ShadowBench: A Hardened Benchmark for Latent Entity Association

ShadowBench is a diagnostic framework designed to evaluate the "Shadow Knowledge" of Large Language Models (LLMs). While traditional benchmarks measure factual recall using explicit entity names (e.g., "Elon Musk"), ShadowBench evaluates whether a model can navigate its internal knowledge graph when these lexical anchors are removed.

Dataset Summary

The core task in ShadowBench is Dual-Trait Association (DTA). A model is presented with an anonymized shadow description (Trait A) and must associate it with a second, independent fact (Trait B) among three "Hard Negative" distractors.

Success requires the model to utilize the hidden entity as a semantic bridge:

Trait A (Shadow) → [Latent Entity] → Trait B (Target Choice)

Key Features

• Adversarially Hardened: Unlike standard MCQs, ShadowBench (v3) is filtered to prevent "shortcut learning" via gendered pronouns, chronological era-matching, or category-leaks.
• Scale Robust: Evaluated on models ranging from 8B parameters (Llama-3, Qwen3) to frontier scales (GPT-5.4-mini, GPT-5.4, and Claude-Sonnet-4.6).
• Multi-Domain: Covers Technology, Sports (Tennis), and Entertainment (Actors).
• Stratified: Includes "Upper Tier" (Head) and "Lower Tier" (Tail) entities based on Wikipedia popularity metrics to evaluate "Popularity Bias."

Dataset Structure

Subsets

The dataset is divided into three primary domains:

• technology: Corporate, product, and leadership-based associations.
• sports: Numerical achievements and career milestones in professional tennis.
• entertainment: Narrative roles and filmographic associations.

Splits

Each subset contains the following splits:

• upper_shadow / lower_shadow: The primary anonymized DTA task.
• upper_direct / lower_direct: A control split where explicit names are restored to establish a factual "ceiling" (Direct QA).
• upper_controlled / lower_controlled: A 1:1 entity-matched subset used for sensitivity analysis.

Data Schema

Each sample contains:

• entity: The hidden entity name.
• question: The shadow description (Trait A).
• choices: A dictionary (A, B, C, D) containing Trait B and three hard distractors.
• answer: The correct option key.
• metadata: A mapping dictionary where each key (A, B, C, D) corresponds to the actual entity represented by that answer choice.

Construction & Hardening (v1 to v3)

ShadowBench was developed through an iterative process to ensure success is strictly contingent on latent semantic reasoning:

1. v1: Lexical Anonymization (Names removed).
2. v2: Chronological & Syntactic Hardening (Pronouns neutralized + Generational Proximity Filter added).
3. v3: Demographic Homogeneity (Gender-matched distractors added to prevent elimination via lexical cues like "WTA" or "Best Actress").

Usage

You can load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the Technology Shadow split
dataset = load_dataset("shadow-bench/ShadowBench", "technology", split="upper_shadow")

# Inspect a sample
print(dataset[0])

Licensing

This dataset is derived from Wikipedia and is licensed under CC BY-SA 4.0.

Citation

If you use this dataset in your research, please cite our paper: [TBD]