Aggregate Benchmarks Collapse Under Individual Variation

Federal agencies are deploying LLMs to summarize public comments during rulemaking. If these systems treat identical comments differently based on demographic signals, they could systematically distort democratic input.

Method: Researchers held comment content constant and varied only demographic attribution across 182 public comments, generating over 106,000 summaries from eight LLMs. Occupation produced consistent differential treatment: the same comment attributed to a street vendor received summaries that preserved less original meaning, used simpler language, and shifted emotional tone compared to attribution to a financial analyst. This pattern held across all names, prompts, models, and regulatory contexts. Race and gender effects were inconsistent or absent.

Caveats: Tested on public comment summarization only. Other government text processing tasks may show different bias patterns.

Reflections: Do occupation-based biases persist when LLMs are fine-tuned on government-specific corpora? · Can prompt engineering reduce socioeconomic bias without degrading summary quality? · How do these biases compound when LLMs are used for downstream decision-making, not just summarization?

LLMs are answering patient medical questions, but semantic similarity metrics don't measure medical accuracy or reveal health equity risks embedded in their responses.

Method: The VB-Score framework evaluates medical QA across four components: entity recognition, semantic similarity, factual consistency, and structured information completeness. Testing three widely-used LLMs on 48 public health topics revealed severe performance failures and a major discrepancy between semantic and entity accuracy. Models exhibited 13.8% lower performance on chronic conditions affecting older and minority populations compared to overall average—evidence of condition-based algorithmic discrimination.

Caveats: Framework tested on public health topics from authoritative sources. Clinical diagnostic accuracy may require additional evaluation dimensions.

Reflections: Can targeted fine-tuning on underrepresented health conditions close the performance gap? · Do these disparities persist when LLMs are used as retrieval systems rather than generative answerers? · How do entity recognition failures translate into patient harm in real-world deployment?

Composing prompts as linear text strings becomes unwieldy when users need to express multifaceted intents, iterate on components, or reuse prompt structures across contexts.

Method: Object-Oriented Prompting (OOPrompt) treats prompts as structured, manipulable artifacts rather than text strings. A formative study with 20 participants informed a design space, followed by a validation study of the full prototype. Users could create, edit, iterate, and reuse prompt components as discrete objects, unifying several existing point systems into a coherent interaction paradigm.

Caveats: Validation study focused on understanding added values and trade-offs. Long-term adoption patterns and learning curves remain uncharacterized.

Reflections: What's the cognitive overhead of learning object-oriented prompting versus traditional text-based prompting? · How do teams collaborate when prompts are structured artifacts rather than shared text? · Can OOPrompt patterns be automatically extracted from existing text-based prompts?