A new study pits two fixes for AI sycophancy against each other, and only one earns its keep.
Researchers tested large language models that use long-term memory to personalize answers, a feature that can backfire when stored user beliefs override facts. They built a defense-in-depth setup comparing internal "steering" of the model's activations against external filtering of retrieved memories, including a Router Gate that keeps, rewrites, or drops each memory item. Across four open-weight models and 1,550 benchmark items judged by three separate LLM judges, the Router Gate preserved far more accuracy than simply wiping all memory. On Llama 3.1 8B, nudging the model's internal activations away from sycophancy did lower the sycophancy score slightly, from 35.80% to 31.32%, but the researchers found the change was not statistically significant.
That matters because AI assistants are increasingly sold on remembering you - your preferences, your past arguments, your opinions. This research suggests the fix for a memory system that flatters rather than informs isn't tinkering with the model's internals, it's being smarter about what gets retrieved in the first place.
It's a modest, unglamorous finding, but an honest one: the fancier internal fix added nothing the data could confirm, while the boring filter did the actual work.