PERSOMA: PERsonalized SOft ProMpt Adapter Architecture for Personalized Language Prompting

Our method is grounded in the unified text-to-text approach proposed by T5 (Raffel et al., 2020), where tasks are conceptualized as text generation. Formally, for an input sequence of tokens $X$, the output sequence $Y$ is modelled probabilistically as $Pr_{\theta_{LM}}(Y|X)$ parameterized by the language model’s learned weights $\theta_{LM}$. Prior research in text-to-text tasks can be divided into two predominant prompting methodologies: text-based prompting, which integrates textual directives into the input (Mishra et al., 2022; Chung et al., 2022; Wei et al., 2022), and soft-prompting, introducing a series of fixed trained tokens preceding the model’s input tokens to capture information about a specific task (Lester et al., 2021; Li and Liang, 2021).

Conventionally, soft-prompting employs a static, task-specific soft-prompt to enhance parameter efficiency across various linguistic tasks, optimizing the likelihood $Pr_{\theta_{LM}}(Y|[T;X])$, where $T$ represents a fixed set of trainable task tokens (called task ”soft prompts”) and $X$ is the task input. Given a set of historical interactions $H_{u}$ described in natural language for user $u$, we further train a soft prompt adapter network to jointly compress and resample these interactions into a set of personalized user=specific soft-prompt tokens $P_{u}$ where $|P_{u}|\leq|H_{u}|$. The generated soft-prompt tokens can then be utilized to optimize $Pr_{\theta_{LM}}(Y_{u}|[P_{u};T;X])$, where $Y_{u}$ is a personalized text output conditioned on user $u$’s history.

In practice, PERSOMA (depicted in Figure 1) consists of three primary components: the history encoder, soft prompt adapter and resampler and a large language model decoder. We first encode each natural language user interaction $h_{i}\in H_{u}$ using a SentenceT5 text embedding model (Ni et al., 2022), our history encoder, creating $H^{\prime}_{u}$. We then feed the set of $H^{\prime}_{u}$ into our soft prompt adapter network to generate the set of personalized soft prompt tokens $P_{u}$. In our results, we experiment with architectures which jointly encode the sequence of $H^{\prime}_{u}$ embeddings (Perceiver (Jaegle et al., 2021), Transformer (Vaswani et al., 2017)) as well as those that encode embeddings individually (MLP projection). Finally, the combined soft prompt $[T;P_{u}]$ is fed into a frozen PaLM 2 Language Model to generate personalized responses to the given task parameterized by $T$. To ensure our work is comparable to the prior art, all mentions of PalM 2 refer to the smallest XXS version.

To train and evaluate PERSOMA, we utilize the personalized genre prediction task proposed by (Doddapaneni et al., 2024) using the MovieLens dataset (Harper and Konstan, 2016). We briefly describe it here for completeness. The MovieLens dataset contains the viewing habits of users and metadata about the movies they are watching. We formulate the genre prediction task on this dataset as follows: given a set of natural language titles $m^{t}$, descriptions $m^{d}$ and review scores $m^{r}$ for each movie $m\in H_{u}$ user $u$ watched, predict the set of genres the user likes $G_{u}^{+}$ and dislikes $G_{u}^{-}$. The task is suitable for evaluating personalization as the model must understand the user’s preferences to predict personalized genre suggestions. We removed all films with less than 20 reviews to ensure that each movie was high quality and could be seen sufficiently across user sequences in the dataset. After pruning, the final dataset includes 14.4 M reviews from 127 K users spanning 8.2 K unique movies and 19 genres.

To apply PERSOMA and other embedding-based personalization baselines on this task, we construct our set of item embeddings $H^{\prime}_{u}$ by concatenating the history encoder embeddings for $m^{t}$, $m^{d}$ and $m^{r}$. For text-only baselines, we instead represent the input history by concatenating the direct text description of each movie, formatted as

Finally, to frame this task as a text-to-text generation task, we formulate the target genre prediction output $Y$ as

where a comma delimits each genre in $G_{u}^{+}$ and $G_{u}^{-}$ (ex: ”Liked Genres: Action, Comedy. Not Liked Genres: Romance”).

Rather than use conventional text generation metrics such as BLEU and ROUGE, the predicted genre names are extracted from the generated string to evaluate class-weighted Recall, Precision, and F1. This ensures that we explicitly consider the model’s performance on the target task, irrespective of the order in which the genres are predicted. Given that users can like or dislike each of the 19 genres, our evaluation metrics evaluate the problem as a 38-class multi-classification problem¹¹1The resulting multi-label distribution can be seen in the Appendix Table 5. We divide the MovieLens dataset into sets of 114K/5K/5K users for training, validation and test respectively, following (Doddapaneni et al., 2024).

A vital component of the PERSOMA architecture is how we process, resample, and represent the user’s historical interactions for personalization. We evaluate PERSOMA’s performance towards this capability by comparing it against various personalization methods when faced with a user history of the latest 5, 10, and 50 movies watched, respectively. We compare against end-to-end finetuned models UEM Base and UEM Large from (Doddapaneni et al., 2024), which use T5, and a PaLM 2 (Anil et al., 2023) XXS model following the text input formulation described in section 3.2 as embedding-based and text prompt-based baselines respectively. We also include a zero-shot Gemini 1.5 Pro model baseline to compare against PERSOMA methods without end-to-end finetuning.

