SHARE: Social-Humanities AI for Research and Education
| Authors | João Gonçalves et al. |
| Year | 2026 |
| HF Upvotes | 1 |
| arXiv | 2604.11152 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
This intermediate technical report introduces the SHARE family of base models and the MIRROR user interface. The SHARE models are the first causal language models fully pretrained by and for the social sciences and humanities (SSH). Their performance in modelling SSH texts is close to that of general purpose models (Phi-4) which use 100 times more tokens, as shown by our custom SSH Cloze benchmark. The MIRROR user interface is designed for reviewing text inputs from the SSH disciplines while preserving critical engagement. By prototyping a generative AI interface that does not generate any text, we propose a way to harness the capabilities of the SHARE models without compromising the integrity of SSH principles and norms.
Engineering Breakdown
Plain English
This paper introduces SHARE, a family of causal language models specifically pretrained on social sciences and humanities (SSH) texts, along with MIRROR, a UI for engaging with those models. The key finding is that SHARE models achieve performance comparable to general-purpose models like Phi-4 on SSH-specific text tasks while using 100 times fewer training tokens, as measured by a custom SSH Cloze benchmark. Rather than building another text-generation system, the authors designed MIRROR as a read-only interface that leverages SHARE's understanding of SSH texts without actually generating new content, addressing the concern that generative AI compromises scholarly integrity. This represents a fundamental rethinking of how AI interfaces should work for academic disciplines with strong norms around original authorship.
Core Technical Contribution
The core innovation is domain-specialized pretraining: the SHARE models are the first causal language models trained from scratch on SSH corpora rather than general web text, achieving dramatically better token efficiency (100x fewer tokens than Phi-4 for equivalent performance). The architectural novelty lies in the MIRROR interface design, which inverts the typical generative AI paradigm by building a non-generative system that preserves critical engagement with source material instead of replacing human authorship. This combination—domain-adapted pretraining plus constraint-based interface design—represents a new category of AI system that prioritizes disciplinary principles over raw capability. The custom SSH Cloze benchmark itself is a contribution, providing the first standardized evaluation specific to social science and humanities text understanding.
How It Works
SHARE begins with data collection: curating a large corpus of SSH texts (academic papers, books, primary sources) that represent the actual distribution of language in these disciplines, rather than internet-scale data. These texts are then used as the pretraining corpus for causal language models trained with standard next-token prediction objectives, but with vocabulary and tokenization tuned to SSH language patterns. The SSH Cloze benchmark works by masking tokens in held-out SSH texts and measuring how accurately the model predicts them—similar to BERT's masked language modeling evaluation but applied in the causal (left-to-right) setting. The MIRROR interface takes text input from a user (e.g., a research question or passage from a paper) and uses the SHARE model to compute embeddings, highlight relevant concepts, surface related citations, and annotate passages—but crucially, it never generates new text, only surfaces and organizes existing content. The interface acts as an enhanced search and sense-making tool rather than a content producer.
Production Impact
For teams building research tools or academic software, this approach eliminates the need to license expensive general-purpose models or fine-tune on proprietary institutional data—SHARE models can be deployed directly on SSH-specific tasks with minimal compute overhead and dramatically better performance per token. The non-generative interface design solves a real adoption problem: universities and publishers can integrate these systems without triggering authorship concerns, institutional review, or the overhead of plagiarism detection systems, making adoption friction much lower than with ChatGPT-style systems. The 100x token efficiency gain translates directly to lower inference costs, faster response times, and the ability to run models locally on institutional hardware rather than relying on cloud APIs, which is critical for preserving academic data privacy. The trade-off is architectural: you lose the flexibility of free-text generation but gain interpretability, auditability, and alignment with institutional values—a worthwhile exchange for many academic settings. Integration requires only standard embedding and retrieval pipelines familiar to most ML engineers, though dataset curation to maintain domain coverage over time remains a non-trivial ongoing cost.
Limitations and When Not to Use This
The paper does not address how SSH Cloze performance translates to downstream task performance in real research workflows—better token prediction on held-out text does not necessarily mean the system will be useful for literature review, hypothesis formation, or teaching, which would require different evaluations. SHARE models are built on curated SSH corpora that may overrepresent certain disciplines, geographic regions, or publishing venues, potentially biasing recommendations and annotations toward mainstream scholarship and away from marginalized or emerging perspectives in those fields. The non-generative constraint, while philosophically aligned with SSH principles, may limit the system's usefulness for tasks like summarization, translation, or synthesis where generation is genuinely valuable—the paper does not explore this boundary clearly. Scale limitations are unclear: the report does not specify the size of the training corpus, model parameters, or how performance scales, making it difficult to predict whether SHARE will handle long documents, specialized subfields, or rapidly evolving SSH discourse. Follow-up work is needed on: (1) evaluation beyond Cloze metrics, (2) handling of languages beyond English, (3) adaptation mechanisms when new SSH subdisciplines emerge, and (4) quantitative studies of whether MIRROR actually changes researcher behavior compared to baseline tools.
Research Context
This work builds on the broader trend of domain-adapted language models (e.g., SciBERT for computer science, PubMedBERT for biomedical text) but extends the idea to an entire cross-disciplinary domain (SSH) and, importantly, challenges the generative-AI-as-default assumption that has dominated recent research. It responds to growing criticism within academia and AI ethics of large language models as threats to authorship and scholarly integrity, offering a concrete alternative that maintains linguistic understanding without enabling plagiarism or displacement of human thought. The paper sits at the intersection of natural language understanding (embedding and ranking SSH content) and human-computer interaction (interface design that enforces non-generation), opening a research direction on AI systems designed to augment rather than replace domain expertise. The SSH Cloze benchmark becomes a standard that future work in scholarly AI can build on, similar to how GLUE and SuperGLUE standardized evaluation for general NLP; this could catalyze further investment in SSH-specific model development and evaluation.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::