Infinite-Parameter LLMs: Writing Live Interaction Into Model Weights

On September 16, 2026, Jinli Hu, Ross M. Clarke, Yichuan Zhang, and José Miguel Hernández-Lobato of the University of Cambridge published arXiv:2609.18842, proposing an LLM architecture where effective model weights are generated from live session data rather than retrieved from a fixed bank. The paper calls this the Infinite-Parameter LLM. The core mechanism is a compact hypernetwork that converts runtime interaction data into low-rank modulations of a frozen base model, with a Bayesian belief over the generator’s latent state updated online as each session proceeds. The stored parameter footprint stays fixed. The weights the model can effectively compile are, by design, unbounded.

This is a research proposal. There are no production deployments, no third-party replications, and the evaluation protocol exists but results are not yet independently verified. What makes the paper worth reading carefully is not that it solves a problem but that it names a structural limitation in every deployed LLM today and proposes a concrete mechanism to address it.

The Problem With Frozen Weights at Inference Time

The scaling laws that have driven LLM progress assume a sharp division: training is when the model learns, inference is when it does not. Mixture-of-Experts architectures have extended this — activating only a fraction of an enormous stored parameter bank for each token — but the bank itself is fixed after pretraining. Every fact, correction, or behavioral constraint that a user supplies at runtime is placed in the prompt or in a retrieval store, and re-read from scratch on every token generation step.

The paper’s abstract is precise about why this is a structural problem rather than just an engineering inconvenience: “A deployed model faces a different world, where much of the data that would make it more useful is not in its training set but in the live interaction it is currently handling, such as the facts a user supplies or the corrections they give. A conventional model cannot learn from this data, because its weights are frozen after training.”

Three concrete costs follow from this. First, computation is not amortized: information a user supplies in turn 1 of a conversation must be re-attended to in turn 10, and again in turn 20, at the same compute cost each time. Second, the context window is consumed by information that could, in principle, be parameterized. Third, token-level representation of knowledge does not generalize the way learned representations do. A user who describes a constraint five different ways across a session relies on the model re-attending to all five instances, rather than having learned the constraint once.

The Hypernetwork Mechanism

The Infinite-Parameter LLM’s core component is a hypernetwork: a compact auxiliary network whose job is not to produce outputs but to produce weight modifications for the base model. At runtime, the hypernetwork takes data from the current session and generates a low-rank modulation of the base model’s feed-forward weights. Those modulated weights — not the original frozen weights — execute the forward pass.

The mechanism that distinguishes this from prior work on hypernetworks and session-level adapters is the Bayesian update layer. Earlier systems that generated weights from context read the context once and produced a fixed modulation. The Infinite-Parameter LLM maintains a latent belief state — a probability distribution over the generator’s latent code — and updates that belief online as each new piece of information arrives in the session. The hypernetwork draws on the updated belief to re-derive effective weights, so the model’s computation changes across turns without any gradient step.

The analogy the paper draws is to MoE routing. MoE models get the effective capacity of a large parameter bank while paying the compute cost of a small active subset. The Infinite-Parameter LLM extends this: the parameter bank is not stored but generated from live data, making the effective bank unbounded relative to a fixed stored footprint. From the abstract: “The stored footprint stays fixed, yet the weights the model can compile are effectively infinite.”

What This Claims Over RAG and Long-Context Models

The paper identifies three specific advantages over keeping knowledge in the prompt or in a retrieval store.

The first is compute amortization. Information written into model weights as low-rank modulations does not need to be re-attended on every generation step. A user-supplied constraint that is parameterized costs compute once; a constraint kept in the context costs compute on every token generation for the remainder of the session.

The second is context window efficiency. A 1M-token context window is valuable, but it is finite. Information that can be parameterized frees that window for content that cannot. The paper does not claim that long contexts are useless — only that parameterized persistence is strictly more expressive than token-level persistence for the same compute budget, for information that can be learned.

The third is generalization. A user who corrects a model’s behavior in several different ways across a session benefits from a model that has learned the underlying pattern, not one that has memorized five separate surface-level corrections and must re-read each. The authors argue that parameterized representations generalize across phrasing variations in a way that exact token retrieval cannot.

