Can separate teams each build their own small add-on for one shared AI model, without breaking each other?
Can different teams train specialist adapters that share one frozen base model?
This explores whether separate teams can each build their own small specialist add-on (a LoRA-style adapter) for one common frozen base model, and whether those pieces can coexist.
This explores whether separate teams can each build their own small specialist add-on for one common frozen base model, and whether those pieces can coexist. The corpus supports the 'one base, many adapters' half directly. It has no note on independently trained adapters being combined, so for that half the closest evidence comes from neighboring work.
The sharing half is the strongest. Can lightweight adapters replace millions of personalized models? treats adapters as durable behavioral deltas, so one strong base plus millions of lightweight adapters can replace millions of full models. That note is about personalizing for users rather than coordinating teams, but the economics are the same: the base is expensive and shared, and the adapter is cheap and owned by whoever trained it. The catch is that it only works when scaling up, down and out reinforce each other. Can editing hidden representations beat weight updates for finetuning? suggests each team's deliverable could be even smaller. It edits frozen hidden representations instead of weights, and its low-rank variant beats LoRA with 10-50x fewer parameters.
The independent-training half has a useful analogue. Can asynchronous expert training beat synchronized distributed LLM training? trains domain experts in parallel with no synchronization, then merges their feed-forward layers as mixture-of-experts components and learns token-level routing. That beat both synchronized training and routing-free merging. The lesson is that teams who never talk can still combine their work, as long as something learns to route between the pieces afterward. Plain merging leaves performance on the table. That note describes merging feed-forward layers, though, not stacking adapters on a frozen base, so it shows the workflow can succeed but not that adapters compose the same way.
The main risk is interference. Can isolating task-specific parameters prevent multi-task fine-tuning interference? finds that tasks trained into one model step on each other. What works is identifying each task's core parameters, clustering tasks that overlap, freezing the cores and geometrically merging the rest. Simply scheduling tasks in a better order is not enough. That is a single team with full visibility of every task, whereas separate teams would have to find their overlaps after the fact. Independent adapters would likely need a deliberate structure to avoid colliding, not just a shared base.
A looser version of the pattern keeps the specialist entirely outside the model. Can a separate trained curator improve skill libraries better than frozen agents? trains a curator separately from a frozen executor, and the curator carries over to different executor backbones. Can a separate memory model inject knowledge without touching the LLM? does the same with a memory model that works with frozen proprietary models, at the price of upfront training and limited capacity. But Can an external manager handle context for frozen agents? shows the fit is not universal: stronger agents want high-fidelity context, and weaker ones want aggressive compression. A specialist built for one frozen base is quietly tuned to that base's strengths, so the shared base works as a contract only while everyone agrees on which base it is.
Sources 7 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.
ReFT learns task-specific interventions on frozen model representations rather than updating weights, with LoReFT (low-rank linear subspace variant) dramatically outperforming LoRA across reasoning, instruction-following, and NLU benchmarks while using far fewer parameters.
Branch-Train-MiX trains domain experts in parallel without synchronization overhead, merges their feed-forward parameters as MoE experts, and learns token-level routing, achieving better accuracy-efficiency tradeoffs than synchronized training or routing-free merging.
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.
SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.
Show all 7 sources
MeMo trains a dedicated memory model to encode new knowledge, eliminating inference-time search costs that scale with corpus size. It avoids fine-tuning risks and works with frozen proprietary models, but trades this for up-front training cost and capacity limits.
AdaCoM trains an external RL-based manager to prune and preserve context for frozen agents. The key finding: stronger agents benefit from high-fidelity preservation, while weaker agents need aggressive compression—optimal context management is agent-specific, not task-universal.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- ReFT: Representation Finetuning for Language Models
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Compositional Skill Routing for LLM Agents: Decompose, Retrieve, and Compose
- Context-PEFT: Efficient Multi-Modal, Multi-Task Fine-Tuning
- Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMs
- SkillOS: Learning Skill Curation for Self-Evolving Agents
- MeMo: Memory as a Model
- Learning Agent-Compatible Context Management for Long-Horizon Tasks