PaperlayerView exact revision on arXiv

Source-linked AI summary

ProTTT: Learning to Learn Semantic User Memory with Test-Time Training

Sejun Park, Hyoungjo Bhang, Hyein Jeong, Yohan Jo

arXiv:2609.32564v1cs.AI

TL;DR

Personalization requires models to capture implicit user knowledge from growing histories, while existing context-based and parametric approaches have drawbacks in cost or updating. PROTTT learns a shared memory initialization with profile supervision and adapts it through test-time training on user history. It outperforms full-history ICL and parametric baselines across benchmarks, with analyses showing it tracks and retains evolving preferences.

  • Problem

    Context-based personalization incurs increasing inference costs or depends on retrieval and summaries, while parameter-based approaches often reprocess histories as user data grows.

    Source: §1 · pp. 1–2

  • Method

    PROTTT learns a shared initialization using textual user profiles, then adapts lightweight user memory through next-token training on user history.

    Source: §1 · p. 2

  • Results

    PROTTT consistently outperforms full-history ICL and all parametric baselines across diverse personalization benchmarks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The analyses indicate that test-time training can construct and incrementally update semantic user memory that tracks and retains evolving preferences.

    Source: §1 · pp. 2–3

    Deployment risks include unauthorized profiling, privacy violations, inaccurate representations, and manipulative personalization; practical use should require consent, memory protection, and user inspection, correction, and deletion mechanisms.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Personalization requires language models to capture user-specific knowledge from a growing user history. Existing context-based approaches incur increasing inference costs as user history accumulates and rely on separate retrieval or summarization stages, while parametric-based approaches often require reconstructing user representations when new user data is added. We introduce ProTTT, a profile-supervised meta-learning framework for learning semantic user memory. The memory construction starts from a shared initialization and is updated for each user through test-time training on user history, allowing it to evolve continuously as the history grows. However, since test-time training alone does not explicitly encourage the memory to capture semantic user knowledge necessary for personalization, we learn this shared initialization using textual user profiles as supervision, so that test-time training on user history captures semantic knowledge more effectively. ProTTT consistently outperforms both full history ICL and all parametric baselines across diverse benchmarks, while substantially reducing inference cost by compressing user history into a lightweight parameterized memory. Our analysis also shows that profile supervision is a reliable objective for learning semantic user knowledge and that the resulting memory can track and retain evolving user preferences, while remaining robust across different history sizes. Overall, we demonstrate the effectiveness of test-time training for personalization and establish ProTTT as a baseline for continuously evolving user memory.

1 INTRODUCTION

Personalization requires models to infer semantic user knowledge, such as preferences and behavioral tendencies, across user histories. PROTTT learns a shared memory initialization through profile supervision, then adapts it on user history; it outperforms full-history ICL and parametric baselines across benchmarks while reducing inference cost.

1 INTRODUCTION

  • Full-history prompting scales inference cost with history length and is context-window bounded, while retrieval and summaries add record-selection or revision dependencies.Retrieval also adds inference-time overhead; summary effectiveness depends on summary quality.Source: §1 · p. 1
  • Parameter-based approaches encode user history compactly, but existing methods typically reprocess the full history when new data arrives.This motivates memory that can be updated continuously from its current representation.Source: §1 · pp. 1–2
  • PROTTT adapts a shared memory initialization through next-token prediction on user history, using textual profiles during training to encourage semantic user knowledge.At inference, it adapts on the target user’s history alone and uses the resulting memory for personalized generation.Source: §1 · p. 2, Figure 1 · p. 2
  • PROTTT consistently outperforms full-history ICL and all parametric baselines across personalization benchmarks, with 10× lower latency than full-history ICL on CMV.Evaluations cover persuasiveness prediction, product-rating prediction, and recommendation and generalization to new scenarios.Source: §1 · p. 2
  • Profile prediction is the most consistent meta-learning objective across benchmarks, and the memory tracks recent preferences while retaining earlier knowledge more robustly than full-history ICL as history grows.Personalization remains relatively stable across history sizes; reducing the memory scaling factor helps preserve general capabilities while largely preserving personalization knowledge.Source: §1 · p. 2
  • The results support test-time training as a lightweight mechanism for constructing and incrementally updating semantic user memory.The paper presents this as a foundation for continuously evolving user memory in personalized language models.Source: §1 · p. 3

2 RELATED WORK

Prior personalization work largely uses context-based methods that supply user data at inference or parameter-based methods that encode it in learned representations. PROTTT builds on test-time parameter updates as memory, shaping them with profile-supervised meta-learning for semantic information that transfers across personalization tasks.

