DREAM method trains dense retrievers using frozen LLMs, outperforms contrastive baselines

·1 views

A new retrieval training method called DREAM derives supervision signals directly from a frozen large language model's next-token prediction objective, bypassing the need for manually constructed positive and negative document pairs. Unlike widely used dense retrievers such as DPR, ANCE, and ColBERT, DREAM avoids expensive data mining pipelines that typically require billions of training examples. The approach works by injecting retriever scores into selected attention heads of a frozen LLM, tightly coupling the retriever with the model's attention mechanism. Tested across backbone sizes of 0.5B, 1B, and 3B parameters, DREAM consistently outperformed existing baselines on the BEIR and RTEB benchmarks. If validated at scale, the method could eliminate the contrastive data construction stage from standard retrieval-augmented generation pipelines.

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)

How Embeddings Are Replacing Traditional Indexes in the AI Search Era

For decades, databases used indexes to locate structured information quickly by matching exact words or values. Traditional indexes struggle when users phrase queries differently from how information is stored, such as asking 'how do I recover my account?' when documentation says 'credential reset procedure.' Embeddings solve this by representing text as mathematical vectors, placing semantically similar content close together regardless of wording. This allows AI-powered search systems to retrieve relevant results based on meaning rather than exact keyword matches. Embeddings now underpin major AI applications including retrieval-augmented generation, enterprise search, recommendation engines, and conversational AI memory.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Speculative Decoding Can Speed Up LLM Inference by 20-50% Without Changing Output

Speculative decoding is a technique that accelerates large language model inference by generating multiple tokens per forward pass instead of one at a time. Standard autoregressive generation is bottlenecked by memory bandwidth, as a 70B model on an H100 GPU produces only one token every 30-50 milliseconds due to sequential dependencies. The method works in draft-verify cycles, where a lightweight draft mechanism proposes several candidate tokens and the target model verifies all of them in a single forward pass. Accepted tokens are statistically identical to those produced by standard decoding, meaning output quality is preserved. The approach is supported across popular inference frameworks including llama.cpp, vLLM, SGLang, and TensorRT-LLM, with multiple implementation variants available as of 2026.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How One Frontend Team Built Visual Regression Tests for 40+ UI Components

A frontend team managing a React 19 monorepo library of over 40 real estate card UI components faced recurring layout breakages caused by shared style changes that went undetected until release or production. To address this, they implemented a visual regression testing workflow combining Storybook, Playwright, and Jest Image Snapshot to photograph and compare component renders pixel by pixel. Each UI variant is defined as a Storybook story, which the test runner renders in headless Chrome and compares against a baseline image. The team opted against traditional DOM snapshot tests, citing their inability to catch visual issues like z-index overlaps or flex-wrap failures. Two npm scripts were added to streamline baseline creation and test execution, enabling the full suite of 40-plus components to be tested in under a minute.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

User Scanner: Open-Source OSINT Tool Gains Traction for Identity Reconnaissance in 2026

A command-line OSINT tool called user-scanner, developed by kaifcodec, has attracted attention in the security research community for its approach to identity reconnaissance. The tool supports over 295 scan vectors spanning more than 110 email-integrated sites and 185 username platforms, allowing analysts to run email, username, or combined scans from a single command. Unlike many free tools that rely on basic HTTP status checks, user-scanner targets signup endpoints and parses JSON responses to extract metadata such as account creation dates, avatar URLs, and privacy settings. An optional flag integrates with Hudson Rock threat intelligence to cross-reference targets against infostealer malware logs at no additional API cost. The tool also includes proxy validation and Nix shell support, and is noted for reporting failures transparently rather than returning false positives.

0 comments Read more at DEV Community