When AI Reviews AI: A Case Study in Benchmark Contamination

Date: December 19, 2025
Method: UKE_G Recursive Triangulation
Target: “Evaluating Large Language Models in Scientific Discovery” (SDE Benchmark)

Two days ago, a new benchmark paper dropped claiming to evaluate how well large language models perform at scientific discovery. The paper introduced SDE (Scientific Discovery Evaluation)—a two-tier benchmark spanning biology, chemistry, materials science, and physics. Models were tested on 1,125 expert-vetted questions and then asked to autonomously run research projects: propose hypotheses, execute simulations, interpret results.

The headline findings seemed solid. State-of-the-art models scored 0.84-0.86 on general science Q&A but only 0.60-0.75 on SDE’s discovery-grounded questions. A 10-25 point gap. The authors’ claim: conventional benchmarks miss what matters for actual scientific work.

I ran the paper through UKE_G—a structured analysis protocol I use for high-stakes technical review. The first pass raised some concerns: weak evidence for claimed scaling plateaus, model performance varying wildly across scenarios, unclear causality in the project-level results. But nothing that felt like a dealbreaker. The methodology seemed rigorous enough given the acknowledged computational constraints.

Then I tried what I’m calling Staged Adversarial Review.

The Technique

Staged Adversarial Review is a four-phase method for forcing perspective-taking when single-pass analysis risks model collapse:

Anchor: Analyze the document with standard protocols
Injection: Get independent analysis from external sources (different AI models, human experts)—critically, without showing them your Phase 1 results
Reconciliation: Compare the analyses, map disagreements, maintain tension rather than synthesizing
Deep Probe: Re-analyze assuming all previous reviewers share systematic blind spots

Operational notes: Phase 2 used identical prompts across models to get uncorrelated error modes. Phases were separated by ~24 hours to reduce anchoring effects. Phase 4 used explicit adversarial framing: “Both previous reviews accepted X at face value—what did they miss?”

The key insight: a model reviewing its own analysis tends to reinforce its initial framing. Introducing uncorrelated external perspectives—then explicitly prompting for “what did everyone miss”—breaks that convergence.

What Phase 1 Found

Standard UKE_G analysis identified the obvious issues:

Scenario dependency: Same model scored 0.23 on NMR structure elucidation but 0.85 on retrosynthesis
Weak scaling evidence: Claims about reasoning plateaus relied on single-digit percentage differences
Project-scenario disconnect: Models succeeded at TMC optimization despite poor performance predicting TMC properties

Nothing damning. The paper openly acknowledged computational cost limitations. The expert panel represented real domain knowledge. It looked like a reasonable first attempt at a hard problem.

What Gemini Saw (Phase 2)

I fed the same paper to Gemini without showing it my analysis. Gemini caught different issues:

Statistical power: With ~20 questions per scenario, confidence intervals overlap completely for most model rankings
Causal confusion: Paper attributed shared failures to pre-training data without evidence—questions might just be objectively hard
Oracle validity: Some “ground truth” came from ML models trained on similar data distributions as the LLMs being evaluated

Good critiques. Orthogonal to mine. The oracle validity point was especially interesting—Gemini flagged the issue but didn’t name specific oracles or trace the circularity mechanism. Both reviews had caught concerns, but neither had pushed hard enough on the evaluation framework itself.

What Phase 4 Uncovered

Then came the adversarial probe: “Assume both previous reviews missed something systematic. Look specifically for issues invisible in the first two passes.”

That’s when the benchmark contamination risk surfaced—an issue neither Phase 1 nor Gemini’s analysis had questioned.

The Contamination Problem (Phase 4-Unique Finding)

The paper generates test questions from public datasets:

USPTO reaction database (standard in retrosynthesis training)
ZINC molecular repository
MatBench materials dataset
NIST spectroscopic data

Every one of these is likely in LLM training corpora. The paper’s defense: “Questions were templated from data, not directly copied.”

But that doesn’t address whether models saw the underlying molecules, reactions, or spectra during training. No deduplication checks. No temporal cutoff enforcement (e.g., “only data published after each model’s training date”). No perplexity-based contamination testing.

Critical gap: The paper evaluates models with different training cutoffs—GPT-4o (2023), Claude Opus 4.1 (2024), GPT-5 (2025)—but doesn’t verify which test molecules are temporally outside each model’s training window.

The Oracle Circularity Problem (Gemini Flagged, Phase 4 Sharpened)

Gemini’s oracle validity concern turned out to be more serious than initially apparent. Phase 4 added specificity by naming the actual oracles and tracing the circularity mechanism:

The paper uses ML models as “ground truth”:

CHGNet for materials property prediction (neural network trained on Materials Project database)
GFN2-xTB for geometry optimization (force field approximation)
molSimplify for transition metal complex generation (heuristic-based)

When you evaluate LLMs against ML oracles trained on similar data distributions, you’re not measuring scientific accuracy—you’re measuring LLM-oracle agreement. This is circular reasoning dressed as validation.

The paper reports no oracle accuracy versus experimental measurements. No cross-oracle agreement checks. No oracle failure mode analysis.

The Temperature Confound

One more gap neither initial review caught: the GPT-5 API doesn’t permit temperature=0.7 (the setting used for other models). This makes cross-model comparisons on retrosynthesis tasks invalid by construction.

It explains a weird anomaly: older GPT-4o (60% solve rate) beats newer GPT-5 (53%). The paper attributes this to “validity checking failures” but doesn’t explain the mechanism. Temperature mismatch is the simpler explanation.

Why This Matters

None of these are individually fatal. Benchmark contamination is hard to detect. Oracle validity is genuinely difficult for computational domains. Temperature mismatches happen in API-constrained evaluation.