2 RELATED WORK

  • Context-based personalization supplies user data in prompts through retrieval or textual profiles, including task-aware summaries and joint retrieval-summarization approaches.LaMP studies lexical, semantic, and time-aware retrieval; later work uses downstream-model feedback to improve retrieval.Source: §2 · p. 3
  • Parameter-based methods encode user data in learned representations, including per-user modules, shared parameter fragments, hypernetwork-generated LoRA, and user embeddings.Examples include OPPU, Personalized Pieces, P2P, and approaches that keep the base model frozen.Source: §2 · p. 3
  • Prior test-time-training work uses online updates as memory, while PROTTT applies this framework to personalization and introduces profile-supervised meta-learning to target semantic user information.Related approaches include TTT layers, Titans, TTT-E2E, PERK, GradMem, and δ-mem.Source: §2 · p. 3

3 METHOD

PROTTT represents user memory with lightweight adapters and learns a shared initialization through profile-supervised meta-learning. At inference, it adapts that initialization sequentially on user history and can continue updating from new chunks without reprocessing prior history.

3.1 PARAMETRIC USER MEMORY

  • Each memory module normalizes a layer’s hidden representation, applies a low-rank transformation, and adds its output to the residual stream before the next transformer layer.The adapter parameters across layers constitute the user memory, separate from the transformer layer parameters.Source: §3.1 · pp. 3–4, Figure 2 · p. 3

3.2 TRAINING

  • For each user, training starts from the shared initialization and updates memory sequentially on chronologically ordered history chunks using next-token prediction.After all chunks, the resulting user-specific memory generates a profile; profile-token loss is backpropagated through the adaptation sequence to update the shared initialization.Source: §3.2 · p. 4
  • Profile-supervised meta-learning addresses the fact that next-token prediction alone does not explicitly encourage memory to capture semantic user knowledge.Textual user profiles supervise the shared initialization so that test-time training constructs semantic user memory more effectively.Source: §3.2 · p. 4

3.3 INFERENCE

  • At inference, the target user’s history alone is used to adapt memory from the learned initialization, after which the memory is fixed for personalized responses.New history chunks can update the current memory without repeatedly processing previously observed history.Source: §3.3 · p. 5

4 EXPERIMENTS

PROTTT is evaluated on persuasion, rating prediction, recommendation, and generalization, against in-context and parameter-based baselines. It achieves the strongest reported personalization results while substantially reducing prompt length and latency relative to full history.

4.1 EXPERIMENTAL SETUP

  • The evaluation covers CMV persuasion, LaMP-3 rating prediction, and PersonaMem Types 6 and 7 for recommendation and generalization to new scenarios.Baselines include full-history and retrieval-based in-context methods, user-specific LoRA, P2P, and δ-Mem.Source: §4.1 · p. 5

4.2 PERSONALIZATION PERFORMANCE AND EFFICIENCY

  • In-context methods do not consistently benefit from explicit textual profiles: profile-augmented generation with full history improves CMV but degrades LaMP-3 and PersonaMem Type 7.Across in-context baselines, the strongest method varies by benchmark; retrieval performs best on PersonaMem Type 6.Source: §4.2 · p. 5
  • PROTTT outperforms full history and all parametric baselines across datasets, with the best results on CMV, LaMP-3, and PersonaMem Type 7.Without profile-supervised meta-learning, naive TTT produces inconsistent gains across tasks.Source: §4.2 · p. 6, Table 1 · p. 6
  • 25.3× shorter prompts and 10.4× lower per-query latency than full history on CMV accompany a one-time memory-construction cost of 2.82 seconds per user.Memory construction can be reused and amortized across subsequent queries.Source: §4.2 · p. 6
  • PROTTT also outperforms full-history ICL across reported evaluations when applied to Llama-3.2-1B-Instruct with a separately meta-learned memory initialization.The paper reports representative efficiency measurements on CMV and provides full benchmark results in the appendix.Source: §4.1 · p. 5, §4.2 · pp. 6–7

5 ANALYSIS

The analysis finds that profile prediction is the most consistent meta-learning objective, and that PROTTT can acquire and retain information as histories grow. It also examines the trade-off between personalization and general model behavior under different history sizes and memory readout scales.

