Engineering
Memory Tax: Dissecting the Net Value of Agent Memory
The authors evaluate representative agent memory systems against a naive retrieval-augmented generation baseline that searches raw conversation history directly, across…
Engineering
RealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents
RealCLI is a paired benchmark that scores both task execution and human-agent collaboration quality on realistic command-line tasks, using a…
Engineering
Deep Fusion Memory: Giving LLMs External Knowledge Without Retrieval at Inference
Researchers have proposed Deep Fusion Memory (DFM), a technique that lets large language models draw on external knowledge without the…
Engineering
Redesigning Source Control at Anthropic
Anthropic has published details of Scion, an in-house source control system it built to replace Git for its own engineering…
Engineering
What Git Does Not Record: Provenance Legibility in Two AI-Assisted Research Repositories
A new study audits two single-investigator research repositories to examine how well git history captures the involvement of AI coding…
Engineering
Baidu Open-Sources Unlimited OCR 2 With Sparse Reference Sliding Window Attention
Baidu released Unlimited OCR 2, a 3.34-billion-parameter open-weight model under the MIT license for parsing very long documents. Its Sparse…
Engineering
NVIDIA’s Nemotron 3.5 Embed, Rerank, and Parse Push Agentic Retrieval Benchmarks
NVIDIA released a new Nemotron 3.5 retrieval suite: Embed (8B and 1B variants, plus a 1B NVFP4 quantised version), a…
Engineering
LASAGNA Study Finds RAG Techniques Don’t Reliably Combine, Even When Each One Works Alone
Researchers introduce LASAGNA, an open-source framework for controlled evaluation of Retrieval-Augmented Generation pipelines, testing 450 configurations across eight benchmarks with…
Engineering
Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
Researchers tested whether complex agentic memory systems are needed for long-term conversation understanding by comparing them with Raw-RAG, a simple…
Engineering
Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents
Researchers introduce Doc2Harness, a two-stage benchmark that directly evaluates how well AI models can build the software harness surrounding a…
Engineering
SWE-at-Scale: Scalable Issue Synthesis From Historical Bug Fixes
Researchers have introduced SWE-at-Scale, a framework that generates realistic software engineering training tasks for coding agents by reverse-engineering historical bug…
Engineering
ColGrep-Agent: Cutting Frontier Coding Agents’ Tokens by a Third
LightOn AI has introduced ColGrep-Agent-2B, a locally running 2-billion-parameter model that sits between a coding agent and a code repository,…