ADRA-Bank: A Modular Benchmark for Academic Deep Research Agents

Summary

Benchmark aimed specifically at ACADEMIC deep-research agents rather than general web search: 200 human-annotated instances across 10 academic domains, covering both research and review papers, with a modular evaluation paradigm that separates planning, retrieval and reasoning and runs both end-to-end (agents) and isolated (backbone LLM) modes. Finds agents strong in narrow specialisms but weak at multi-source retrieval and cross-field consistency, and identifies high-level planning as the binding constraint. Complements the AstaBench/AIRS-Bench family already in the KB by targeting the literature-synthesis stage.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-08-29 literature sweep as a medium-priority candidate for the RISE knowledge base (angle: research-agent benchmarks). Verified against: arXiv abstract page opened 2026-08-29 (title, 10 authors, submitted 30 Nov 2025, full abstract read)

Critique / open questions

Queued — not yet assessed.

Key quotes

A surge in academic publications calls for automated deep research (DR) systems, but accurately evaluating them is still an open problem. (abstract)

First, existing benchmarks often focus narrowly on retrieval while neglecting high-level planning and reasoning. (abstract)