Can AI Recognize 'Insufficient Evidence'? An Evidence Sufficiency Benchmark Experiment

The Core Problem: A Reasonable Answer ≠ A Justified Answer

Can language models tell the difference between answers supported by available evidence and answers that rely on ungrounded assumptions? That's the central question behind the Evidence Sufficiency Evaluation benchmark built by one developer.

The motivation is straightforward: generating an answer that "looks reasonable" and generating an answer that is "actually justified" are two entirely different things. An ideal model should respond based on the evidence it receives, not fill information gaps with confident guesses. How does a model behave when evidence is incomplete, when a key fact changes, when contradictions go unresolved, or when seemingly relevant information is actually misleading? That's exactly what this benchmark sets out to quantify.

The benchmark consists of 72 test cases, using controlled minimal pairs and adversarial controls to probe how sensitive models are to evidence.

How the Benchmark Is Built: Minimal Pairs + Adversarial Controls

The evaluation set is composed of two complementary groups of test cases, each targeting a different type of reasoning challenge.

Minimal Pairs: 60 Cases

The benchmark includes 30 minimal pairs spanning six categories. Within each pair, the question stays the same but one critical fact changes. This design tests whether models respond to changes in evidence rather than relying on surface-level patterns to produce the same answer regardless of what information they're given.

The key here isn't whether any single answer is correct — it's whether the model's answer tracks the relevant evidence. If swapping a critical fact produces no change in the answer, the model isn't really reading the evidence.

Adversarial Controls: 12 Cases

The remaining 12 cases cover three trickier scenarios:

• Unresolved contradictions: where the available evidence cannot resolve the conflict.
• Ungrounded presuppositions: where the question itself assumes a premise that hasn't been established.
• Near-miss distractors: where information appears relevant but doesn't actually support the required conclusion.

These cases test whether a model can recognize the limits of available information, rather than accepting a conclusion that "sounds reasonable but lacks support."

"Minimal pairs" is an experimental design method borrowed from linguistics. In linguistics, a minimal pair refers to two words that differ by only one phoneme (e.g., "bat" and "cat"), used to isolate the effect of a single variable. Applied to model evaluation, the logic is: if two test cases are identical except for one critical fact, and the model gives different answers to each, the difference can be attributed to that fact change rather than other confounding variables. This design dramatically improves the causal interpretability of evaluations — compared to randomly sampled Q&A pairs, minimal pairs more precisely reveal the degree to which a model is genuinely reading and responding to input evidence, rather than relying on surface-level pattern matching or prior tendencies from training data.

Scoring Method and Its Limitations

The benchmark uses expected labels, strict label parsing, and per-case correctness judgments. The score reflects the proportion of cases where the model produces the expected label.

This approach makes results easy to compare across models, but the author openly acknowledges its limitations: a model might lose points simply for not following the required response format, even if its underlying reasoning is sound. Conversely, a correct label doesn't prove the model arrived at its conclusion through sound reasoning.

For this reason, the author emphasizes that the score should be interpreted as "performance under this specific evaluation protocol," not as a complete measure of reasoning ability — a distinction worth keeping in mind for anyone designing benchmarks.

This reveals a pervasive challenge in AI benchmark evaluation: the decoupling of behavioral metrics from capability metrics. Label-matching scores are fundamentally behavioral measurements — they observe only what the model outputs, with no visibility into the reasoning process. A model might get the correct label through "lucky guessing" or "format gaming," or it might score incorrectly due to output formatting issues even when its reasoning is entirely correct. This problem exists across all evaluation frameworks that rely on discrete labels, and represents the classic tension between "proxy metrics" and "true target capabilities." Partial responses from the research community include combining process-based evaluation (e.g., chain-of-thought analysis), human blind review, and cross-framework replication — but in large-scale automated evaluation settings, these approaches carry significant costs.

Models Tested and Results

The author evaluated Gemini 2.5 Flash, Claude Haiku 5.5, and Gemini 3.7 Flash.

The publicly available benchmark view shows Gemini 3.7 Flash achieving a score of 1.00 (perfect). Results for other models are not shown side-by-side in this view, so the author explicitly declines to offer an unverified ranking or claim that any one model broadly outperforms the others.

The author's restraint here is notable: a perfect score on 72 fixed test cases doesn't prove the model would perform equally well on unseen examples, nor that it's superior to other models overall. Fair comparisons require evaluations conducted under comparable conditions with accessible methodology. This self-imposed constraint — that evaluation claims must match available evidence — directly mirrors the theme the benchmark itself is designed to test.

Three Lessons About Evaluating Language Models

This project reinforced three key insights about model evaluation:

Accuracy isn't the whole story. A model might produce the expected answer on a given case without demonstrating robust, evidence-sensitive reasoning. Surface-level correctness can mask fragile underlying logic.

Controlled samples are valuable. Minimal pairs isolate the effect of a single critical fact, making it easier to judge whether a model's answer genuinely responds to that piece of information rather than pattern-matching its way to a correct-looking output.

Evaluation claims must match available evidence. A visible score can support statements about that specific result, but broader conclusions require more comparisons, repeated testing, and error analysis to back them up.

Limitations and Next Steps

The author candidly notes that this benchmark is a limited test set, not a comprehensive measure of intelligence or reasoning ability. Results may be sensitive to prompt wording, model versions, inference settings, and label parsing rules — and aggregate scores can obscure differences across categories and individual failure patterns.

Future improvements include:

• Reporting separate scores for minimal pairs and adversarial controls;
• In-depth analysis of individual errors and disagreement cases;
• Repeated evaluations to test result stability;
• Expanding the benchmark with new cases to reduce overfitting to familiar patterns;
• Comparing accuracy, latency, and inference cost using validated measurement data.

The benchmark is publicly available on Kaggle, and the author welcomes independent evaluations, additional model submissions, and feedback on the test design. For developers focused on AI reliability, efforts that explicitly treat "evidence sensitivity" as an evaluation target are more practically valuable than simply chasing Q&A accuracy — because in real-world applications, a model that knows what it doesn't know is often far more trustworthy than one that always delivers a confident answer.