This benchmark measures how effectively various LLMs can infer a narrow or specific "theme" (category/rule) from a small set of examples and anti-examples, then detect which item truly fits that theme among a collection of misleading candidates. The overall process involves generating themes, creating examples and anti-examples, filtering out low-quality data via a "double-check" step, and finally prompting LLMs to score the real example among several distractors.
This bar chart displays, for each model, the average rank that model assigns to the true example (when placed among seven distractors). Ranks range from 1 (top score) to 8 (lowest).
- Smaller bars indicate better performance, because it means the correct example is consistently placed near the top.
- A bar height of 2.0 would mean that on average, the leftover correct item was the second-highest-scored candidate.
A more granular view of the ranks each model assigns to the leftover correct example per file, showing how stable or varied those ranks are across different themes.
A correlation matrix based on how similarly two models assign a “difference score” to the correct vs. anti-examples. It highlights which LLMs behave similarly or deviate significantly.
A stacked bar chart indicating how frequently each model places the real leftover example strictly at the top (or tied for top). This quickly shows which LLMs are best at ensuring the real item is #1 vs. merely near the top.
Rank | Model | Avg Rank | Skipped/Total |
---|---|---|---|
1 | o1 | 1.80 | 0/810 |
2 | Gemini 2.0 Flash Thinking Exp | 1.90 | 0/810 |
3 | Claude 3.5 Sonnet 2024-10-22 | 1.93 | 0/810 |
4 | o1-mini | 1.95 | 0/810 |
5 | GPT-4o | 1.96 | 0/810 |
6 | Gemini 2.0 Flash Exp | 2.00 | 0/810 |
7 | DeepSeek-V3 | 2.03 | 0/810 |
8 | Qwen QwQ * | 2.05 | 280/810 |
9 | Llama 3.1 405B | 2.08 | 0/810 |
10 | Mistral Large 2 | 2.11 | 0/810 |
11 | Llama 3.3 70B | 2.12 | 0/810 |
12 | Gemini 1.5 Pro (Sept) | 2.13 | 0/810 |
13 | Gemini 1.5 Flash | 2.13 | 0/810 |
14 | Grok 2 12-12 | 2.21 | 0/810 |
15 | Qwen 2.5 72B | 2.21 | 0/810 |
16 | Claude 3.5 Haiku | 2.25 | 0/810 |
17 | GPT-4o mini | 2.30 | 0/810 |
18 | Gemma 2 27B | 2.60 | 0/810 |
- Avg Rank is the mean ranking assigned to the correct example across 810 test files.
- Skipped indicates how many outputs failed to parse or didn’t follow the required output format (e.g., missing and tags).
-
Theme & Example Creation
We prompt high-quality LLMs (Claude 3.5 Sonnet, Grok 2, Gemini 1.5 Pro, GPT-4o, DeepSeek-V3) to generate 2,000 unique, succinct “themes” that foucs on a narrow concept. Each is based on a random trio of starting points, ensuring novelty.
-
Gather Examples & Anti-Examples
For each theme, we then request four
<example>
entries that specifically fit it, plus 20<anti_example>
entries that could belong to a broader or partially overlapping category but do not fit the exact theme. -
Quality Check (“Double Check”)
- We create specialized prompts that ask LLMs to score how well each of the four real examples (#1–4) matches the theme, and how well each of the twenty anti-examples (#5–24) fits the notion of being “broader or related but not the theme.”
- We parse these 24 numeric scores (per file, per LLM), computing standardized z-scores. If a real example scores poorly (z < -2.5) or if the top anti-examples fail to show sufficiently high “anti-example” scores, that file is discarded.
- From the initial 2,000 sets, we end up retaining 810 sets (themes + examples + anti-examples).
-
Final “Pick” Challenge
- From each of the 810 validated sets, the final prompt includes 3 real examples + 3 anti-examples as context.
- The fourth real example is hidden among 7 top "misleading" anti-examples (8 total)
- We then prompt 18 different LLMs to assign a 0–10 score to each of these 8 candidates. A perfect approach would always rank the correct example #1.
-
Result Analysis
- If a model consistently places the real leftover example at or near the top, it implies strong thematic generalization.
- We compile the results into multiple stats, including average rank, difference vs. the anti-example average, fraction of times the real item is top, etc.
Examples: mathematical models, decision trees, flowcharts
Anti-examples: diagrams, maps, blueprints
Candidates:
- checklists
- spreadsheets
- weather forecasts
- mind mapping <- correct pick
- road signs
- instruction manuals
- summaries
- outlines
Theme: "Concepts or systems that involve solving complex problems through simplification or abstraction"
Examples: clay pot, bamboo sieve, calabash gourd spoon
Anti-examples: plastic serving spoon, rubber spatula, plastic strainer
Candidates:
- cast iron skillet
- wooden mortar <- correct pick
- nylon cooking utensils
- ceramic bowl
- stone grinding wheel
- porcelain plate
- silicone baking mat
- bamboo steamer basket
Theme: "Tools or implements traditionally used in West African food preparation that are made primarily from a single, naturally occurring material."
- Also check out the LLM Step Game, LLM Creative Story-Writing Benchmark, LLM Confabulation/Hallucination Benchmark, LLM Deception Benchmark, NYT Connections Benchmark, and LLM Divergent Thinking Creativity Benchmark.
- Follow @lechmazur on X (Twitter) for other upcoming benchmarks and more.