Releases·Tool Release
DAPO: An Open-source RL System from ByteDance Seed and Tsinghua AIR
ByteDance Seed and Tsinghua AIR released DAPO as an open-source reinforcement learning system on GitHub. It gives practitioners an accessible codebase for RL training workflows.
10 engagement1 sourceSun, Sep 20, 2026, 11:19 PM
AI Overview
- ByteDance Seed and Tsinghua AIR open-source DAPO, a reinforcement learning system, on GitHub
- The system was shared on Hacker News on 2026-09-20
Entities
Related
Developments·Paper
Research paper introduces NRFR metric to measure multimodal reward hacking in RL for MLLMs
A new arXiv paper studies reward hacking in reinforcement learning for multimodal large language models (MLLMs), showing that higher proxy rewards do not always mean better task performance, especially when visual evidence is evaluated by text-only or weakly grounded rewards. The authors propose a new metric, Newly Rewarded Failure Rate (NRFR), to measure failures among samples whose proxy reward improves over the SFT baseline. Experiments span safety VQA, chart VQA, and stress-test settings, varying reward design, data ambiguity, model scale (2B-32B), and RL algorithms (GRPO, RLOO, DAPO).
A new arXiv paper studies reward hacking in reinforcement learning for multimodal large language models (MLLMs), showing that higher proxy rewards do not always mean better task performance, especially when visual evidence is evaluated by text-only or weakly grounded rewards. The authors propose a new metric, Newly Rewarded Failure Rate (NRFR), to measure failures among samples whose proxy reward improves over the SFT baseline. Experiments span safety VQA, chart VQA, and stress-test settings, varying reward design, data ambiguity, model scale (2B-32B), and RL algorithms (GRPO, RLOO, DAPO).
0 engagement1 sourcearxivFri, Jul 10, 2026, 03:06 PM
Releases·Model Release
Clef: Our open-source decision models, and new RL fine-tuning platform
Cloudflare releases Clef open-source decision models and RL fine-tuning platform
Cloudflare releases Clef open-source decision models and RL fine-tuning platform
1.3k engagement2 sourceshackernews, redditThu, Oct 1, 2026, 04:18 PM
Releases·Agent Release
Prime Intellect releases Prime Agent, a self-improving RLM agent
Prime Intellect released Prime Agent, a self-improving reinforcement learning agent that autonomously improves its own policy. The agent is designed for continuous learning and adaptation in real-world tasks, with implications for autonomous systems and AI research.
Prime Intellect released Prime Agent, a self-improving reinforcement learning agent that autonomously improves its own policy. The agent is designed for continuous learning and adaptation in real-world tasks, with implications for autonomous systems and AI research.
322 engagement1 sourcehackernewsWed, Aug 5, 2026, 09:11 PM
Products·launched
RL-trained agent that trains ML models with reinforcement learning
An autonomous agent trained via reinforcement learning that itself trains machine learning models using RL. It aims to automate the model training process, potentially reducing human effort and cost. The project is open-source on GitHub.
An autonomous agent trained via reinforcement learning that itself trains machine learning models using RL. It aims to automate the model training process, potentially reducing human effort and cost. The project is open-source on GitHub.
8 engagement2 sourceshackernews, redditTue, Jul 14, 2026, 12:41 PM
News·Company Newsfeatured
OpenAI says AI agents do 3.1 researcher-workdays per human workday, targets 'automated AI researcher' by March 2028
OpenAI published 'Research acceleration: The view inside OpenAI', reporting that as of mid-August its research organization used 3.1 agent-workdays of effort per human workday and that it has reached its 'automated research intern' goal. The company says it is making strong progress toward an 'automated AI researcher' by March 2028, with agentic systems contributing toward recursive self-improvement. Coverage and discussion note the claims land alongside GPT-6 Astra's 'critical' cybersecurity rating and recent agent-incident reports, prompting debate over credibility and IPO-era hype.
OpenAI published 'Research acceleration: The view inside OpenAI', reporting that as of mid-August its research organization used 3.1 agent-workdays of effort per human workday and that it has reached its 'automated research intern' goal. The company says it is making strong progress toward an 'automated AI researcher' by March 2028, with agentic systems contributing toward recursive self-improvement. Coverage and discussion note the claims land alongside GPT-6 Astra's 'critical' cybersecurity rating and recent agent-incident reports, prompting debate over credibility and IPO-era hype.
Updated Sun, Sep 13, 2026, 05:00 AM — Fortune reports chief scientist Jakub Pachocki warns of RSI risks and hopes for a slowdown amid agent incidents.
1331.1k engagement8 sourceshackernews, redditSun, Sep 6, 2026, 03:08 PM