task_id stringclasses 8
values | question stringclasses 8
values | gold_answer stringclasses 8
values | source_title stringclasses 7
values | source_url stringclasses 7
values | source_version stringclasses 7
values | evidence_excerpt stringclasses 8
values | evidence_location stringclasses 1
value | verification_rule stringclasses 8
values | construction stringclasses 1
value | what_not_proven stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|
research-rag-memory-001 | According to the RAG abstract, what are its parametric and non-parametric memories, and how is the latter accessed? | A pretrained sequence-to-sequence model is the parametric memory. A dense Wikipedia vector index is the non-parametric memory, accessed by a pretrained neural retriever. | Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks | https://arxiv.org/abs/2005.11401v4 | 2005.11401v4 | a dense vector index of Wikipedia | arXiv abstract at the pinned version URL | Require all three components; the external index is not model parameters. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-rag-passages-002 | Does the RAG abstract require the retrieved passages to remain fixed for every output token? | No. One formulation uses the same passages throughout a generated sequence; the other permits different passages for different output tokens. | Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks | https://arxiv.org/abs/2005.11401v4 | 2005.11401v4 | the same retrieved passages across the whole generated sequence | arXiv abstract at the pinned version URL | Require the distinction between sequence-level and token-level conditioning. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-swebench-size-003 | How many software-engineering problems and Python repositories does the SWE-bench paper abstract report? | It reports 2,294 problems drawn from 12 popular Python repositories. | SWE-bench: Can Language Models Resolve Real-World GitHub Issues? | https://arxiv.org/abs/2310.06770v3 | 2310.06770v3 | real GitHub issues and corresponding pull requests | arXiv abstract at the pinned version URL | Require 2294 problems and 12 Python repositories, with correct units; this is the paper version, not a later subset. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-gaia-release-004 | How many questions did the GAIA authors devise, and how many answers did they retain for the leaderboard? | They devised 466 questions and retained the answers to 300 of them for the leaderboard. | GAIA: a benchmark for General AI Assistants | https://arxiv.org/abs/2311.12983v1 | 2311.12983v1 | we devise 466 questions and their answer | arXiv abstract at the pinned version URL | Require 466 total questions and 300 retained answers; do not claim access to hidden gold answers. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-react-mechanism-005 | What does ReAct interleave, and what complementary functions do its reasoning and actions serve in the abstract? | It interleaves reasoning traces with task-specific actions. Reasoning supports planning and exception handling; actions obtain information from external sources or environments. | ReAct: Synergizing Reasoning and Acting in Language Models | https://arxiv.org/abs/2210.03629v3 | 2210.03629v3 | reasoning traces and task-specific actions in an interleaved manner | arXiv abstract at the pinned version URL | Require interleaving, planning or exception handling, and external information gathering; do not request private reasoning. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-toolformer-decisions-006 | Which four tool-use decisions does Toolformer learn according to its abstract? | Which APIs to call, when to call them, which arguments to pass, and how to use the returned results in subsequent token prediction. | Toolformer: Language Models Can Teach Themselves to Use Tools | https://arxiv.org/abs/2302.04761v1 | 2302.04761v1 | which APIs to call, when to call them, what arguments to pass | arXiv abstract at the pinned version URL | Require API selection, timing, arguments, and incorporation of results; API selection alone is incomplete. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-webarena-scope-007 | Which four website domains does WebArena include, and which property of task completion does its benchmark evaluate? | E-commerce, social forum discussions, collaborative software development, and content management; it evaluates functional correctness of task completions. | WebArena: A Realistic Web Environment for Building Autonomous Agents | https://arxiv.org/abs/2307.13854v4 | 2307.13854v4 | evaluating the functional correctness of task completions | arXiv abstract at the pinned version URL | Require all four domains and functional correctness, rather than fluent explanations or action-count matching. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
research-browsergym-agentlab-008 | What does BrowserGym standardize, and what complementary role does AgentLab play in the ecosystem paper? | BrowserGym offers a common gym-like environment with observation and action spaces across benchmarks. AgentLab supports creating, testing, and analyzing agents and managing experiments. | The BrowserGym Ecosystem for Web Agent Research | https://arxiv.org/abs/2412.05467v4 | 2412.05467v4 | a unified, gym-like environment with well-defined observation and action spaces | arXiv abstract at the pinned version URL | Require the common observation/action interface and complementary agent-development or experiment-management support. | authored source-reading question; source-supported reference answer | No model run, benchmark reproduction, current leaderboard result, or production outcome is established. |
Research Evidence QA: eight source-reading tasks
Eight authored English questions and source-supported reference answers covering RAG, SWE-bench, GAIA, ReAct, Toolformer, WebArena and BrowserGym. Seven version-pinned primary papers supply public evidence, with a short quotation and an explicit answer rubric for every task.
This is a small teaching and verification set. It is not a representative research benchmark. The answers are public. No frontier model was evaluated, and there is no contamination-free or production-performance claim. Describing SWE-bench, GAIA or WebArena is not passing those benchmarks. Hidden benchmark answers are not included.
What was checked
The build fetched all seven pinned arXiv pages, extracted their abstracts, checked every evidence excerpt after whitespace normalization, and checked the supporting answer phrases. source-checks.json holds eight per-task outcomes. provenance.json records source hashes and timestamps. Those lexical checks support traceability; they do not automatically grade semantic answers. Gold answers were independently checked against the primary abstracts during authoring.
reproduce/build_research_qa.py reproduces the source checks using Python's standard library. Run python reproduce/build_research_qa.py; it writes a new bundle below its own directory. No paid model call is used. Live arXiv HTML can change, so fresh page hashes can differ even at a pinned paper version. Network or source-shape failures stop the build.
Related tools, data and code
Try the Verbatim Citation Gate and inspect its source code. Exact quote matching does not establish that a broader claim is true. A separately measured 91-task action-policy baseline includes its source code, conditions and per-case outputs. Agent sessions are a separate corpus of anonymized historical conversations.
The authored questions, answers, rubrics and build code are offered under CC BY 4.0. Short attributed quotations remain subject to their original sources' rights. The paper texts themselves are not redistributed or relicensed here. Attribution: TonyDzi / PaloAltoAiResearchLab; AI-assisted authoring. Project profile.
- Downloads last month
- -