The Evaluation Protocol

The paper specifies an evaluation protocol designed to test these claims against in-context learning and retrieval-augmented generation on tasks requiring persistent knowledge across turns. The protocol exists; results are reported in the paper. What does not yet exist is independent replication by groups outside the authoring team, and the paper was submitted to arXiv on September 16, 2026 — five days before this writing. The results should be read as the authors’ own measurement of their own system.

The Open Questions That Determine Whether This Scales

Three engineering questions determine whether the Infinite-Parameter LLM produces results at production scale that match its theoretical appeal.

The first is numerical stability. Low-rank modulation of weights is well-understood in the fine-tuning literature (LoRA, QLoRA, and related methods apply exactly this technique). The difference here is that modulations are applied online, repeatedly, in a session, without a training step that provides gradient feedback to stabilize updates. Whether iterative online modulations remain numerically stable at the parameter counts of frontier models — hundreds of billions of parameters, not the research scale of this paper — is not established.

The second is interference. Each online update potentially modifies the same low-rank subspace that earlier updates addressed. Whether cumulative updates within a session degrade existing capabilities — the model’s ability to perform tasks unrelated to what it has been “learning” — is the catastrophic forgetting question applied to inference rather than fine-tuning. The paper addresses this through the Bayesian update mechanism, which is designed to smooth belief updates, but the properties of this mechanism under adversarial or pathological session content are not tested.

The third is inference latency. A hypernetwork that generates weight modulations on every turn adds compute to the forward pass. At the parameter counts and serving throughputs required for production deployments, whether that overhead fits within latency budgets that users and operators require is not yet demonstrated.

What This Means for Engineering Teams

If this architecture scales, it affects application-layer engineering in a specific way. The engineering investment that today goes into session state management, context packing, retrieval pipeline design, and context compression exists because the model itself cannot retain learned state across turns. That investment is substantial: chunking strategies, embedding stores, re-ranking logic, context window budget management, and memory routing all solve the same underlying problem — the model’s weights are frozen and cannot absorb what the session teaches.

Teams building on large language model infrastructure invest significant engineering budget in these layers. Some portion of that investment would become unnecessary if the model itself could parameterize session knowledge. The architectural implication is not that retrieval and context management disappear — broad knowledge recall from a retrieval store is a different problem than session-level adaptation — but that the session-management layer could become dramatically thinner.

For teams working on domain-specific LLM systems, the Infinite-Parameter LLM architecture is worth tracking as a candidate for long-session applications where coherence across many turns matters more than broad knowledge recall: multi-session user assistants, autonomous agents that operate over hours or days, or systems that must adapt to domain-specific corrections without full fine-tuning cycles.

The practical posture for 2026: treat this as a research result that identifies a real structural limitation in current architectures and proposes a concrete mechanism. Watch for replications that test the three open questions above — stability, interference, and latency — at production parameter scale. Do not redesign retrieval infrastructure around it yet.

Key Takeaways

• The paper (arXiv:2609.18842, submitted September 16, 2026) proposes a hypernetwork that generates low-rank modulations of a frozen base model from live session data, so effective weights evolve during a session without gradient updates.
• A Bayesian belief over the hypernetwork’s latent state is updated online as each session turn proceeds, making the weight modulation sensitive to the full session history rather than a single context read.
• The claimed advantages over in-context learning and RAG are compute amortization, context window efficiency, and better generalization of session-supplied knowledge across phrasing variations.
• Three open questions determine scalability: numerical stability of iterative online modulations at frontier parameter counts, interference between cumulative session updates, and inference latency overhead of the hypernetwork at production throughputs.
• Results are the authors’ own measurements; independent replication has not yet been published as of the paper’s September 16 submission date.
• If it scales, it reduces the engineering investment required for session state management, context packing, and retrieval pipelines in long-session applications — but it does not replace retrieval for broad knowledge recall.

Work With Origins AI

Origins AI builds production AI systems for engineering teams. If your team is evaluating how emerging LLM architectures affect retrieval pipeline design or session state management in long-running AI applications, talk to our team.