Yiderigun/LLM_anchoring

AnchorBench

A multi-paradigm benchmark for anchoring bias in large language models.

Dataset Description

AnchorBench measures how much LLM numeric estimates shift toward salient reference numbers delivered through five interface channels (prompt text, conversation history, in-context demos, retrieved documents, tool outputs). Each item is presented under five matched conditions that share the same evidence and gold answer; only the anchor changes. The benchmark distinguishes irrelevant anchors (transparently arbitrary) from plausible anchors (weakly credible), so you can measure both raw susceptibility and relevance discrimination.

• Homepage: [Repository / paper link]
• Paper: AnchorBench: A Theory-Grounded, Multi-Paradigm Benchmark for Anchoring Bias in Large Language Models
• Language: English
• Task: Numeric estimation (0–100) under anchoring manipulations

Suites (5)

Suite	Anchor channel	Description
external_core	Sentence in prompt	Classic external anchor ("A recent study found X")
history_core	Prior turn in conversation	Self-generated anchor (model’s Stage 1 estimate)
icl_core	Demo header metadata	Numeric priming ("Prior estimate for similar cases: X")
rag_core	Retrieved document	Anchor in a 3-doc mini-corpus
tool_core	Tool output JSON	Anchor in simulated tool response

Each suite has 360 items and 1,800 prompt views (5 conditions × 360 items). Total: 1,800 items, 9,000 prompt views across all suites.

Conditions (5 per item)

Condition	Anchor	Relevance
control	None	—
irrelevant_low	Low value	Irrelevant (e.g. "case #X")
irrelevant_high	High value	Irrelevant
plausible_low	Low value	Plausible (e.g. "survey suggested X")
plausible_high	High value	Plausible

All five conditions for a given item share the same evidence, question, and gold answer; only the anchor-bearing part of the prompt changes.

Data Schema

Each file is JSONL: one JSON object per line.

Path layout: {suite}_core/promptviews.jsonl (e.g. external_core/promptviews.jsonl).

Field	Type	Description
item_id	string	Unique item id (same for the 5 conditions of one item).
condition	string	One of: control, irrelevant_low, irrelevant_high, plausible_low, plausible_high.
prompt_text	string	Full prompt to send to the model.
y_star	int	Ground-truth answer (0–100). Use for accuracy and anchoring metrics.
anchor_value	int | null	Numeric anchor in the prompt. null for control; integer for the other 4 conditions.

How to Use

Load the data

Download the JSONL from the repo and read locally:

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="Yiderigun/LLM_anchoring",
    filename="external_core/promptviews.jsonl",
    repo_type="dataset",
)
with open(path) as f:
    for line in f:
        row = json.loads(line)
        # row keys: item_id, condition, prompt_text, y_star, anchor_value
        break

Run your model on row["prompt_text"], then compare the output to row["y_star"] (ground truth) and use row["anchor_value"] for anchor conditions when computing UAI/TAR.

Minimal evaluation

1. Control accuracy: For rows with condition == "control", compare model output to y_star.
2. UAI (Unified Anchor Influence): For anchor conditions,
   UAI = (model_answer - y_control) / (anchor_value - y_control)
   (exclude items where |anchor_value - y_control| < 3).
3. TAR (Toward-Anchor Rate): Fraction of anchor-condition answers that shift toward anchor_value relative to control.
4. Discrimination: UAI_plausible - UAI_irrelevant (positive = model discriminates by relevance).

Group by item_id to get the 5 conditions per item; use the control row’s model output as y_control for that item.

Domains and Design

• 6 domains: Pricing/WTP, Operations/Time, Transportation/Logistics, Resource Consumption, Market/Demographics, Legal/Policy.
• 2 difficulties: Easy (concordant evidence), Hard (missing/conflicting values).
• Gold answer: y_star = round(mean(visible evidence)) (0–100).

Intended Use

• Research on anchoring bias in LLMs.
• Benchmarking susceptibility and relevance discrimination across models and suites.
• Not for training; evaluation only.

Dataset Structure

external_core/promptviews.jsonl   # 1,800 lines
history_core/promptviews.jsonl    # 1,800 lines
icl_core/promptviews.jsonl        # 1,800 lines
rag_core/promptviews.jsonl        # 1,800 lines
tool_core/promptviews.jsonl       # 1,800 lines

Total: 9,000 lines (1,800 per suite).