Replacing the fixed reference policy in DPO with a moving geometric anchor, because the reference drifts out of calibration
AspireAI community · arXiv
Direct Preference Optimization and related methods align large language models from pairwise preferences by regularising updates against a fixed reference policy. As the policy drifts, a static reference becomes increasingly miscalibrated, which creates distributional mismatch and can amplify spurious preference signals under noisy supervision. Reference-free variants avoid the mismatch but then risk unconstrained reward drift. GAPO, Geometric Anchor Preference Optimization, replaces the fixed reference with a dynamic, geometry-aware anchor: an adversarial local perturbation of the current policy within a small radius that serves as a pessimistic baseline. The anchor enables an adaptive reweighting mechanism that modulates the importance of each preference pair based on its local sensitivity. The paper introduces the Anchor Gap, the reward discrepancy between the policy and its anchor, and shows under smoothness conditions that it approximates worst-case local margin degradation, so optimising a logistic objective weighted by that gap downweights geometrically brittle instances while emphasising robust preference signals. Across diverse noise settings GAPO consistently improves robustness while matching or improving performance on standard LLM alignment and reasoning benchmarks. The work was accepted at NeurIPS 2026. The limitation is that the anchor's radius is a real hyperparameter with a genuine accuracy-versus-drift trade-off, and the evidence comes from the noise settings and benchmarks the authors selected.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.