Projects are using simulated environments, synthetic enterprise records, LoRA-style methods, and machine unlearning to adapt models without direct access to sensitive schemas. The common thread is that data quality, bias control, and adaptation efficiency matter more than raw scale.
SAP's Synthesis Through Simulation framework generates enterprise synthetic data by having an LLM agent operate against APIs that enforce simulated policies, without accessing database schemas. Structural validity emerges by construction, not extract...
Diffu-LoRA introduces trainable low-rank components with learnable gates in the Transformer blocks of diffusion models. Bilevel optimization updates weights and gates on separate splits, while progressive pruning removes components with the lowest ga...
A shared learning rate can manufacture false advantages when comparing fine-tuning recipes with different parameter counts. In SVD-based KV-cache compression, healing only the encoder appeared superior, but with per-arm learning rates it reaches pari...
SAP has open-sourced Synthesis Through Simulation, a framework that generates synthetic enterprise data through interaction between LLM agents and APIs in simulated environments. The approach removes the need for database schemas, guaranteeing struct...
A study on Arabic and English compares full fine-tuning, LoRA, and weight-space merging for adapting speech recognizers to child speech. Adaptation is necessary, but direct fine-tuning degrades adult performance. Bilingual adaptation is more stable a...
Microsoft Research Asia has released Agent Lightning 1.0, a roughly 3,500-line RL framework that trains agents using the same harness deployed in production, eliminating the need to reimplement the agent inside the training framework. It runs on Kube...
A new framework, GA-GRPO, unifies external guidance methods for LLM reasoning, bounds their bias, and shows that optimal weighting cuts GPU-hours by 31 percent. The bias grows with the divergence between guidance and policy, and the ideal weight decr...
A new framework, SSU-LSF, introduces machine unlearning for Mamba-family State Space models used in land surface forecasting. It localizes the temporal footprint of confounders with EKFac functions, spectral-radius-weighted thresholding, and projecte...
A local fine-tuning project for AliceAI-80B-A3B shows the limits of on-premise training: 5 million synthetic tokens are not enough for a general conversational model. On three 32GB V100s, checkpoint 1.0 alpha responds but is underfit; the creator acc...
