Adding an 'copy exactly' instruction did not make AI essay quotes more accurate

·3 views

A researcher tested whether adding an explicit instruction to copy student words exactly would improve how precisely two small AI models quote English-learner essays. Both models already quoted accurately about 94% of the time, and the extra line changed that rate by a negligible amount, -0.2 points. The preregistered verdict was no clear effect, with a 95% interval ruling out gains above roughly 2 points. Performance differed far more between the two models, with one quoting about 98-99% exactly and the other about 88-90%. The study covered only two 3B models and one prompt wording, so larger models may behave differently. The author suggests checking quotes after generation rather than relying on prompt wording.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Proxy toggle can't explain search result changes; compare run inputs first

A developer comparing search datasets considered rerunning with a different proxy setting to explain changed results. The article argues this is a flawed experiment, since every run uses a managed proxy regardless of the proxyConfiguration value, and result rows carry no proxy telemetry. It recommends separating changes in the submitted search, the observed results, and transport assumptions before spending on more runs. A table maps common symptoms, such as market mismatches, query coverage gaps, and result window differences, to the evidence that distinguishes them. It also suggests inspecting archived run manifests and confirming runs succeeded before drawing conclusions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How xterm.js Uses a WebGL Glyph Atlas to Render Multiple Busy Terminal Panes

A developer's article explains why running several AI coding agents in split terminal panes places a heavy rendering load on the browser. It describes how xterm.js, the terminal component used in VS Code, batches screen redraws to one per animation frame and only draws visible rows. It contrasts the DOM renderer, which creates many spans and triggers costly layout work under heavy colored output, with the WebGL renderer, which draws each glyph once into a texture atlas and reuses it. The piece also includes a small measurement page so readers can test rendering performance on their own machines.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Guide Shows How to Build a Rust Client for Detecting Structural PDF Tampering

A tutorial explains how to integrate a PDF tamper detection REST API into a Rust service using reqwest, serde, and related crates. The API checks a file's structure and metadata for signs of post-creation editing, without verifying the sender's identity or the document's authenticity. The client exposes two endpoints: a POST that submits a PDF URL and returns a check ID, and a GET that returns an intact, modified, or inconclusive status. The guide recommends sending idempotency keys and retrying only specific transient errors, and it notes that the API flags every verdict as unsuitable for automated adverse decisions. Equivalent integrations are described for several other languages and frameworks.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tips for Solo Indie Developers Facing Tight Time and Budgets

The article advises indie creators to start with a small, testable core feature set that can be shipped quickly and refined through early feedback. It recommends breaking large goals into micro-tasks under two hours each to maintain visible progress and motivation. To control costs, it suggests relying on free tiers such as static hosting, serverless functions, and self-hosted open-source CI, keeping monthly expenses low. Simple automations like pre-commit hooks, nightly builds, and changelog generators are presented as time-savers that pay off over several weeks. The piece also argues that a small, engaged user community offers more useful feedback than a large audience, and encourages fixed work hours to prevent burnout.

0 comments Read more at DEV Community