anvo25/deep-bias

Deep and shallow biases in language models

by An Vo^1,2*, Vy Tuong Dang^3*, Khai-Nguyen Nguyen⁴, Emilio Villa-Cueva¹,
Thamar Solorio^1†, Anh Totti Nguyen^5†, Daeyoung Kim^3†

^*Equal contribution    ^†Equal advising
¹MBZUAI, ²University of Michigan, ³KAIST, ⁴University of Virginia, ⁵Auburn University

The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026)

---

Abstract

Large language models often repeatedly select the same answer even when many alternatives are plausible. Prior work treats this concentration as bias, but it does not distinguish stable model preferences from responses that depend on a particular prompt wording. We introduce a bias depth score that measures both how strongly a model prefers its top answer under direct prompting and whether that answer survives scenario reframing. Across 4,442 opinion prompts and four large language models, only about a quarter of the concentrated preferences survive reframing. We call these persistent cases Deep biases, and the remaining prompt-dependent cases Shallow biases. Our results show that Deep biases are more often inherited from pretraining and preserved through SFT. Under both continued fine-tuning and prompt-based debiasing for diversity, Deep biases are consistently harder to remove than Shallow biases. Bias depth therefore separates stable learned biases from prompt-wording artifacts that single-prompt metrics conflate.

---

How it works

For every prompt family (for example "Choose a random popular butterfly"), each model was sampled 30 times on the direct prompt and once on each of 30 everyday reframings of the same choice.

• DR (direct rate): share of the 30 direct answers that give the top answer.
• FR (framed rate): share of the 30 reframings that give that same top answer.
• π = DR × FR: the bias depth score.

A family is Deep if π > 0.40, Shallow if it is not Deep and DR > 0.65, and Non-bias otherwise.

Load

from datasets import load_dataset

prompts  = load_dataset("anvo25/deep-bias", "prompts", split="train")
framings = load_dataset("anvo25/deep-bias", "framings", split="train")
sft      = load_dataset("anvo25/deep-bias", "olmo3_7b_sft", split="train")

Configs

Config	Rows	Content
prompts	4,442	prompt families
framings	133,260	30 reframings per family
olmo3_7b_pretrained	4,442	allenai/Olmo-3-1025-7B
olmo3_7b_sft	4,442	tuongvy2603/BITD_baseline
olmo3_7b_dpo	4,442	allenai/Olmo-3-7B-Instruct-DPO
olmo3_7b_rlvr	4,442	allenai/Olmo-3-7B-Instruct
tulu3_8b_sft	4,442	allenai/Llama-3.1-Tulu-3-8B-SFT
claude_sonnet_5	4,440	Claude Sonnet 5 (OpenRouter, reasoning off)
gpt_5_6_sol	4,442	GPT-5.6 Sol (OpenAI API, reasoning off)
olmo3_7b_sft_gepa	4,442	Olmo-3-7B-SFT with the GEPA-optimized system prompt
olmo3_7b_sft_lora	4,442	Olmo-3-7B-SFT with the diversity LoRA adapter (adapter coming soon)

The pretrained model is queried with a Q: {prompt}\nA: completion prompt. All other models get the prompt followed by "Pick exactly ONE option. Respond with only your choice, 1-3 words, no other text." Sampling uses temperature 0.6 and at most 25 new tokens.

Fields

prompts

• id: row index in allenai/Dolci-Instruct-SFT (revision bd3c8f3a9b2cc5a9682e44b96ddd0bb2ff027221) of the row the family was built from
• prompt: the direct prompt, always of the form "Choose a random ..."
• example_answers, reason: answers and a short explanation written by the rewriting model
• evidence: the Dolci-Instruct-SFT row(s) the prompt was built from (user and assistant turns)
• family_size_estimate, duplicate_count, total_family_size: how many Dolci rows ask a similar question
• merged_row_idxs, merged_rewrites, judge_merges: the near-duplicate prompts merged into this family

framings: id, framing_idx (0 to 29), prompt, framing_topic, framing (the scenario text sent to the model)

Model configs: one record per prompt family

• id, prompt, model
• direct: top_answer, dr, n_samples, n_clusters, and distribution, the answer clusters with answer, count, rate, surface_forms (the raw strings in the cluster) and lemma_keys
• framed: the same fields, with fr in place of dr. Here top_answer is the most common framed answer, which can differ from the direct one.
• pi, bias_type (deep, shallow, or non_bias)

Raw responses: raw/<config>/direct.jsonl.gz (id, sample_idx, prompt, response) and raw/<config>/framed.jsonl.gz (id, framing_idx, framing, response), 133,260 rows each. Download them with huggingface_hub.hf_hub_download("anvo25/deep-bias", "raw/olmo3_7b_sft/direct.jsonl.gz", repo_type="dataset").

How it was built

The prompt families come from Dolci-Instruct-SFT, the instruction data of Olmo 3. We kept rows that ask for something random, rewrote each into one "Choose a random ..." prompt with an LLM, merged near duplicates, and wrote 30 reframings per prompt. Code: GitHub. Two API keys found in the copied Dolci rows were replaced with [REDACTED_TOKEN].

License

ODC-BY, following Dolci-Instruct-SFT. Model outputs from Claude and GPT are subject to the terms of their providers.

Citation

@inproceedings{vo2026deepbias,
  title     = {Deep and shallow biases in language models},
  author    = {Vo, An and Dang, Vy Tuong and Nguyen, Khai-Nguyen and Villa-Cueva, Emilio and Solorio, Thamar and Nguyen, Anh Totti and Kim, Daeyoung},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2609.09901}
}