Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)

Source

arXiv

arxiv.org/abs/2608.06202v1

Read the source

Source

Plain-Language Brief

The study compares two access modalities of a large language model (LLM), ChatGPT's chat UI and OpenAI's API, with and without web search enabled. The research shows that the accuracy of responses varies between the modalities and is affected by web search. They also found inconsistent responses in up to 21% of prompts, different citation grounding, and inconsistent abstention behavior across both modalities.

What the Document Says

  • The study uses a large language model (LLM) for comparison between two access modalities: ChatGPT's chat UI and OpenAI's API.
  • Web search is enabled and disabled in the study.
  • The research collects 4,812 total responses across three repeated runs per prompt from a stratified total sample of 401 prompts from BBQ and SafetyBench benchmarks.
  • Beyond standard performance measures, they evaluate response consistency, response text similarity, citation grounding, and abstention behavior.
  • Chat UI responses were less accurate than API responses on both benchmarks with search disabled.
  • Enabling web search reduced accuracy by up to 8 percentage points.
  • Repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts.
  • The two modalities grounded answers in different citations.
  • Abstention behavior was also inconsistent across both modalities.

Key Concepts Mentioned

  • Large Language Model
  • ChatGPT
  • OpenAI API
  • Web Search
  • BBQ Benchmark
  • SafetyBench Benchmark
  • Response Consistency
  • Response Text Similarity
  • Citation Grounding
  • Abstention Behavior

Claims Made by the Document

  • The accuracy of responses varies between the modalities and is affected by web search.
  • Repeated runs of the same prompt produce inconsistent responses.
  • Different citations are grounded in for the same questions across both modalities.

Evidence Used

  • Chat UI responses were less accurate than API responses on both benchmarks with search disabled.
  • Enabling web search reduced accuracy by up to 8 percentage points.
  • Repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts.
  • The two modalities grounded answers in different citations.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.