A new paper on arXiv challenges the dominant approach to aligning language models with human values. The authors note that while human values shape how we expect models to behave, most alignment techniques treat values as a single, global objective. They argue this is a mismatch, because the appropriate value orientation is personal and varies by context.
The paper focuses on multi-turn conversations, where the give-and-take of dialogue makes value alignment especially dynamic. Rather than relying on additional training, the authors propose a training-free alignment method that can steer model behavior during inference. The abstract is truncated, so full details of the method are not available from the source alone.
If the premise holds, it suggests that alignment should be more adaptive—responding to the user's values and the situation—rather than assuming one universal standard works for everyone. The preprint is available on arXiv under ID 2610.19817.