Since you all asked for this list: – 25 Feb, SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution, https://
arxiv.org/abs/2502.18449
– 24 Feb, Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment,
@rasbt
-
Latest AI Papers: SWE-RL and LoRA Improvements
By
–
-
Token increase, parameters, and training methods improvements
By
–
sure, but due to more tokens, more parameters, new training methods, or all of the above 😛
-
New Supervision Techniques Combined with SFT and RLHF Training
By
–
So, it's some new sort "supervision techniques" > "We trained it using new supervision techniques combined with traditional methods like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to those used for GPT-4o."
-
Rolling dice to randomly select an LLM for daily use
By
–
Soon I'll need to roll one of these 20-sided dice to decide which LLM to use on a given day
-
AI Model Releases Accelerate: Too Much Progress in One Week?
By
–
Can there be too much of a good thing? The 7-day view so far:
– Grok 3 was released
– Claude 3.7 was released
– DeepSeek shipped some awesome stuff in their open-source week
– DeepSeek announced to release R2 soon
– OpenAI made Deep Research free for Plus users
– Google is made -
Claude 3.7 Model: Clarifying Distillation vs RL Training Scaling
By
–
> – Claude 3.7 is the new o3-mini-high
Wait, 3.7 is a distilled model? I thought the system card said it was RL-trained plus it uses inference time scaling? -
GPT4.5 Technical Details and OpenAI’s Lack of Documentation
By
–
GPT2 and GPT3 were the good old times with technical reports. GPT3.5 had at least the InstructGPT report. GPT4 had strong bigger+MoE rumors.
Does anyone know what's new in GPT4.5? -
Question about dataset consistency and paper link issue
By
–
Thanks for sharing! Are those based on the same dataset as described in the Phi-4 paper in December? I really appreciated that one for all the insights into the synthetic data usage. PS: Unfortunately, the paper link in your post doesn't seem to work.
-
Running efficiently on a 2020 MacBook Air
By
–
Haha you never know . I am running this in a 2020 MacBook Air so efficiency is key.
-
AI Lowers Development Barriers for Custom Applications
By
–
That's true. I used to do that for years, where I write my own command line tools. I never developed smartphone apps etc though because of the overhead. I can see though how AI can lower the barrier here to make certain custom apps.