The Prompt Files
Historical trail · continuously updated

White papers that built the field

From Attention Is All You Need to this month. Each entry has a verified link and one sentence on why it matters to someone who writes prompts for a living. Essentials are marked; the rest are the connective tissue between them.

—
papers and reports
—
years covered
Most recent entry
—

The timeline

Showing —. Click a title for the abstract; PDF links go straight to arXiv.

    How to read this list

    Essentials first

    If you only read the starred entries you will understand every major turn in the field: attention, scaling, few-shot prompting, chain-of-thought, RLHF, open weights, long context, test-time compute, reasoning models, agents, and the alignment results that complicate all of it.

    Vendor reports count

    Since 2023 much of the most important work ships as a company blog post or system card rather than an arXiv paper. Those are included and linked to the primary source, never to coverage of it.

    Corrections welcome

    The list is maintained in a single data file in the site repository. If a landmark paper is missing or a link is wrong, open an issue or a pull request.

    Benchmarks and community evaluation

    Where capability claims get checked.

    OpenAI Evals

    The open evaluation framework this site has contributed to since 2023. Pull requests · Issues

    SWE-bench

    Real GitHub issues, real repositories. The yardstick for coding agents, with Verified and Pro variants.

    LMArena (Chatbot Arena)

    Blind pairwise human votes. Imperfect, gameable, and still the leaderboard everyone watches.

    ARC Prize

    Chollet's abstraction-and-reasoning tasks, now in a second generation, as a measure of novel problem solving rather than memorized skill.

    Humanity's Last Exam

    Expert-written questions at the frontier of human knowledge, designed to be the last closed-ended academic benchmark.

    GAIA and τ-bench

    General-assistant and tool-agent benchmarks that measure the thing users care about: did the agent finish the job, and does it do so reliably?

    Research program

    The site's own agenda, carried forward from 2023.

    Thread 1 · since 2023

    Multilingual transfer for low-resource languages

    Adaptive Language Weighting (ALW): reweighting the contribution of high-resource languages during adaptation so that transfer to low-resource targets improves without catastrophic forgetting. Linear (ALRMM) and non-linear (N-ALRMM) weighting schedules. Status: working notes and code; see the low-resource prompting guide for the practitioner side.

    Thread 2 · since 2023

    Prompt-chain structuring

    Methods for decomposing a goal into a chain of model calls whose intermediate outputs are typed and checkable. This thread grew into the Prompt Constructor and the agentic-workflow guide in the library.

    Thread 3 · 2025 →

    Evaluating reasoning traces

    Given the 2025 findings that chains of thought are often unfaithful, how should a practitioner read them? Collecting cases where the visible reasoning predicted the answer and where it did not.

    Code

    Repositories

    All research code is published under github.com/JLColeman2792.