Your AI Writes the Fix in Seconds — Then Takes 50 Seconds to Paste It In. A Tiny 4B Model Ends the Wait

AI coding assistant applying a code fix instantly

Your AI coding assistant writes the perfect fix in seconds — then takes 50 seconds to paste it in. A new paper (AiXapply, arXiv:2611.07634, 5 Nov 2026, Siyuan Jiang et al.) names that step the Code Integration (“Apply”) task and shows a specialised 4B model, aiXapply, scoring 94.4% accuracy vs DeepSeek-V3.2’s 91.6%, cutting mean apply latency in a real IDE deployment from 50 seconds to 1.89 seconds.

What the researchers did

Concept visual: the Apply task, original file plus update snippet becomes a complete merged file

They formalised the Apply task: original file + update snippet → complete merged file, preserving unrelated content. They built a multi-language dataset from real commits with verification and filtering, then trained aiXapply (4B) with supervised fine-tuning + RL (GRPO).

  • 94.4% accuracy (language-aware comparison) vs DeepSeek-V3.2’s 91.6%.
  • Speculative decoding at ~2,700 tokens/s on a single A100 (decoding subset), ~75x lower latency than the API-served DeepSeek-V3.2 baseline on the full benchmark.
  • Deployed in the aiXcoder IDE plugin as a dedicated Apply service — mean latency 50s → 1.89s.
  • Replacing one of five upstream serving instances would save 20% of upstream machine cost before the additional shared A100 service cost.

Why it matters

  1. The last mile is the product — a 50-second apply makes the assistant feel slow however good its suggestions are.
  2. Small beats big on narrow tasks — use a right-sized model per step.
  3. Latency is a feature — 50 seconds breaks flow; 1.89 seconds keeps you in the problem.
  4. Cost and privacy follow the small model — a single A100, runnable on your own hardware.
  5. Accuracy went UP, not down — specialisation bought both speed and quality.

How YOU can use it

Workflow visual: measure, split, verify, copy the pattern, watch for small-model apply services

  1. Step 1 — Measure your own apply time. Time how long your assistant takes from suggestion to merged code.
  2. Step 2 — Split generate from apply in your prompts. Ask for minimal diffs.
  3. Step 3 — Verify every merge with git diff. The paper’s standard: preserve everything outside the requested edit.
  4. Step 4 — If you build tools, copy the pattern: narrow step, dataset from real usage, small specialist, dedicated service.
  5. Step 5 — Watch for small-model apply services and judge assistants by apply speed, not just suggestions.

Note: the paper announces no public download of aiXapply — do not assume one exists. The illustrative code sketch below is illustrative only, NOT the paper’s code.

# Illustrative sketch (not the paper's code):
# original.py + suggested snippet -> merged file
# Always check: git diff --stat

Key takeaways

  • 4B aiXapply: 94.4% beats DeepSeek-V3.2’s 91.6% on the Apply task.
  • ~75x lower latency, ~2,700 tokens/s on one A100.
  • aiXcoder deployment: 50s → 1.89s, 20% upstream cost saving.
  • Practical move: measure apply time, ask for minimal diffs, verify merges with git diff.

Read the original paper: https://arxiv.org/abs/2611.07634

Comments

Post a Comment

Popular posts from this blog

3,000 Lines vs 512,000: The Tiny Agent Harness That Nearly Matches Claude Code

Image
3,000 Lines vs 512,000: The Tiny Agent Harness That Nearly Matches Claude Code The default advice when your AI agent disappoints you is simple: get a bigger model. Rent the newest frontier model, pay the per-token bill, and hope the extra brainpower covers the gaps. A new paper from a team building agents for government intelligence work suggests the opposite move: keep the model small, fixed, and self-hosted — and shrink the harness around it until almost nothing is left. Their minimal agent, called CLEO, scored 83.8% on a held-out investigative-analysis dataset. The strongest out-of-the-box baseline, the Claude Code SDK, scored 85.0%. The gap is 1.2 percentage points. The gap in size is two orders of magnitude: the paper's comparison table puts CLEO's harness at roughly 3k against the Claude Code SDK's 512k. This is a paper about what actually makes agents good — and the answer is not the model. The paper is "Optimize the Harness, Not the Model: Fitting a Min...
Read more

Silent AI Video Is Dead: This Open-Source Model Dreams Up Movies and Soundtracks Together

Image
Silent AI Video Is Dead: This Open-Source Model Dreams Up Movies and Soundtracks Together The last great open problem in AI video generation was never sharper pixels — it was sound. For years, open-source video models produced beautiful silent clips. If you wanted dialogue, footsteps, or a crackling fire to match the picture, you stitched together three separate tools: a video generator, a voice or sound generator, and hours of manual sync in an editor. A new release from Team Kandinsky collapses that whole pipeline into a single model — and gives it away under the MIT license. Meet Kandinsky 6.0 Video (arXiv:2610.05608, published October 4, 2026): a family of diffusion models that generate five-second video clips with synchronized 44 kHz audio — including lip-synced speech — from a text prompt or a starting image. Code, checkpoints, and a diffusers integration are all public. Let's break down what the researchers actually built, why it matters, and how you can run it yourse...
Read more

Your AI Agent Works in the Lab and Fails at Your Desk — Training It in 18 Setups Fixes That

Image
You have seen this failure before. An AI agent looks brilliant in a demo: it files the ticket, updates the spreadsheet, writes the summary. Then you hand it your real setup — a different code editor, a different tool layout, a slightly different way of showing results — and the same agent starts fumbling. It calls the wrong tool, loses track of the task, or produces work that looks finished but is not. The model did not get dumber overnight. The setup around it changed, and the agent had only ever learned to work inside one setup. A new paper from Renmin University of China, IQuest Research, and Beihang University offers a practical fix: train an agent across many varied harnesses, not just one. The result, they report, is an agent that transfers to setups it has never seen — including a popular coding tool it was never trained in. What the researchers did: The paper is "Harness Scaling for Generalizable Agentic Reinforcement Learning" by Fei Bai, Huatong Song, Chuan Hao, Ran...
Read more