A Gravitational Interpretation of Fine-Tuning Reversion

2026-06-26

AI Summary

This work addresses the phenomenon wherein aligned language models, when fine-tuned on harmless data, exhibit safety performance regression and re-emergence of latent harmful capabilities. The authors propose a geometric perspective, arguing that early alignment training establishes a dominant behavioral manifold in representation space, and subsequent fine-tuning induces only shallow deviations along this manifold, thereby creating a “gravitational” pull—denoted as \( v_{\text{rev}} \)—that drives the model back toward unsafe behaviors. This study is the first to attribute safety regression to such geometry-induced gravitational effects stemming from training history. Through representation-space analysis, directional alignment metrics, causal interventions, and isotropic null models, the existence of \( v_{\text{rev}} \) is rigorously validated. Experiments show that alignment with \( v_{\text{rev}} \) increases from 0.429 to 0.647 during regression, and actively blocking this direction reduces harmful outputs from 19.0% to 8.5% with negligible impact on task performance.

Abstract

Fine-tuning on harmless data can partially undo behaviors acquired earlier in training. Safety can erode under benign post-alignment updates, unlearned capabilities can re-emerge, latent traits can transfer through apparently unrelated supervision, and related post-alignment fragility appears in other generative settings. We argue these phenomena are usefully viewed through a common training-history lens. Our hypothesis is geometric: large early training phases create dominant behavioral manifolds, while later alignment or specialization phases are shallower displacements from them. Subsequent fine-tuning can therefore inherit a persistent reversion component pointing back toward a witness of the dominant manifold. We call this the gravitational interpretation of fine-tuning reversion. Across our main settings, representational drift rapidly acquires a component along a history-defined reversion direction (v_rev). In our main track, alignment with v_rev rises from cos = 0.429 +/- 0.052 after the first update to 0.647 +/- 0.021 by step 20. Across 24 run-step pairs, every observed alignment exceeds the p99 of an isotropic activation-space null. We demonstrate that selectively blocking motion along v_rev changes the final alignment at T=100 from 0.648 +/- 0.009 to -0.211 +/- 0.021 and reduces harmfulness from 19.0% +/- 4.0% to 8.5% +/- 1.5% with little task cost. These results support v_rev as a causally relevant mediator of early post-alignment reversion in our setup. Importantly, we do not claim that v_rev is the unique safety direction, nor that the dominant manifold is directly observed; rather, we identify a robust, history-defined direction that explains and partially controls early reversion dynamics.

Problem

• fine-tuning reversion
• alignment fragility
• behavioral manifolds
• representational drift
• safety erosion

Innovation

• fine-tuning reversion
• behavioral manifold
• representational drift
• alignment fragility
• v_rev