Can per-user adapters remain consistent without drifting or leaking?
This explores whether you can give every user their own lightweight fine-tuned adapter on a shared base model and keep each one stable over time (no drift) without their learned traits bleeding into other tasks or users (no leaking).
This explores whether per-user adapters can stay stable and self-contained — and the corpus suggests the answer is conditional: adapters are a genuinely viable way to run millions of personalized models off one base, but consistency and isolation are separate engineering problems you have to solve deliberately, not free properties of the approach.
Start with why the idea is attractive. The corpus reframes PEFT adapters not as a training trick but as durable, portable behavioral state — a small "delta" that carries a user's learned preferences and can be swapped in on demand, letting one strong base plus millions of lightweight adapters stand in for millions of full models Can lightweight adapters replace millions of personalized models?. That same mechanism works at the trait level: an adapter touching every transformer layer with under 0.1% extra parameters can reliably steer personality without prompting, which is exactly the kind of consistent, architecture-level control a per-user adapter would need Can we control personality in language models without prompting?.
The leaking problem is the sharpest. When multiple fine-tunings touch overlapping weights, they interfere — your adapter for one task quietly degrades another. The corpus's most direct answer is structural: identify each task's core parameter regions, freeze them, and only merge the non-core parts geometrically. This explicit isolation consistently beats naive multi-task fine-tuning, and notably, just scheduling tasks at different times doesn't fix it — you need the structural separation Can isolating task-specific parameters prevent multi-task fine-tuning interference?. The lesson transfers to per-user adapters: isolation has to be designed into which weights each adapter is allowed to move.
Drift is the other half, and here the corpus warns against a false comfort. You might assume that pinning settings (zero temperature, fixed seed) gives you a stable, consistent adapter — but consistency of output is not the same as reliability; a deterministic setup just replays one draw from the distribution repeatedly, which can be a consistently wrong one Does setting temperature to zero actually make LLM outputs reliable?. Worse, keeping an adapter consistent fights against keeping it current. Deployed agents seem to need two adaptation timescales — fast skill injection and slower gradient updates — to keep learning without freezing Can agents adapt without pausing service to users?, and adaptive memory that continuously forms and prunes links via execution feedback outperforms fixed structures Should agent memory adapt dynamically based on execution feedback?. So "no drift" can't mean "never changes" — it means changing in the right direction.
The quietly surprising thread is what should go into the adapter in the first place. Personalization works best when profiles are built from a user's past outputs rather than their input queries — meaning the durable signal is style and preference, not semantic content Do user outputs outperform inputs for LLM personalization?. That reframes the whole consistency-vs-leaking tension: an adapter encoding stable stylistic preferences has a much smaller, better-bounded thing to keep consistent and contain than one trying to absorb everything a user ever typed. And separately, the corpus shows that for on-device personal agents, task success, privacy-compliant behavior, and preference reuse are statistically distinct capabilities — no model wins all three — so "doesn't leak" is a measurable axis you have to test for directly, not infer from the adapter working well Do phone agents succeed at all three critical tasks equally?.
Sources 8 notes
PEFT adapters function as durable behavioral deltas carrying learned user experience, enabling a single strong base plus millions of lightweight adapters to replace millions of full models—but only when scale-up, scale-down, and scale-out reinforce simultaneously.
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.
Research shows that identifying core parameter regions per task, clustering overlapping tasks, and freezing core parameters while geometrically merging non-core parameters consistently outperforms standard multi-task fine-tuning. Temporal task scheduling alone proves insufficient without explicit structural parameter isolation.
Fixed seeds and zero temperature replicate the same output repeatedly, but that output remains one draw from the model's probability distribution. McDonald's omega testing across 100 repetitions reveals that consistency does not equal reliability.
MetaClaw demonstrates that deployed agents require both rapid skill injection from failures (seconds, zero downtime) and slower gradient-based optimization during idle windows (minutes to hours). The two mechanisms reinforce each other, with better policies producing more informative failures and richer skills enabling higher-reward trajectories.
FluxMem demonstrates that adaptive memory topology—where links form, refine, and consolidate based on closed-loop execution feedback—consistently reaches state-of-the-art across three distinct benchmarks. Dynamic connectivity outperforms fixed retrieval by aligning abstraction and eliminating interference.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
MyPhoneBench demonstrates that task success, privacy-compliant completion, and saved-preference reuse are statistically distinct capabilities with no model dominating all three. Success-only rankings do not predict privacy or preference performance.