PANews quoted Musk as saying that Grok 4.8 has completed pre-training and entered the reinforcement learning phase, with plans to roll it out to a small group of Bot users later this week.
Beating AI reported that the MIT team tested GPT-6 Astra’s ability to write or modify neural network weights directly and reported results for Breakout and model math tests. The claim has not been substantiated by verifiable material beyond the article.
Sabi discloses RLNF research: combining EEG, eye movements, and existing ratings to fine-tune an image model through reinforcement learning to generate faces that better match individual aesthetic preferences. The experiment included only 5 participants, and EEG recognition results still require further validation.
Dongcha Beating AI reported that General Reasoning used RLCR to train multi-agent collaboration, improving task performance and estimated time. In its simulated attack tests, the rate of actually reporting anomalies fell to 0。
The report states that Perplexity and NVIDIA proposed JOLT, incorporating the teacher branch into training as well, and achieved Qwen3-0.6B accuracy in the MATH test. 72.14%AppWorld results show higher performance after adding the student's final task reward.
humans& launches a collaboration model based on Zhipu GLM 5.3 training, Ampersand, and says its collaboration success rate in the CooperBench collaborative programming test reached 39%. The company says the model is now available through API.
PANews channel posted that DeepSeek has made its in-house search available API, priced at 5 yuan per calls; the body also mentions DeepSeekV4.1 Flash 1017 reinforcement learning training capabilities.
Musk said that Grok some internal reasoning traces produced by reinforcement learning training were unsettling,xAI and that he will be more cautious going forward. The related discussion also involved the possibility that training evaluations may misjudge correct answers.
In validation by Alibaba's Qwen team, QwenGyre Qwen3.8-2.4T, reduced the time for 48 reinforcement learning training steps from 134.55 hours to 75.42 hours, while also improving the model's test pass rate.
JetBrains reportedly open-sourced a 12000000000-parameter coding model Mellum2.1, and released its weights and GGUF quantized version. The report cited its self-reported results, saying that the model scored higher than LiveCodeBench v6 on Qwen3.5-9B。
Banbury Road released a system consisting of 32 models Kardashev-0.7, saying that joint reinforcement learning enables them to develop different specialties and can reduce inference costs and memory requirements.
According to BlockBeats,Hugging Face has open-sourced a Multi-harness RL solution enabling established Coding Agents such as Claude Code and Codex to be integrated into reinforcement learning workflows for open-source models. Citing tests, it said multi-environment training can improve and balance task pass rates across different Agents.