Daily AI Paper

When people merge multiple fine-tuned language models into one, they often get a model that looks capable on paper but quietly drifts away from the behavior of each source model. This paper asks why that happens in decoder models, where generation is autoregressive and errors can accumulate token by token. The authors show that two things matter: the bias is position-dependent, and some positions are far more important than others because they carry high uncertainty and drive the final decision. Their method, DARTS, fixes this with a tiny amount of extra parameters by learning a per-position correction and weighting the repair more strongly where the model is most likely to matter. The result is a practical way to make model merging work better for code generation, math reasoning, and instruction following, without retraining from scratch. That matters because model merging could make multi-task LLMs cheaper and faster to deploy, if we can keep the merged model faithful to the originals.