LoRA adapters, deletion tuning, energy-based weight edits and PACT analyses point toward model updates without full retraining or duplicate weights. This modular approach reduces duplication and gives self-hosted users finer control.
A new study shows that zeroing fewer than 0.05% of a pretrained LLM's weights, with no other modification, matches full fine-tuning across three post-training regimes: reinforcement learning with verifiable rewards, supervised fine-tuning, and on-pol...
An open-source project shows how to transfer a new capability into the weights of Qwen 3.5 0.8B using an energy-based decision model. The author compares the result to LoRA and reports a 3-millisecond intervention, without catastrophic forgetting. Th...
The proposal to distribute LoRA adapters instead of entire modified LLM copies signals a pivotal shift towards modularity in the AI ecosystem. For organizations managing self-hosted infrastructures, this evolution promises a significant reduction in ...
A Reddit proposal suggests distributing only abliteration LoRAs instead of full uncensored LLM versions, cutting duplication and costs. The technical point: one adapter works across every model of the same architecture and every quantization. For sel...
NousResearch post-trained Alibaba's Qwen3.8-27B using only supervised fine-tuning, without RL or architectural changes. The resulting model, Arete-27B, cuts propaganda answers from 22% to 0.9% and raises the fact score from 2.32 to 2.98 out of 5, wit...
PLatour shows that part of the gains in PACT systems appears even with non-informative senders or when only the receiver is trained. PACT-Check measures the real contribution of inter-model communication. For on-premise deployments, the difference be...
An open-source framework challenges the PACT premise: passing latent representations between models can improve benchmarks without real information transfer. PLatour adds trainable zero-initialized bridges and PACT-Check, a method to separate genuine...