But together, they raise a more fundamental question: Is this benchmark measuring discovery capability or training data overlap plus ML-ML agreement?

The paper’s core claim—that there’s a 10-25 point gap between general science knowledge and discovery-relevant skills—depends entirely on the assumption that test questions are clean and oracles are valid. Phase 4 found reasons to doubt both.

The Serendipity Signal

The Staged Adversarial Review protocol includes a success criterion: “If Phase 4 produces an Omega variable that fundamentally alters the decision, the protocol has succeeded.”

In this case, the decision context is: Should the SDE benchmark be adopted as a field standard for evaluating LLM scientific discovery capabilities?

The contamination and oracle circularity issues don’t just tweak the interpretation—they undermine confidence in the benchmark’s core measurement claims:

Contamination risk: If test molecules are in training data, the claimed 10-25 point capability gap becomes unverifiable. We can’t distinguish “discovery ability” from “training data overlap.”
Oracle circularity: If ML oracles evaluate LLM outputs, project-level success may reflect oracle-LLM agreement rather than scientific validity.

These are structural flaws, not addressable weaknesses. They change the evaluation from “this benchmark shows LLMs have limited discovery capability” to “this benchmark’s measurement validity is unverified.”

That shift—from accepting measurement claims to questioning measurement validity—is what defines a successful Phase 4 outcome.

Mechanism Explanation

Why does this technique work?

Adversarial prompting: Phase 4 succeeds primarily because the prompt explicitly frames the task as “outsmart the previous reviewers” rather than “summarize the paper.” The model optimizes for finding gaps, not confirming conclusions. This is the core mechanism.

Perspective injection: Introducing uncorrelated external analysis (Phase 2) breaks the initial probability path by forcing comparison rather than reinforcement.

Lens saturation: Single passes rarely exhaust all critical perspectives. Phase 1 saturates content extraction. Phase 3 saturates perspective-taking. Phase 4 is reserved for edge detection and grounding verification—the dimensions most likely to harbor invisible assumptions.

Note on temporal separation: There was no temporal separation. These results with immediate re-prompting using adversarial stance. Completed within 20 minutes.

Cost-Benefit Tradeoff

Staged Adversarial Review triples token count and analysis time. It’s inappropriate for routine document review. But for high-stakes evaluation—where a single missed confound invalidates the entire analysis—the overhead is justified.

Decision heuristic for when to use this technique:

Document characteristics: Dense technical papers with methodological claims, benchmarks that could become field standards, binding agreements with irreversible consequences
Consequence threshold: If accepting the document’s claims at face value could lead to systematic misallocation of resources (research funding, model selection, policy decisions)
Red flags in Phase 1: Multiple Omega variables flagged, contradictory evidence that seems reconcilable but isn’t fully explained, foundational assumptions accepted without verification
External perspective availability: Can you access genuinely independent analysis (different models, domain experts, adversarial reviewers)?

For this paper, multiple criteria were met: it’s a proposed field-standard benchmark, Phase 1 flagged several Omegas (oracle validity, causality direction), and the contamination risk—if real—would systematically bias capability assessments across the entire AI research community.

Lessons for AI-Assisted Analysis

Model collapse is real: Without external contradiction, AI analysis reinforces its initial framing. The first interpretation becomes the anchor.
Comparative tension is productive: Introducing uncorrelated external perspectives doesn’t “confuse” the model—it breaks premature convergence.
Unknown unknowns require adversarial stance: Passive review asks “Is this claim supported?” Active review asks “What systematic blind spots do all reviewers share?”
Transcript techniques surface structure: The multi-phase transcript itself becomes a diagnostic instrument, exposing role drift and hidden assumptions.
Governance requires separation: The reviewer who generates analysis should not be the reviewer who validates it. Temporal separation (Phase 1 vs Phase 4) is a minimum requirement.

Final Assessment

The SDE paper represents serious effort by domain experts. The benchmark concept is valuable. The two-tier evaluation (scenario + project) captures something conventional benchmarks miss.

But the execution has architectural flaws that compromise the core claims:

Contamination risk makes the 10-25 point capability gap unverifiable
Oracle circularity makes project-level success ambiguous
Statistical power failures make model rankings unreliable
Temperature mismatches make cross-model comparisons invalid

These aren’t nitpicks. They’re foundational issues that only surfaced through adversarial multi-phase review.

The technique that found them—Staged Adversarial Review—proved its value. Not because it found “the truth” (benchmark evaluation remains genuinely hard), but because it systematically uncovered what passive review missed.

The technique isn’t truly “recursive” in the computer science sense—Phase 4 outputs don’t feed back into Phase 1 for re-analysis. It’s better described as staged: each phase exhausts a different critical lens, with Phase 4 reserved for adversarial questioning of foundations that earlier phases accepted.

That’s the signal: when Phase 4 fundamentally alters your decision context, you’ve justified the overhead.

Omegas Logged:

Ω: cost_benefit_threshold — The decision heuristic above is provisional. Does it reliably predict when Staged Adversarial Review yields high-value Phase 4 findings? Requires empirical validation across document types.

Ω: technique_generalization — Does this technique work beyond complex technical benchmarking papers? Testing needed across domains (legal contracts, policy documents, strategic plans) to establish boundary conditions.

Ω: tacit_knowledge_transfer — Can practitioners execute this technique from the documented procedure, or does effective application require undocumented expertise from iterative refinement?

Status: All open questions. Systematic testing required to operationalize decision criteria and validate cross-domain effectiveness.

This analysis used UKE_G v1.4 with Staged Adversarial Review extension. The technique is documented in the field notes and available under CC0-1.0 license. All claims about the SDE paper are derived from the published manuscript and supplementary materials.