@rohanpaul_ai: A 4B coding agent reached 61.5% on SWE-bench Verified without frontier-model distillation by combining a simpler tool i…
Summary
FrogNano is a 4B coding agent trained via online task synthesis that achieved 61.5% on SWE-bench Verified, using a simplified tool interface and adaptive synthetic tasks.
View Cached Full Text
Cached at: 09/21/26, 01:33 PM
A 4B coding agent reached 61.5% on SWE-bench Verified without frontier-model distillation by combining a simpler tool interface with synthetic tasks that keep adapting to what the model can currently learn.
FrogNano starts from Qwen3.5-4B and is trained with RL on about 1,500 synthetic software-engineering tasks.
The important part is how those tasks are chosen.
As the model improves, the system generates fresh problems that are challenging but still learnable, so the curriculum improves with the agent.
The interface matters just as much.
Switching to a simpler 5-tool setup moved the base model from 8.3% to 37.2% on SWE-bench Verified.
After 5 rounds, FrogNano reached 61.5%.
– arxiv. org/abs/2609.07925
Title: “FrogNano: Training a 4B Coding Agent via Online Task Synthesis”
Similar Articles
Microsoft trained a 4B coding agent almost entirely with Reinforcement Learning, without a bigger teacher
Microsoft Research presents FrogNano, a 4B coding agent trained exclusively via reinforcement learning with online task synthesis, achieving competitive performance without distillation from larger models.
I built a coding agent that gets 87% on benchmarks with a 4B parameter model, here's how
The author built SmallCode, a coding agent optimized for small local models, achieving 87% benchmark success with a 4B parameter model using techniques like compound tools, improvement loops, and token budgeting.
@xdotli: my friend @xeophon thinks coding is solved here's validation that a 3b model is trained with focus on algo efficiency a…
Nanbeige 4.1, a 3B model, outperforms Qwen3-30b-A3b and Qwen 3.5 4b in coding tasks with focus on algorithmic efficiency, achieving long horizon tasks with 600+ tool calls.
@rohanpaul_ai: Meta paper shows that coding agents get much better when they reuse short summaries of past attempts instead of raw log…
A Meta paper shows that coding agents improve significantly when they reuse short summaries of past attempts instead of raw logs, achieving strong gains on SWE-Bench and Terminal-Bench with Claude 4.5 Opus.
@IntologyAI: Can coding agents do research? We release NanoGPT-Bench, an internal eval we’ve used to test agents on an AI R&D proble…
IntologyAI releases NanoGPT-Bench, an internal benchmark to evaluate coding agents on AI R&D tasks. Current agents recover only 9.3% of human progress, mostly through hyperparameter tuning, highlighting gaps in algorithmic research capabilities.