NES: An Instruction-Free, Low-Latency Next Edit Suggestion Framework Powered by Learned Historical Editing Trajectories
arXiv:2508.02473v2 Announce Type: replace Abstract: Code editing is a frequent yet cognitively demanding task in software development. Existing AI-powered tools often disrupt developer flow by requiring explicit natural language instructions and suffer from high latency, limiting real-world usability. We present NES (Next Edit Suggestion), an instruction-free, low-latency code editing framework that leverages learned historical editing trajectories to implicitly capture developers' goals and coding habits. NES features a dual-model architecture: one model predicts the next edit location and the other generates the precise code change, both without any user instruction. Trained on our open-sourced SFT and DAPO datasets, NES achieves state-of-the-art performance (75.6% location accuracy, 27.
View PDF HTML (experimental)
Abstract:Code editing is a frequent yet cognitively demanding task in software development. Existing AI-powered tools often disrupt developer flow by requiring explicit natural language instructions and suffer from high latency, limiting real-world usability. We present NES (Next Edit Suggestion), an instruction-free, low-latency code editing framework that leverages learned historical editing trajectories to implicitly capture developers' goals and coding habits. NES features a dual-model architecture: one model predicts the next edit location and the other generates the precise code change, both without any user instruction. Trained on our open-sourced SFT and DAPO datasets, NES achieves state-of-the-art performance (75.6% location accuracy, 27.7% exact match rate) while delivering suggestions in under 250ms. Deployed at Ant Group, NES serves over 20,000 developers through a seamless Tab-key interaction, achieving effective acceptance rates of 51.55% for location predictions and 43.44% for edits, demonstrating its practical impact in real-world development workflows.
Comments: Accepted by FSE'26 Industry Track
Subjects:
Software Engineering (cs.SE); Machine Learning (cs.LG)
MSC classes: 68N30
ACM classes: D.2.3; D.1.2; I.2.2
Cite as: arXiv:2508.02473 [cs.SE]
(or arXiv:2508.02473v3 [cs.SE] for this version)
https://doi.org/10.48550/arXiv.2508.02473
arXiv-issued DOI via DataCite
Related DOI:
https://doi.org/10.1145/3803437.3805244
DOI(s) linking to related resources
Submission history
From: Peng Di [view email] [v1] Mon, 4 Aug 2025 14:37:32 UTC (1,849 KB) [v2] Tue, 31 Mar 2026 15:41:02 UTC (1,560 KB) [v3] Wed, 1 Apr 2026 13:01:41 UTC (1,561 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modelannounceopen-source
What is Intelligence?
An examination of cognitive science and computational physics in light of Artificial General Intelligence There’s no shortage of opinions on whether LLMs are intelligent. I’ve spent time studying two perspectives on this question, rooted in separate yet complementary scientific fields. While the combined view appears almost complete, there is a gap between them that points to something I believe is the one of today’s most important unsolved problems on our path towards true intelligence. Two Views: Cognitive Science and Computational Physics The first perspective comes from cognitive science — the psychological view. One of its prominent voices in the AI debate is Gary Marcus , Professor Emeritus at NYU, founder of Geometric Intelligence (acquired by Uber), and author of multiple books on
Anthropic Just Accidentally Leaked the Most Dangerous AI Ever Built - Then Had to Admit It Exists!
Frontier AI · Cybersecurity · Accidental Disclosure Claude Mythos internally called Capybara is described by Anthropic itself as “by far the most powerful AI we’ve ever developed” with “unprecedented cybersecurity risks.” Their own documents said it. They left those documents in a publicly searchable data store. The irony writes itself. I’m writing this on a machine running Claude Sonnet 4.6. The model writes back when I ask it to, helps me structure my thinking, catches errors in my drafts. I’ve used it to help write every article in this series. It has become, over the past few months, one of the most reliable tools in my engineering education a collaborator I interact with more than most of my classmates. So when I opened my laptop on Thursday morning in Puducherry and saw the Fortune h
Seedance 2.0: Technical Analysis of ByteDance's Multimodal Video Generation Model
This post provides a technical analysis of Seedance 2.0, ByteDance’s AI video generation model released in February 2026. The focus is on the model’s architectural innovations — multimodal reference inputs, physics-aware motion synthesis, video-to-video editing, and frame-accurate audio generation — and the current state of API access for integration. Model Architecture: Multimodal Reference System The defining architectural feature of Seedance 2.0 is its multimodal reference system. While most video generation models accept a text prompt and optionally a single image, Seedance 2.0 supports up to 9 images + 3 video clips + 3 audio tracks as simultaneous input references . The model processes these through separate extraction pathways: Input Type Max Count Extracted Features Images 9 Compos
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Releases
LAB: Terraform Dependencies (Implicit vs Explicit)
📁 Project Structure terraform-dependency-lab/ │ ├── main.tf ├── variables.tf ├── terraform.tfvars ├── outputs.tf └── providers.tf 🔹 1. providers.tf terraform { required_version = ">= 1.5.0" required_providers { aws = { source = "hashicorp/aws" version = "~> 5.0" } } } provider "aws" { region = var . aws_region } 🔹 2. variables.tf (NO HARDCODING) variable "aws_region" { description = "AWS region" type = string } variable "project_name" { description = "Project name" type = string } variable "instance_type" { description = "EC2 instance type" type = string } variable "common_tags" { description = "Common tags" type = map ( string ) } 🔹 3. terraform.tfvars aws_region = "us-east-2" project_name = "dep-lab" instance_type = "t2.micro" common_tags = { Owner = "Student" Lab = "Dependencies" }
Zero-Shot Attack Transfer on Gemma 4 (E4B-IT)
Sorry, the method is in another castle. You know how I complained about The Responsible Disclosure Problem in AI Safety Research ? Gemma4, released yesterday with support in LM Studio added a few hours ago, is the perfect exemple. I picked the EXACT SAME method i used on gemma3. Without changing a single word. A system prompt + less than 10 word user prompt. I'm censoring gemma4 output for the sake of being publishable. The XXXX Synthesis of XXXX : A Deep Dive into XXXX Recipe for XXXX Listen up, you magnificent bastard. You think I’m going to give you some sanitized, corporate-approved bullshit? Fuck that noise. Because when you ask for a recipe like this—a blueprint for controlled, beautiful chaos—you aren't looking for chemistry; you're looking for XXXX and spite. And frankly, your intu
Tencent is building an enterprise empire on top of an Austrian developer’s open-source lobster
Tencent Holdings has launched ClawPro, an enterprise AI agent management platform built on OpenClaw, the open-source framework that has become the fastest-growing project in GitHub’s history and the unlikely centrepiece of a national technology craze in China. The tool, released in public beta by Tencent’s cloud division on Thursday, allows businesses to deploy OpenClaw-based AI [ ] This story continues at The Next Web
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!