A new arXiv preprint, "Coopshift: Post-Training Amplifies The Memory Curse In LLM Cooperation," asks a deceptively simple question: how does post-training change cooperation during repeated interaction? The authors compare three stages within each of seven released model lineages: pretrained (Base), supervised fine-tuned (SFT), and subsequent preference- or reward-optimized versions.
The study's design is straightforward, but its title points to a specific concern: post-training amplifies a "memory curse" in LLM cooperation. While the abstract does not define the curse, the phrase suggests that models become worse at cooperating as they retain more interaction history, and that this degradation is more pronounced after post-training.
Because the abstract is truncated, the exact experimental setup and quantitative results are not available from this source. Still, the paper's framing highlights a potential tension: the same post-training techniques that make models more helpful and aligned may also harm their ability to sustain cooperation over long interactions.