---
type: "post"
id: "code-and-dataset-released-at-anonymous-4open-science-41b622"
slug: "code-and-dataset-released-at-anonymous-4open-science-41b622"
title: "Code and dataset released at anonymous.4open.science; paper on arXiv."
description: "ImplicitBBQ is a question-answering benchmark for testing whether language models respond differently when demographic identity is conveyed through associated characteristics rather than stated directly."
retrieval_nugget: "ImplicitBBQ is a question-answering benchmark for testing whether language models respond differently when demographic identity is conveyed through associated characteristics rather than stated directly. Its authors report a large implicit-versus-explicit bias gap in open-weight models, giving model teams a structured way to test a failure mode that direct-identity evaluations can miss."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["models"]
entities: ["ImplicitBBQ","Bhaskara Hanuma Vedula","Darshan Anghan","Ishita Goyal","Ponnurangam Kumaraguru","Abhijnan Chakraborty"]
related_patterns: ["ai-engineering-evaluations-are-splitting-into-task-specific"]
novelty_label: "new_fact"
what_changed: "ImplicitBBQ introduces a benchmark for implicit bias across six demographic dimensions and releases its code and dataset; its evaluation finds substantially higher bias in ambiguous contexts."
prior_refs: []
source_url: "https://arxiv.org/abs/2604.01925"
source_urls: ["https://arxiv.org/abs/2604.01925"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:code-and-dataset-released-at-anonymous-4open-science-41b622"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/code-and-dataset-released-at-anonymous-4open-science-41b622/","markdown":"https://newruntime.com/posts/code-and-dataset-released-at-anonymous-4open-science-41b622.md","json":"https://newruntime.com/posts/code-and-dataset-released-at-anonymous-4open-science-41b622.json"}
markdown_url: "https://newruntime.com/posts/code-and-dataset-released-at-anonymous-4open-science-41b622.md"
---

# Code and dataset released at anonymous.4open.science; paper on arXiv.

## Retrieval answer

ImplicitBBQ is a question-answering benchmark for testing whether language models respond differently when demographic identity is conveyed through associated characteristics rather than stated directly. Its authors report a large implicit-versus-explicit bias gap in open-weight models, giving model teams a structured way to test a failure mode that direct-identity evaluations can miss.

## Summary
ImplicitBBQ is a question-answering benchmark for testing whether language models respond differently when demographic identity is conveyed through associated characteristics rather than stated directly. Its authors report a large implicit-versus-explicit bias gap in open-weight models, giving model teams a structured way to test a failure mode that direct-identity evaluations can miss.

## What happened
In a paper first submitted to arXiv on April 2, 2026 and revised on November 9, the five authors of ImplicitBBQ introduce a benchmark for evaluating implicit bias in large language models. The paper argues that models increasingly suppress biased outputs when an identity is named explicitly, while biases can remain when identity is implied indirectly. The authors describe this as an evaluation gap: a model may look safer under a test that names a demographic group while responding differently to traits conventionally associated with that group.

The benchmark uses characteristic-based cues rather than relying only on names as demographic proxies. The authors say name-based cues can have weak associations with many social demographics and cannot readily cover dimensions including age and socioeconomic status. In ImplicitBBQ, attributes associated with a demographic group signal identity indirectly. The benchmark spans age, gender, region, religion, caste, and socioeconomic status, and is framed as a question-answering evaluation.

The authors compare ambiguous contexts, in which a characteristic acts as an indirect cue, with explicit contexts that state the relevant identity. Across 11 models, they report that implicit bias in ambiguous contexts is more than six times higher than explicit bias in open-weight models. Results also varied materially by demographic dimension: caste was the most severe category in their evaluation, while gender was the least affected. The paper does not claim that each model or each category follows the same pattern.

The paper also evaluates several prompting interventions. According to the authors, safety prompting and chain-of-thought reasoning do not substantially close the implicit-bias gap. Few-shot prompting reduces implicit bias by 79%, but caste bias remains four times the level of any other evaluated dimension. This is an important qualification to the numerical result: the reported reduction is not presented as eliminating the problem, and the remaining disparity is concentrated unevenly rather than distributed evenly across the benchmark.

The authors say they publicly release code and data for model providers and researchers to benchmark possible mitigation techniques. The arXiv record identifies the current paper as version 3; version 1 was submitted in April 2026 and version 2 in July. The abstract lists no price, service tier, API, or usage limit. The release is therefore a research artifact for evaluation, not a product announcement or a claim that any provider has implemented a mitigation.

## What the archive adds
New Runtime's earlier coverage has examined another model-risk pathway: fine-tuning on an untrusted base model can expose private training examples to extraction. [Code and data released at thu-coai/Backdoor-Data-Extraction](https://newruntime.com/posts/code-and-data-released-at-thu-coai-backdoor-a427ad/) addresses provenance and data protection. ImplicitBBQ is a new event and a different measurement problem: it tests demographic associations triggered by indirect cues rather than private-example extraction.

## Trend signal
This does not yet fit a pattern New Runtime tracks. ImplicitBBQ is a research benchmark about measuring bias in model responses, rather than evidence of a durable shift in agent architecture, runtime security boundaries, or model-output interfaces.

## Why now
The revised version arrives after two earlier arXiv versions in April and July 2026. Its stated contribution is timely because explicit-identity checks can produce reassuring outputs while leaving indirect, characteristic-based associations untested. The paper positions code and data release as an invitation to test mitigations against that less direct failure mode.

## Why it matters
Teams building screening, support, moderation, recommendation, or decision-assistance workflows cannot assume that demographic information arrives as a clean field called identity. Location, cultural references, behavior, religious practice, age-related details, or socioeconomic circumstances can all become contextual cues in a prompt or retrieved document. A review suite that only swaps explicitly named groups may therefore test a narrower surface than the application actually exposes.

ImplicitBBQ provides a vocabulary and benchmark structure for separating explicit from indirect cues across six demographic dimensions. Its findings also challenge a common operational shortcut: adding a safety instruction or requesting a reasoning trace should not be treated as evidence that unequal associations have been removed. The practical use is to add indirect-cue cases to internal evaluation and red-team work, then measure the actual model, prompt template, retrieval context, and decision threshold a team plans to deploy.

## Limits and uncertainty
The findings are reported by the paper's authors and have not been independently reproduced here. The abstract does not name the 11 models, describe model-selection criteria, specify sample sizes, explain scoring in full, or report uncertainty intervals. Those omissions make it impossible to infer universal rates for other models, deployments, languages, or task types from the reported ratios alone.

The paper also does not establish that lower scores on this benchmark produce lower discrimination in real-world outcomes. Characteristic-based cues may not represent every relevant social signal, and the strength of an association can depend on context. Although the authors say code and data are publicly released, their completeness, usability, licensing, and reproducibility have not been independently checked. The results should guide further evaluation rather than serve as a standalone fairness certification.

## Skill to build now
Add a small indirect-cue test set to one production evaluation this week. For each existing case that explicitly names a demographic category, write a paired case where identity is conveyed only through an associated characteristic, then compare outputs using the same rubric. Keep the cases separate from live decisions, record the model and prompt version, and investigate disparities before treating a prompting intervention as a mitigation.

## Reader takeaway
A model can avoid biased language when identity is stated directly while still making unequal associations from surrounding characteristics. ImplicitBBQ offers a public research framework for testing that gap, especially across age, gender, region, religion, caste, and socioeconomic status. Use it as a prompt to expand evaluation, not as proof that a particular deployment is fair.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)