Our results, shown in Table 1, comparing MLP, Perceiver, and Transformer variants of PERSOMA, demonstrate that both MLP and Transformer variants outperform UEM Base and UEM Large, particularly with larger history sizes. For example, PERSOMA MLP achieves a 0.19 higher F1 score than UEM Large when processing 50 items. Even with a frozen language model, PERSOMA MLP with Frozen LM still outperforms UEM Large by 0.16 F1 with the same history size.

Since PERSOMA MLP with Frozen LM trains far fewer parameters than UEM Large, our results indicate that finetuning the entire model may be unnecessary with a stronger base LLM (PaLM2 vs T5). Instead, training the adapter to create expressive user embeddings within the LLM language space is sufficient to create significant personalization gains, even at smaller history sizes. This result is further reinforced when resampling with a Perceiver model, unlocking further improvements in token efficiency.

Next, we inspect the performance of PERSOMA against text-prompting baselines. Here, it is essential to note that text prompting requires significantly more computing, scaling according to text token length rather than having a single token per history item. This is best seen in Table 2, where we note that text prompting would require $\approx$ 16 000 input tokens to model a user history of 50 items in MovieLens, whereas PERSOMA would only need 130, further reduced to 100 with Perceiver resampling. Further, it is worth noting that PaLM 2 is finetuned end-to-end toward the target task, tuning many more parameters than our Frozen variants.

Despite utilizing several orders of magnitude less computing, the gap between PERSOMA and PaLM 2 text prompting is only a marginal 0.007 F1 and 0.019 F1 difference when using a history size of 5 and 10, respectively. Comparing PERSOMA Frozen and Gemini, we see that PERSOMA MLP achieves 0.308 higher F1 at a history length of 50. This demonstrates the effectiveness of using in-domain soft prompts for personalization over text prompts.

Finally, we compare the performance of various PERSOMA soft prompt adapter architectures. Surprisingly, we find that PERSOMA MLP, a smaller and simpler architecture, outperforms PERSOMA Perceiver and PERSOMA Transformer. This is especially the case when utilizing a large history size of 50, where PERSOMA MLP outperforms PERSOMA Transformer and PERSOMA Perceiver by 0.024 F1 and 0.198 F1, respectively.

There are a few possible reasons for this loss in optimal performance with sequential adapter models. First, it is essential to note that while MovieLens does provide an ordering of watched films, the order is self-reported by users on the platform via a survey (Harper and Konstan, 2016). Given that ordering is not strict, architectures incorporating positional encodings, such as Transformer and Perceiver, may be better suited for tasks that heavily depend on interactions’ temporal order. Second, the lack of performance of the Perceiver may be attributed to the reduced number of prompt tokens, which may not have sufficient expressivity as input to the LLM. However, we note that the Perceiver result resampled to a history length of 20 is better than MLP with a history length of 10.

To assess the effectiveness of parameter-efficient finetuning techniques for PERSOMA, we conducted experiments using Low-Rank Adaptation (LoRA) (Hu et al., 2022) and frozen language model weights, comparing them to a fully finetuned PERSOMA model. For LoRA experiments, we use rank four adaptation for both attention and transformer feedforward layers.

As demonstrated in Table 4, when using recency sampling with a history size of 50, both LoRA and frozen PERSOMA variants achieved strong F1 scores of 0.533 and 0.541 respectively, closely matching the performance of the end-to-end finetuned model (0.569 F1). This indicates that parameter-efficient techniques can be effectively applied to PERSOMA, offering a promising avenue for reducing computational costs without sacrificing performance.

Finally, inspired by the LAMP personalization benchmark (Salemi et al., 2023), we extensively evaluate the impact of various methods sampling strategies on the performance of PERSOMA, PaLM 2 and Counting. For a given history size $H$, we evaluate the following sampling strategies:

• •

  Recency: Select the latest $H$ watched movies

• •

  Random: Uniformly sample $H$ watched movies

• •

  Long Tail: Uniformly sample $H$ watched movies that are below the global 90th quartile of popularity

• •

  Top-K Popularity: Select the top $H$ globally popular movies the user watched

• •

  Genre Sample: Sample $H$ movies the user watched according to the users genre density

The results of each sampling strategy can be seen in Table 3. We can see that PERSOMA performance improves significantly as the history size increases regardless of sampling strategy. PERSOMA achieved the highest performance with Top-K popularity, achieving 0.593 F1 and 0.294 F1 at history 50 and 5, respectively. Top-K sampling bested Recency sampling, which achieved 0.569 F1 and 0.278 F1 with a history of 50 and 5.

One notable outlier sampling strategy was long tail sampling, achieving a lower F1 across PERSOMA and PaLM 2. This performance gap is likely due to many esoteric films not being in the popular zeitgeist and, therefore, not being well represented in the training data for the language model (unlike Top-K sampling, which features only popular movies). Besides long-tail sampling, all other strategies perform within $\pm$ 0.03 F1 of each other, demonstrating PERSOMA’s robustness.