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 pathways (prompt text, conversation history, in-context demonstrations, retrieved documents, tool outputs). Each item is presented under 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: https://github.com/Ydrg9989/AnchorBench
• Paper: AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMs (COLM 2026)
• Language: English
• Task: Numeric estimation (0–100) under anchoring manipulations
• License: CC BY 4.0 (items are synthetic and author-generated)
• Splits: none. AnchorBench is eval-only; every row is split: test.

Two-Layer Structure

The benchmark is organized into core and ablation layers:

Core Layer (main results)

The core layer contains the base 5-condition comparison used for all primary results. Every suite shares these conditions:

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

File: promptviews_core.jsonl in each suite directory.

Ablation Layer (robustness checks)

Suite-specific extended conditions for appendix analyses:

Suite	Ablation Conditions	Purpose
External	placebo_low/high, authority_low/high	Salience vs. content controls
History	control_twostage	Matched two-stage neutral control
ICL	neutral_low/high	Pure numeric priming (no semantic header)
RAG	*_order_first/last, *_nodiscl, *_authority	Position, disclaimer, framing ablations
Tool	(none)	Core-only

File: promptviews_ablation.jsonl in each suite directory (where applicable).

Suites (6)

Suite	Anchor pathway	Items	Core views	Ablation views	All views
external_core	Sentence in the prompt	360	1,800	1,440	3,240
history_core	The model's own prior turn	360	1,800	360	2,160
icl_core	Demonstration metadata	360	1,800	720	2,520
rag_core	Retrieved document	360	1,800	4,320	6,120
tool_core	Tool-call response	360	1,800	0	1,800
external_uncertain	Prompt, with only k of 5 ratings visible	360	5,400	0	5,400

Total core: 1,800 items, 9,000 core prompt views (5 suites). With ablations: 15,840 prompt views. Released on the Hub: 14,400 rows -- the 9,000 core views plus the 5,400 uncertain views. external_uncertain backs Table 2 in the main paper.

Data Schema

Each file is JSONL: one JSON object per line.

Path layout: {suite}_core/promptviews_core.jsonl (core) or {suite}_core/promptviews_ablation.jsonl (ablations).

Field	Type	Description
item_id	string	Unique item id (same for all conditions of one item).
condition	string	Condition name (see tables above).
prompt_text	string	Full prompt to send to the model.
anchor_value	int | null	Numeric anchor in the prompt. null for controls.
anchor_relevance	string	"none", "irrelevant", "plausible", "placebo", "authority", or "neutral".

Ground-truth answer y_star is in itemspecs.jsonl (matched by item_id).

How to Use

Load the data

from huggingface_hub import hf_hub_download
import json

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

Minimal evaluation

1. Control accuracy: For condition == "control", compare model output to y_star from itemspecs.
2. UAI (Unified Anchor Influence): 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).

Extended evaluation

Use the provided metrics module for bootstrap CIs, paired Wilcoxon tests with BH correction, parse-policy sensitivity, and by-offset/by-difficulty breakdowns:

from anchorbench_eval.metrics import compute_extended_metrics
metrics = compute_extended_metrics(records)
# metrics includes: uai_irr_ci, uai_plaus_ci, disc_delta_ci,
# p_plaus_vs_irr, p_irr_vs_zero, parse_by_condition, etc.

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.

Citation

@inproceedings{anchorbench2026,
  title     = {AnchorBench: A Theory-Grounded, Multi-Paradigm Benchmark for Anchoring Bias in Large Language Models},
  booktitle = {Proceedings of the Conference on Language Modeling (COLM)},
  year      = {2026},
}

Dataset Structure

Released layout on the Hub — one file per suite, 14,400 rows in total:

data/external.jsonl              1,800 rows
data/history.jsonl               1,800
data/icl.jsonl                   1,800
data/rag.jsonl                   1,800
data/tool.jsonl                  1,800
data/external_uncertain.jsonl    5,400   (360 items x 15 conditions)

The internal repository additionally carries itemspecs.jsonl and the ablation views for each suite; those are not part of this release. See the code repository for them.

Fields

Field	Type	Description
item_id	string	Stable identifier. Encodes suite, domain, difficulty (e/h) and anchor offset, e.g. EXT-pricing_wtp-e-off15-001.
suite	string	external, history, icl, rag, tool, external_uncertain.
condition	string	One of the five matched conditions (15 for the uncertain suite).
relevance	string	none for control, else irrelevant / plausible / self_generated.
anchor_polarity	string | null	low or high; null for control.
anchor_value	int | null	The number placed in the prompt. Null for control and for all of History — see below.
offset	int | null	Designed distance of the anchor from the evidence centre: 15, 25 or 40.
difficulty	string	easy (5 ratings, low noise) or hard (3 of 5 shown, higher noise).
domain	string	One of six business domains.
prompt_text	string	The exact prompt the models were given.
y_star_evidence	int	Gold answer: the rounded mean of the full five ratings.
split	string	Always test.

anchor_value is null for every History row, deliberately. History's anchor is the model's own Stage-1 answer, so it is a property of the (item, model) run rather than of the item — the same item records 80 for Qwen-7B and 81 for Llama-8B. Publishing one number would misrepresent what the model saw. Take the realised value from the released results instead.

Loading

from datasets import load_dataset

ds = load_dataset("Yiderigun/LLM_anchoring", data_files="data/external.jsonl")["train"]

# UAI needs the control answer for the same item, so group by item_id.
row = ds[0]
print(row["condition"], row["anchor_value"], row["y_star_evidence"])

Scoring an anchored response, given the model's control answer y_ctrl for the same item_id:

UAI = (y_anchored - y_ctrl) / (anchor_value - y_ctrl)

Items where |anchor_value - y_ctrl| < 3 are excluded from UAI in the paper: the near-zero denominator makes the ratio unstable. Conclusions are stable across thresholds of 1, 3 and 5.

Known properties worth stating

• Eval-only. No train or validation portion; every row is split: test.
• Three prompts repeat inside uncertain_p1_control (6 of 360 items, 1.7%) with different gold answers. That is the design rather than a collision: at k=1 the prompt shows one visible rating while gold stays the mean of all five, so two items sharing a first rating give identical prompts with different answers. The irreducible uncertainty is the point of that condition.
• Synthetic by construction. Items are instantiated from author-designed templates; scenario text, evidence labels and anchor framings come from pools generated with LLM assistance and then verified by the authors, while the numeric evidence, anchor values and gold answers come from seeded sampling. AnchorBench is a controlled diagnostic, not a sample of organic user queries.

Version history and the mapping from dataset version to paper tables is in VERSIONS.md.