October 19, 2026
Technology Watch

From the laboratory to industry: following new technology on its way

← Front page

arXiv

Personalized embeddings learn without chat histories

Image drawn by AI

A compact user memory could let personalized assistants learn preferences without carrying full chat histories, but the key risk is trusting the model’s own guesses about satisfaction. The authors of COPE gave 256 simulated users learned embeddings and trained a language model with sparse feedback. In a kitchen, dinner suggestions fit a person’s budget and cooking style without repeated corrections when the model’s self-ratings remain accurate.

Today’s personal assistants have two poor ways to remember you. They can carry a long summary of past conversations, or retrieve old snippets and paste them into each new request. That memory takes up space beside your current question, may bring back something irrelevant, and still leaves you correcting the assistant when it treats your tastes as ordinary. If there is little explicit feedback, the provider has little usable evidence for improving the system.

A smaller kind of memory

COPE starts with a different idea: give each user a compact numerical state rather than a readable profile. This state is called a personalized embedding. It is not a list saying “likes cheap meals” or “prefers slow travel.” Instead, it is a set of numbers that guides the language model toward answers that fit one person.

That compactness matters because a shared model serves many people. Training it directly on one person’s feedback might improve that person’s answers while changing how it behaves for everyone else. A separate user state aims to learn individual preferences without filling every prompt with a personal history.

But the state is also hard to inspect. A user or operator cannot easily read it, correct one mistaken belief, or know exactly what it has inferred. And if the model judges its own answers incorrectly, it may learn from its own mistake.

Learning from the missing ratings

COPE, described in the paper, represents each user’s state as virtual tokens placed before the user’s text. They are not words shown to the user. They are extra pieces of numerical guidance that the model reads along with the request.

The method updates this state and the shared model in a repeating “Interact-Collect-Optimize” loop. The assistant answers a request. When the user gives a comment, that text changes the user embedding. When the user gives a number, that number helps calibrate the model’s self-evaluation. For later requests without feedback, the model can use its own predicted score as a stand-in for a missing rating.

COPE scores an answer in two parts: whether it is complete and whether it fits the user. The reward is the product of those two scores. In the paper’s main experiment, 256 simulated users received answers to single-turn tasks. At 50% feedback, COPE scored 4.69 out of 10, compared with 4.47 for a training-based profile prompt and 3.65 for a profile added to the prompt. Combining COPE with sparse retrieval reached 5.23.

The most important number is less flattering. The model’s self-evaluator agreed with user-feedback rankings on 72.63% of pairs when it made a directional judgment. That leaves many chances for a wrong guess to become training data. A system that decides an answer pleased someone, then uses that decision to shape its next answer, could reinforce a preference the person never had.

Refer to caption
Figure 1: Illustration of the dynamic interaction scenario and the iterative “Interact-Collect-Optimize” loop. (a) Interaction & Sparse Feedback: During session tt, for each user, the assistant πθt\\pi\_{\\theta\_{t}} generates a response to the query, guided by the user-specific personalized embedding. If it fails to align with the user’s hidden preferences ηi\\eta\_{i}, explicit feedback (textual fitf\_{i}^{t} and numerical sits\_{i}^{t}) is provided sparsely with probability pp. (b) Data Collection & Optimization: Interaction data 𝒟1t\\mathcal{D}\_{1}^{t} (without feedback) and 𝒟2t\\mathcal{D}\_{2}^{t} (with feedback) are collected to iteratively optimize the model. (c) Subsequent Interaction & Improvement: When encountering a query involving the similar preference in subsequent sessions t+kt+k, the updated assistant πθt+k\\pi\_{\\theta\_{t+k}} successfully incorporates learned preferences to provide a personalized response.## 2 Setup Source: COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation（arXiv）
Refer to caption
Figure 6: Schematic illustration of personalized embedding integration and the forward computation and backward optimization of the three objectives in the optimization stage of COPE. Personalized embeddings are implemented as ll virtual tokens and prepended to the text embeddings to provide user-specific guidance. For clarity, the assistant model is conceptually decomposed into an embedding layer and a decoder to explicitly visualize the embedding of textual inputs and their concatenation with personalized embeddings, although both components belong to the same assistant model. Two different system prompts are used: a response-generation prompt (pink) and a self-evaluation prompt (green), both of which are provided in our code repository. Trainable parameters under each objective are marked with spark icons, while frozen parameters are indicated by snowflake icons.## Appendix E Illustration of Personalized Embedding Integration and Optimization Flow Source: COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation（arXiv）

If the memory becomes trustworthy

In a kitchen, someone could ask for dinner ideas without repeating the same corrections about budget, ingredients, and cooking style every week. The assistant could learn those preferences from scattered comments instead of carrying a full conversation history into each request. In a travel-booking chat, recommendations could fit a person’s preferred pace and accessibility needs without displaying a digest of old conversations each time.

That future depends on the learned state remaining useful when tastes change. It also depends on the provider making persistent preferences safe, controllable, and accurate. COPE’s test used an LLM simulator, pre-generated and manually inspected queries, 51 interactions per user, and single-turn sessions. Real people, longer relationships, and changing preferences still have to show whether the model’s private guess is a safe substitute for the user’s judgment.

Sources

COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation