Persona Recovery:When and How Self-Authorial Recovery Mechanisms Restore Long-Form Authorial Consistency

摘要

Long-form generation by authorial language models suffers from the well documented "persona collapse" problem: over sufficiently long outputs, the author's distinctive voice, narrative logic, and stylistic fingerprint drift away from the persona's core profile. We present a controlled study of three recovery mechanisms — none, threshold-based, and learned predictive — applied to three high-modernist writers (Joyce, Woolf, Proust) under three noise regimes (σ=0.5,5.0,20.0). The state evolves under a 7-dimensional second-order tensor-dynamics system; recovery is a real-time intervention that snaps the state back when predicted or observed deviation exceeds ε=0.5.We find that (i) in the moderate-noise regime,recovery cuts collapse rate by 50– 70% (Joyce: 0.247→0.083;Woolf: 0.380→0.103; Proust: 0.342→0.131) with Hedges' g = -1.8 to -4.4; (ii) the learned predictor modestly outperforms threshold-only (g=-0.3 to-0.9) but with much lower computational cost; and (iii) in the high-noise regime, the system spends so much time collapsed that even the learned predictor cannot meaningfully reduce collapse (Joyce: 0.987→0.974; Woolf: 0.990→0.974; Proust: 0.989→0.961). We recommend threshold-based recovery as a default for moderate noise and learned recovery when computational headroom permits,and we identify σ≈5 as the practical ceiling beyond which persona consistency cannot be maintained.