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 examines the limitations of current AI benchmarks by focusing on factors such as modality (chat UI vs API), search capabilities, consistency of responses, citation grounding, and abstention behavior using ChatGPT and OpenAI's models. The researchers found significant differences in accuracy between the two modalities, with web search enabling reducing accuracy, and inconsistency in responses across repeated runs and modalities.

What the Document Says

  • The study evaluates large language models (LLMs) using two modalities: ChatGPT's chat UI and OpenAI's API.
  • The evaluation considers both cases with web search enabled and disabled.
  • The researchers used a total of 401 prompts from BBQ and SafetyBench benchmarks, resulting in 4,812 responses across three repeated runs per prompt.
  • Beyond standard performance measures, the study evaluates model output dimensions like response consistency, response text similarity, citation grounding, and abstention behavior.
  • API responses were found to be more accurate than chat UI responses on both benchmarks with search disabled.
  • Enabling web search decreased accuracy by up to 8 percentage points.
  • The direction of modality performance trends was reversed for one benchmark when web search was enabled.
  • Inconsistent responses were found in up to 21% of prompts across repeated runs.
  • Different citations were used to ground answers between the two modalities.
  • Abstention behavior was inconsistent across both modalities.
  • The authors suggest that AI safety evaluations should account for modality, multi-run consistency, search conditions, and response-level behaviors.

Key Concepts Mentioned

  • large language models
  • ChatGPT
  • OpenAI
  • API
  • chat UI
  • web search
  • benchmarks
  • accuracy
  • response consistency
  • response text similarity
  • citation grounding
  • abstention behavior

Claims Made by the Document

  • API responses are more accurate than chat UI responses on both benchmarks with search disabled.
  • Enabling web search decreases accuracy.
  • The direction of modality performance trends is reversed for one benchmark when web search is enabled.
  • Inconsistent responses are found in up to 21% of prompts across repeated runs.

Evidence Used

  • Results from comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled.
  • Data collected from a total of 401 prompts and 4,812 responses across three repeated runs per prompt.

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.