5.1 WHICH META-LEARNING OBJECTIVE BEST GUIDES USER MEMORY FORMATION?

  • Profile prediction improves performance across all three evaluation settings, while persuasion judgment, response selection, and response generation each improve some tasks but degrade others.The figure reports signed performance changes and raw scores before and after memory construction; positive changes indicate improvement.Source: §5.1 · p. 7, Figure 3 · p. 7

5.2 CONTINUAL ACQUISITION AND RETENTION OF USER KNOWLEDGE

  • As seven later histories are incorporated, PROTTT retains Profile A’s likelihood more robustly than full-history ICL, whose likelihood progressively declines.The analysis also reports that each profile’s mean token log-probability rises substantially after its associated history is incorporated.Source: Figure 4 · p. 7, §5.2 · p. 8

5.3 SCALING WITH USER-HISTORY SIZE

  • TTT outperforms ICL on CMV and LaMP-3 across comparable history sizes, and its personalization performance shows no degradation as history grows to 300 items.ICL is reported only through k = 100 because of the context limit; MMLU performance also shows no degradation as more history is incorporated.Source: §5.3 · p. 8, Figure 5 · p. 8

5.4 GENERAL CAPABILITY, GENERATION QUALITY, AND USER-INFORMATION RECALL

  • Scaled PROTTT readout at α = 0.75 yields 5.69 history-supported claims per response, versus 1.67 for the base model, 1.24 for user-specific LoRA, and 2.80 for δ-Mem.The scaled readout also reduces general-capability and generation-quality degradation; full-strength readout produces a larger MMLU reduction and more unsupported claims than the scaled model.Source: §5.4 · p. 9, Table 2 · p. 9

6 CONCLUSION

PROTTT combines profile-supervised initialization with test-time training to build lightweight user memory that updates as history grows. The paper reports benchmark gains and lower inference cost, while identifying privacy and profile-quality concerns for deployment.

6 CONCLUSION

  • PROTTT learns a shared initialization from textual user profiles, then uses test-time training to build and continuously update lightweight user-specific memory.Source: §6 · p. 9
  • Across multiple personalization benchmarks, PROTTT outperforms full-history ICL and parametric baselines while reducing inference cost through compact parameterized memory.Source: §6 · p. 9

AI USE STATEMENT

  • The authors report using generative AI for code, literature work, drafting, and specified experimental tasks, and say they reviewed and verified AI-assisted artifacts.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • User histories may contain sensitive information, and internalized memories can risk unauthorized profiling, privacy violations, inaccurate representations, or manipulative personalization.The authors say deployment should require consent, memory protection, and ways for users to inspect, correct, and delete information; generated profiles and judgments may contain errors or biases.Source: ETHICS STATEMENT · p. 10
  • The work uses existing research datasets and reports no new user-data collection or human-subject experiments.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

A.1 HISTORY SERIALIZATION AND CHUNKED TTT UPDATES

User histories are serialized in order and processed through sequential, chunk-level test-time memory updates rather than one full-history pass. ALPSBench uses a distinct loss scope because its histories are short transcripts with substantial assistant content.

A.1 HISTORY SERIALIZATION AND CHUNKED TTT UPDATES

  • Histories are serialized as separate user turns in original order, then split into chunks of up to 2,048 tokens for one TTT update per chunk.The memory is neither updated per token nor written from the full history in one pass.Source: §A.1 · p. 12
  • Records are packed greedily into chunks, splitting a record only when it exceeds the chunk limit; ChatML turn boundaries need no extra separators.Source: §A.1 · p. 12
  • Each chunk's next-token loss updates memory before the next chunk is processed, and the loss covers original user-record tokens rather than ChatML role labels or boundaries.Serialization, masking, and updates are identical during meta-learning and inference.Source: §A.1 · p. 13
  • For ALPSBench, loss covers both user and assistant content because histories are short transcripts and only about 32% of tokens come from user turns.Assistant responses can reflect preferences, experiences, and context expressed by users in preceding turns.Source: §A.1 · p. 13

A.2 TRAINING DETAILS

Training pairs user histories processed through TTT with semantic-memory targets that supervise a shared initialization. The appendix describes data sources, target construction, and differentiable training of that initialization while keeping the backbone frozen.

