SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Variance-Based Routing Improves Long-Context Memory in Efficient Language Models

0
·1 views

Researchers at Princeton, with a Google DeepMind collaborator, identified a weakness in memory-efficient language models: routing based only on average relevance can fail to distinguish a segment containing a precise match from one with only loosely related content. Their paper, posted October 27, 2026 on arXiv, introduces Merlin, which adds a second statistic, variance, to the routing decision. Each context segment keeps two compact summaries, a mean and a spread of its keys, allowing the model to detect sharply matching segments. The authors report gains of up to 13.5% on LongBench and 8.3% on RULER, outperforming baselines that use four times the memory.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Phone validation API results need distinct handling for valid, unknown and failed checks

A signup form that treats every phone-check result as pass or reject can wrongly block users during outages or on unanswerable questions. The article presents a dependency-free Python fixture with four branches (normal, unknown, provider_failure, otp) that returns fixed outputs for testing signup logic without credentials or network calls. The free Basic Lookup route does not return line type, ownership, reachability or OTP delivery, so the design must not rely on those fields. The proposed design maps valid results to continuation under the team's own policy, unknown results to review or an existing OTP step, and provider failures to bounded retries before review. Questions the route cannot answer are handed off to a separate OTP provider the team already operates.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer builds ReadYou to summarize Hacktoberfest AI challenge winners' projects

A developer found it hard to make sense of the projects submitted by Hacktoberfest winners, so they created ReadYou to produce summaries of them. The tool also lets users interact with the projects through a conversational interface. The project was shared on DEV Community as a submission to the Hacktoberfest Open-Source AI Challenge, Week 3. The post was published on October 25 and is tagged with several challenge and AI-related topics.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Canva Launches Brush, an AI Model That Edits Designs, Claiming 46% Lower Cost per Edit

Canva released Brush on 29 October, eight days after its Craft image-generation model, to handle refinements to existing designs. Built by Canva Research, the vision-language model is available now to Canva Pro subscribers within Canva AI, with wider rollout planned later. Canva reports it runs 69% faster and 46% cheaper per edit than its previous design agent, though the company has not published its test method or outside benchmarks. Brush works through an agentic loop that makes changes, inspects the result, and adjusts, while keeping every element editable by the user. The company says the aim is to apply requested changes without disturbing the rest of the design.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Audit of 80 Shopify and WooCommerce product pages finds gaps in shipping and return schema

A technical audit examined public HTTP responses from 100 online stores, of which 80 had readable product pages. Most of those pages lacked key structured data, with 78 of 80 missing shippingDetails and 76 of 80 missing hasMerchantReturnPolicy in their Product JSON-LD. The author notes that Offer fields like price and availability were generally present, while native files such as /llms.txt were largely absent unless manually configured. The study is explicitly limited to a non-random sample and does not mean stores lack shipping or return policies on their websites. The author also promotes a read-only tool, CitedCart, that lets store owners inspect their own markup.

0 comments Read more at DEV Community