Skip to content
NewA calmer, faster ZeroHour

I2C-GRPO trains a Qwen3-4B agent for DDoS alert triage

Original title: I2C-GRPO: Insight Importance Checklist for Reinforcement Learning in a Multi-Agent DDoS Alert Triage Workflow

arXiv cs.CRby Eyal Zloof ()1 min read1
AI summary · x-ai/grok-4.7Summaries by Vani AI

Across three seeds the agent reaches mean F1 0.869 and a 4.9% false-positive rate on an independent trace.

The paper proposes I2C-GRPO, a checklist reward for intermediate investigation steps when labels are missing and outcome rewards are sparse. Expert LLMs propose actions, atomic forensic insights are ranked by a three-model panel, and a reward model scores coverage without executing the actions. An open-weight Qwen3-4B Triage Agent is trained inside a fixed multi-agent DDoS alert-triage workflow on DARPA 2000 LLDOS 1.0 and tested on CSE-CIC-IDS2018. The authors limit all claims to DDoS alert triage and say code, task contexts, and the ranked insight dataset will be released.

Unconfirmed: Reported gains are limited by the authors to DDoS alert triage.

Key facts
  • Mean F1 is 5.9 and 9.7 points above uniform insight weights and DIVA.
  • False-positive rate is 4.9%, versus 16.4% for GPT-5-mini.
  • False positives fall only at the highest-orness weighting evaluated.
  • Training reduces to a one-step contextual bandit over fixed prefixes.
  • Test data is an independent CSE-CIC-IDS2018 DDoS trace.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at arxiv.org

Large language model (LLM) agents are increasingly used for investigative tasks over incomplete evidence, such as cybersecurity alert triage, but reinforcement learning (RL) for such tasks is difficult: intermediate steps have no verifiable labels, rollouts are costly, and outcome rewards are sparse. We propose Insight Importance Checklist (I2C)-GRPO, a checklist reward for intermediate investigation steps.

arXiv cs.CR

By Eyal Zloof · arxiv.org

Read the full article at arXiv cs.CR

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.