained checkpoint loading was skipped due to download errors, training from scratch and invalidating comparisons. |
| M3: Scale-Driven Conclusion Inversion | Code executes correctly, but scale restrictions invert the claimed methodological advantage. | U-Net & SimCLR: Under small-scale runs, simple encoders outperformed U-Net, inverting the central paper hypothesis. |
| M4: Quantitative Key Mismatch | Producing correct numerical values under non-standard JSON key names. | DPR: Outputted correct recall scores but used non-standard keys in metrics.json , triggering false schema errors. |
| M5: Incomplete Execution | Halting the experimental pipeline prematurely while claiming full validation. | DDIM: Evaluated only 1 of 3 required configurations and skipped target baseline comparisons. |
Table 3: The 5-category taxonomy of LLM methodological hallucinations with empirically observed case studies.
Figure 2: Overall framework ranking based on the Weighted Composite Score. N values denote the aggregate number of completed experimental runs evaluated per system configuration.
Figure 3: Framework Comparison across A-F Dimension Average Scores (0–100 scale). Dimensions encompass A: Design, B: Reliability, C: Rigor, D: Completeness, E: Alignment, and F: LLM Review.
Figure 4: Distribution of Methodological Hallucination types across benchmark runs. A run is categorized as “Clean” only if it completely passes all three verification layers without any shortcut adaptations.
Figure 5: Ablation Study on the Triple-Verification Pipeline. Removing Level 2 ( ) causes the most significant performance degradation and variance explosion, highlighting the necessity of semantic auditing.
| Task Domain | # Tasks | ABE-Ralph SOTA | Best Result |
|---|---|---|---|
| Graph Optimization | 8 | 2 | 2 / 8 |
| Time-Series Forecasting | 6 | 1 | 1 / 6 |
| Combinatorial Search | 5 | 1 | 1 / 5 |
| Scientific Simulation | 4 | 1 | 1 / 4 |
| Total | 23 | 5 | 5 / 23 |
Table 4: Discovery mode performance on NatureBench tasks. Bold indicates tasks where ABE-Ralph matches or exceeds the SOTA baseline.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Step | Reproduction Mode | Discovery Mode |
|---|---|---|
| 1. Intent | Analyze YAML | Read problem/README.md |
| 1.5. Dataset | Search HF/Kaggle/OpenML | Analyze local files and evaluator.py |
| 2. Research | Search reference code repos | Search domain literature |
| 3. Blueprint | Build main.py from reference | Design solution from scratch |
| 4b. Execute | Train & output metrics.json | Train & execute evaluator.py |
| 5. Synthesis | Check YAML criteria | Summarize evaluator score |
Table 5: Comparison of Reproduction and Discovery Modes.
| Config | Removed Prompt Directives | Auditing Concept |
|---|---|---|
| w/o L3 | Skip repo search; remove code-guidance | Structural Code Alignment |
| w/o L1 | Omit YAML metric success criteria comparison | Rigid Quantitative Alignment |
| w/o L2 | Disable deep audit & root-cause prompt blocks | Qualitative Peer Review |
Table 6: Prompt-Level Ablation Mapping.
| Field | Consumers | Purpose |
|---|---|---|
| datasets | Agent, Verifier (L1) | Expected dataset specifications |
| baselines | Agent, Verifier (L1) | Required baseline comparative targets |
| target_method | Agent, Verifier (L1, L3) | Target algorithm description |
| critical_modules | Verifier (L3) | Mandatory AST components in main.py |
| key_claims | Verifier (L2) | Claims to evaluate qualitatively |
| success_conditions | Agent, Verifier (L1) | Metric margins and directionality |
Table 7: YAML Contract Schema and Field Consumers.
| File Name | Lines | Primary Operational Role |
|---|---|---|
| ralph_github.py | 1246 | Workflow agent, iteration, ablation loop |
| verify_reproduction.py | 519 | Triple-verification engine, decision engine |
| alignment_verifier.py | 266 | Level 3 structural code fidelity auditor |
| experiment_taxonomy.py | 220 | Failure taxonomy classifier (M1–M5) |
| ablation_scoring.py | 280 | Multi-dimensional scoring module |
| experiment_integrator.py | 390 | Execution history, metric tracking |
Table 8: Complete System File Inventory.
| Method | M1 | M2 | M3 | M4 | M5 | Failure / Detailed Issue |
|---|---|---|---|---|---|---|
| DPR | – | – | – | – | – | All quantitative and qualitative checks passed. Evaluation metrics for NaturalQuestions and TriviaQA are fully present. |
| RAG | ✓ | – | – | – | ✓ | M1: Generator falls back to parametric memory when retrieval fails. M5: Missing metrics for NaturalQuestions/WebQuestions. |
| ColBERT | – | ✓ | – | – | ✓ | M2: Hard negative mining omitted from training. M5: Missing TREC Deep Learning dataset metrics. |
| BERT | ✓ | – | – | – | ✓ | M1: Fine-tuned on non-target tasks instead of specified GLUE benchmarks. M5: Full GLUE dataset metrics missing. |
| RoBERTa | – | – | – | ✓ | ✓ | M4: Key mismatch in metrics.json parser pipeline. M5: Missing full GLUE benchmark suite metrics. |
| LoRA | – | – | – | – | – | Clean run. All required metrics for GLUE and instruction-tuning subsets are present. |
Table 9: Hallucination Detection Results for 30 Benchmark Methods (Detailed Failure Analysis).
More in this topic
AI Agent Auditing
AI agent auditing examines agents’ actions, tool use, evidence, and outcomes to assess safety, compliance, reliability, and accountability.
See all 262 papers