Experiments or Outcomes? Probing Scientific Feasibility in Large Language Models

1. Problem Formulation

The paper addresses the gap in understanding how Large Language Models (LLMs) perform Scientific Feasibility Assessment. This task involves determining whether a scientific claim (hypothesis) is consistent with established knowledge and whether concrete experiments could support or refute it.

The authors frame this as a diagnostic reasoning task where a model must predict a binary label (FEASIBLE or INFEASIBLE) and provide a justification. The core challenge is isolating how different types of evidence—specifically experimental descriptions (methods/designs) versus experimental outcomes (results)—influence the model's judgment, and how robust these judgments are when evidence is incomplete.

2. Methodology

The authors propose a Controlled Knowledge Framework to systematically evaluate LLMs under varying evidence conditions.

A. Experimental Setup

• Models Evaluated: Five proprietary LLMs across different tiers and vendors: gpt-5.1, gpt-4o, Gemini-2.5-Pro, Gemini-2.5-Flash, and Grok-4.1-fast.
• Datasets:
  • MATTER-OF-FACT: Claims paired with source papers where evidence may support, contradict, or partially address the hypothesis.
  • REASONS: Scientific statements explicitly supported by cited literature (positive feasibility setting).
• Data Leakage Control: Applied post-cutoff filtering based on model knowledge cutoffs and paper publication dates to minimize pretraining leakage risks.

B. Controlled Evidence Conditions

The study evaluates models under four distinct context settings ($x$) alongside a hypothesis ($h$):

1. H (Hypothesis-only): $x = \emptyset$. Relies solely on parametric (internal) knowledge.
2. H+E (Hypothesis + Experiments): $x = E^*$. Provides experiment descriptions but no outcomes.
3. H+O (Hypothesis + Outcomes): $x = O^*$. Provides outcome summaries but no experimental procedures.
4. H+E+O (Hypothesis + Experiments + Outcomes): $x = (E^*, O^*)$. Provides complete context.

C. Stability and Robustness Analysis

To test robustness under incomplete information, the authors introduced a Partial Revelation Protocol:

• Parameters: $k_1$ (fraction of experiments revealed) and $k_2$ (fraction of outcomes revealed), tested at levels $\{0, 0.5, 1.0\}$.
• Sampling: For partial levels ($k < 1$), subsets are sampled uniformly at random without replacement. Results are averaged over 5 independent runs to account for variance.
• Metrics: Accuracy, Macro-F1, Matthews Correlation Coefficient (MCC), and ROUGE for explanation similarity. Crucially, they measure non-monotonic degradation (where performance drops as evidence increases) and below-baseline rates (where partial evidence performs worse than no evidence).

3. Key Contributions

1. Controlled-Knowledge Framework: A novel methodology to decouple the effects of experimental design descriptions from outcome results on feasibility judgments.
2. Stability Analysis: A rigorous quantification of model robustness, revealing that partial evidence often leads to brittle reasoning rather than graceful degradation.
3. Empirical Findings: Demonstrated that current LLMs exhibit specific failure modes when handling scientific evidence, challenging the assumption that "more context is always better."

4. Key Results and Findings

RQ1: Internal Knowledge Baseline

Under the Hypothesis-only (H) setting, models achieve non-trivial accuracy, indicating they possess coarse priors about scientific feasibility. However, these judgments are brittle and rely on generic plausibility rather than explicit constraints.

RQ2: Experiments vs. Outcomes

• Finding 1: Outcomes are more stabilizing than Experiments.
  • Providing Outcomes (H+O) consistently improves or preserves accuracy compared to the baseline.
  • Providing Experiments only (H+E) frequently degrades performance, sometimes dropping below the hypothesis-only baseline.
  • Interpretation: LLMs struggle to infer potential outcomes from experimental designs alone. Experiment descriptions introduce ambiguity that models fail to resolve, whereas outcome statements provide explicit directional constraints.
• Finding 2: More evidence is not strictly better.
  • The H+E+O condition does not consistently outperform H+O. In some cases, adding experiment descriptions to outcomes interferes with decision-making.

RQ3: Robustness Under Partial Information

• Finding 3: Non-monotonic and Adversarial Degradation.
  • Performance does not degrade smoothly as evidence is removed.
  • Partial evidence (e.g., 50% reveal) often performs worse than full evidence AND worse than no evidence.
  • This indicates that partial scientific context can actively mislead the model, causing it to over-commit to incorrect conclusions based on fragmented data.

Failure Mode Analysis

The authors identify three systematic limitations in current LLMs:

1. Surface Alignment over Evidential Validity: Models prioritize lexical/topical matches between hypotheses and experiment descriptions, ignoring critical mismatches in variables (e.g., dosage, endpoints).
2. Lack of Evidence Relevance Gating: Models fail to question whether provided evidence actually tests the hypothesis; they attempt to reconcile mismatched context rather than rejecting it.
3. Anchoring under Partial Information: Models anchor on available fragments and over-commit, failing to revert to uncertainty when evidence is incomplete.

5. Significance and Implications

• Diagnostic Value: The study reveals failure modes (brittleness, adversarial partial evidence) that are invisible in standard full-context evaluations or retrieval-augmented generation (RAG) benchmarks.
• Scientific Workflow Deployment: The findings suggest that simply retrieving more papers or experimental details does not guarantee better scientific reasoning. In fact, providing incomplete experimental descriptions can be harmful.
• Future Directions: Improving scientific feasibility assessment requires mechanisms that explicitly model evidence relevance, uncertainty, and variable alignment, rather than relying on the passive ingestion of additional context. The authors call for diagnostic evaluations before deploying LLMs in critical scientific workflows.

Conclusion: Current LLMs treat scientific feasibility as a surface-level classification task. While outcome evidence is highly beneficial, experimental descriptions without results often introduce fragility. Robust scientific reasoning in AI requires moving beyond simple context retrieval to structured evidence validation.