mlsanityData to 14 Oct 2026

SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment

arXiv 2411.10912 ↗First posted 16 Nov 2024cs.CLcited by 6

On the map: LLM alignment / RLHF / preference optimisation → LLM preference optimization methods

Abstract (arXiv)

When different groups' values differ, one approach to model alignment is to steer models at inference time towards each group's preferences. However, techniques like in-context learning only consider similarity when drawing few-shot examples and not cross-group differences in values. We propose SPICA, a framework that accounts for group-level differences during in-context example retrieval. SPICA introduces three designs: scenario banks, group-informed retrieval metrics, and in-context alignment prompts. From an evaluation of SPICA on an alignment task collecting inputs from four demographic groups (n = 544), our metrics retrieve in-context examples that more closely match observed preferences, with the best prompt configuration using multiple contrastive responses to demonstrate examples. In an end-to-end evaluation (n = 120), we observe that SPICA is higher rated than similarity-based retrieval, with groups seeing up to a +0.16 point improvement on a 5 point scale. Additionally, gains from SPICA were more uniform, with all groups benefiting from alignment rather than only some. Finally, we find that while a group-agnostic approach can align to aggregated values, it is not most suited for divergent groups.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 39 of its 65 references that are in the map. First posted on arXiv 16 Nov 2024.

Builds on

39 of 65 references.

Closest ancestors · 3

Same direction, newest first

  1. 2023The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning4 Dec
  2. 2023Aligning Large Language Models through Synthetic Feedback23 May
  3. 2022Training language models to follow instructions with human feedback4 Mar

Other references · 36

  1. 2020Retrieval-Augmented Generation for Knowledge-Intensive NLP Taskscited by 9,297
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  3. 2020Dense Passage Retrieval for Open-Domain Question Answeringcited by 3,824
  4. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 3,050
  5. 2019BioBERT: a pre-trained biomedical language representation model for biomedical text miningcited by 1,947
  6. 2020Don't Stop Pretraining: Adapt Language Models to Domains and Taskscited by 1,565
  7. 2022Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?cited by 1,434
  8. 2022Diffusion-LM Improves Controllable Text Generationcited by 1,027
  9. 2021Ethical and social risks of harm from Language Modelscited by 966
  10. 2023BloombergGPT: A Large Language Model for Financecited by 939
  11. 2022A Survey on In-context Learningcited by 721
  12. 2021Learning To Retrieve Prompts for In-Context Learningcited by 648
24 more
  1. 2022Precise Zero-Shot Dense Retrieval without Relevance Labelscited by 517
  2. 2020Artificial Intelligence, Values and Alignmentcited by 448
  3. 2023Personalized Soups: Personalized Large Language Model Alignment via Post-hoc Parameter Mergingcited by 280
  4. 2022Active Example Selection for In-Context Learningcited by 196
  5. 2024Collective Constitutional AI: Aligning a Language Model with Public Inputcited by 123
  6. 2024Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaborationcited by 113
  7. 2024Reinforcement Learning for LLM Post-Training: A Surveycited by 92
  8. 2024The Political Preferences of LLMscited by 84
  9. 2022Structured Prompting: Scaling In-Context Learning to 1,000 Examplescited by 81
  10. 2021Designing Toxic Content Classification for a Diversity of Perspectivescited by 65
  11. 2024(A)I Am Not a Lawyer, But...: Engaging Legal Experts towards Responsible LLM Policies for Legal Advicecited by 63
  12. 2024Large Language Models Reflect the Ideology of their Creatorscited by 57
  13. 2023The Past, Present and Better Future of Feedback Learning in Large Language Models for Subjective Human Preferences and Valuescited by 54
  14. 2024Ethical Reasoning and Moral Value Alignment of LLMs Depend on the Language we Prompt them incited by 46
  15. 2024Can Language Models Reason about Individualistic Human Values and Preferences?cited by 34
  16. 2024AI Alignment with Changing and Influenceable Reward Functionscited by 33
  17. 2024What are human values, and how do we align AI to them?cited by 32
  18. 2023In-Context Alignment: Chat with Vanilla Language Models Before Fine-Tuningcited by 19
  19. 2023Specialist or Generalist? Instruction Tuning for Specific NLP Taskscited by 17
  20. 2021What Makes Online Communities 'Better'? Measuring Values, Consensus, and Conflict across Thousands of Subredditscited by 13
  21. 2022Enabling Classifiers to Make Judgements Explicitly Aligned with Human Valuescited by 13
  22. 2024ValueScope: Unveiling Implicit Norms and Values via Return Potential Model of Social Interactionscited by 12
  23. 2023How Does In-Context Learning Help Prompt Tuning?cited by 11
  24. 2023Case Law Grounding: Using Precedents to Align Decision-Making for Humans and AIcited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 3; 2026: 3.

6 papers.

  1. 2025Value Profiles for Encoding Human Variationcited by 29
  2. 2025PICACO: Pluralistic In-Context Value Alignment of LLMs via Total Correlation Optimizationcited by 3
  3. 2025Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemblecited by 2
  4. 2026Structuring the Space of Sociotechnical Alignmentcited by 2
  5. 2026Overton Pluralistic Reinforcement Learning for Large Language Modelscited by 1
  6. 2026Evaluating Pluralism in LLMs through Latent Perspectivescited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Aligned at the Start: Conceptual Groupings in LLM Embeddings
  2. 2024Context-DPO: Aligning Language Models for Context-Faithfulness
  3. 2023Beyond Output Matching: Bidirectional Alignment for Enhanced In-Context Learning
  4. 2026A Unifying Framework for Concept-Based Representational Similarity
  5. 2025Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
  6. 2026Group Alignment-Induced Sycophancy: A Two-Sided Evaluation of Steerable Pluralistic Alignment
  7. 2025RPO: Retrieval Preference Optimization for Robust Retrieval-Augmented Generation
  8. 2025Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples