ai alignment engineers

Abstract: How does post-training change cooperation during repeated interaction? Within each of seven released model lineages, we compare pretrained (Base), supervised fine-tuned (SFT), and subsequent preference- or reward-optimized (Aligned) checkpoints across seven social dilemmas. Copies of each checkpoint play one another for 200 rounds, with visible history limited to the last two or forty rounds. The memory curse refers to lower cooperation when agents can see more past rounds. Our central finding is that post-training amplifies this decline. Averaged across the six binary-action games, expanding visible history reduces cooperation at every SFT and Aligned checkpoint. At each post-trained stage, these declines exceed Base's in six of seven lineages. Most of this average amplification is already present at SFT. With long history, both post-trained stages also cooperate less than Base on average in each binary-action game. In the symmetric binary games, their fitted cooperation rates fall more steeply on average than Base's as per-round cooperation costs rise, benefits fall, or noncooperation becomes more rewarding. The difference in cooperation rates between Aligned and SFT also changes from one-shot decisions to repeated play. When given the same past records, their choices depend on whether the instructions describe a final round or possible future interaction. Post-training evaluation should therefore go beyond isolated choices to assess whether cooperation is sustained in repeated interaction and how it depends on payoffs and visible history.

Abstract: Language models are frequently aligned to human preferences via imperfect evaluators whose errors can be exploited during optimization. We train policies to search for comparisons on which an evaluator may reverse the underlying preference, providing candidates for additional supervision. Since human preferences are unavailable during policy training, we use disagreement among an ensemble of preference models as a proxy for prediction error. We formulate this search as a zero-sum Stackelberg game between two opposing policies, a Leader which tries to minimize disagreement, and a Follower which maximizes it. For a general class of finite contextual Stackelberg games, we establish equilibrium existence and, under regularization, uniqueness and a closed-form characterization. We introduce Stackelberg Mirror Descent-Ascent (SMIDA), a two-timescale algorithm with linear last-iterate convergence in the tabular setting, and develop a neural policy-gradient analogue. Experiments with a synthetic oracle support disagreement as an informative proxy for prediction error. On held out data the trained policies outperform reference policies and increase the fraction of comparisons that exceed a disagreement threshold and reverse the oracle's preference.

Abstract: LLM developers post-train their models to exhibit prosocial values and behavioral traits, which are enumerated in an alignment target. However, while recent post-training developments have yielded models that score highly on alignment evaluations, training models on sets of narrow behaviors still influences their behavior across unseen contexts and environments in unexpected ways. In this paper, we establish the task of alignment generalization prediction, i.e., predicting how fine-tuning a model to follow a given value changes its behavior across a wide range of held-out values. We conduct a large-scale analysis of alignment generalization effects across 66 values found in modern alignment targets, and benchmark representational techniques on the alignment generalization prediction task. We find that representations based on model activations when applying values in context significantly outperform methods based on textual descriptions of the values. Specifically, the best activations-based methods achieve correlations of 0.45 with our generalization matrix, compared with 0.05 from description-based baselines. We then show the applicability of representations that predict alignment generalization toward downstream tasks by using them to measure how similar the values in a multi-value alignment target are, which we find is significantly correlated with model robustness. Finally, we show initial evidence towards a shared, model-independent value space, which we use to develop the first taxonomy of LLM values grounded in empirical generalization dynamics. Our work demonstrates the importance of studying value generalization in LLMs and its application toward the more empirical design and training of model behavior.