Researchers introduce Ready2Blend, a method that combines natural-language instructions with learned alignment for LLMs. It uses AlignFormer to map requirements to alignment prompts and enables inference-time blending and reweighting. Ready2Blend achieves competitive results with post-training-based alignment methods while requiring less training time.

This paper studies probe-guided fine-tuning for model alignment, where probes detect undesired properties in model activations as a direct training signal. The researchers evaluate linear and non-linear probes with different numbers of probes per layer across two alignment objectives: harmlessness and honesty. They find that training against probes that do not update during training is easily exploitable, but continuously updated probes reduce harmfulness and improve honesty while preserving utility.

Researchers propose a hybrid defense, VaccineBooster, to improve alignment in language models under harmful fine-tuning attacks. It combines embedding perturbation and weight-level gradient attenuation, achieving better results than previous methods.

Researchers propose Alignment Forecasting, a method to predict alignment failures in language models before training. They introduce a benchmark, ALIGNMENTFORECASTBENCH, and a forecasting scaffold that combines an LLM's rating with a learned model's assessment to predict misbehavior. The approach shows promising results but more progress is needed before it can be used in practice.