A.2 TRAINING DETAILS

  • Each meta-learning example pairs a TTT-processed user history with a semantic-memory target, using ALPSBench session dialogues and CMV user posts and comments.Table 3 summarizes the two data sources.Source: §A.2 · p. 13, Table 3 · p. 14
  • For ALPSBench, Task 1 pairs dialogues with verified memory items, while Task 2 pairs original and extended dialogues with their corresponding original and updated memories.Task 4 supplies profile-extraction and request-conditioned targets.Source: §A.2 · p. 13
  • For CMV, training uses ReCAP users with at least eight history entries and at most their 100 most recent posts and comments; GPT-5-mini generates profiles from histories and current posts.The profiles describe characteristics relevant to persuasion in the discussion and serve as meta-learning targets.Source: §A.2 · p. 13
  • ALPSBench prompts ask the model to recover user information, with verified memory items rendered as one bullet per line; Task 4 also asks for the fact relevant to a new request.Source: §A.2 · p. 13
  • For CMV, the profile-generation prompt additionally supplies the current post.Source: §A.2 · p. 13
  • Training starts each example from the shared memory initialization, applies TTT to the history, then generates the target with Qwen3-0.6B as a frozen backbone.Source: §A.2 · p. 14
  • Teacher-forced target-token loss is differentiated through the complete TTT-update sequence to update the shared initialization, while the language-model backbone remains frozen.Multiple targets are processed separately, with losses combined according to supervised target-token counts.Source: §A.2 · p. 14
  • Each update uses two examples from distinct task-level sources, samples sources by the square root of example counts, and trains for three epochs totaling 6,531 updates.The main-model hyperparameters are reported in Table 4.Source: §A.2 · p. 14

A.3 ADDITIONAL PERSONAMEM RESULTS

The additional PersonaMem analysis focuses on recommendation and responses to new situations, where questions draw on earlier user interactions. Results are less consistent on categories requiring precise historical recall.

A.3 ADDITIONAL PERSONAMEM RESULTS

  • The evaluation covers all seven PersonaMem categories, with Types 6 and 7 as the main focus because they require using earlier interactions for recommendations or new situations.The benchmark contains 589 questions with four answer choices each.Source: §A.3 · p. 14

A.3 ADDITIONAL PERSONAM RESULTS

  • Gains are less consistent on categories requiring precise recall, including specific facts, events, preference order, or whether an item appeared in the history.Source: §A.3 · p. 14
  • The authors characterize PROTTT as emphasizing reusable user knowledge over preserving a detailed record, while noting that some fine-grained details are retained.Source: §A.3 · p. 14

A.4 BASELINE DETAILS

The section specifies shared evaluation conditions and the construction of context-based and parametric baselines. It also identifies the contents of the main hyperparameter and PersonaMem accuracy tables.

A.4 BASELINE DETAILS

  • All methods use Qwen3-0.6B, official benchmark prompts and scoring, and at most each user’s 100 most recent history entries.Full History includes a comparable amount of history without exceeding the context window; context-based methods prompt with history, while parametric methods encode it first.Source: §A.4 · p. 14
  • Table 4 lists main-model training hyperparameters, and Table 5 reports PersonaMem accuracy across released question categories, with question counts shown in parentheses.Table 5 defines before/after as without/with target-user information and notes that the two after-TTT PROTTT rows differ only in memory readout scale.Source: Table 4 · p. 15, Table 5 · p. 15
  • Full History prompts with the full history, Recent-5 with five recent entries, and BM25 with five retrieved entries; Base receives only the task input.BM25 uses the current task input as its retrieval query.Source: §A.4 · p. 15
  • User-Specific LoRA trains a separate rank-16 adapter per user on history, then evaluates with only the task query and a frozen backbone.The adapter is applied to attention and feedforward layers with scaling factor 32; the one-epoch setting is selected on validation data.Source: §A.4 · p. 16
  • PAG generates one query-independent textual profile per user with GPT-5-mini and prepends it to each task input.The prompt requests concise bullet points covering supported interests, recurring topics, opinions and values, reasoning style, and writing style; PAG + Full History adds the retained history.Source: §A.4 · p. 16
  • δ-Mem uses the authors’ official recipe, with its backbone replaced by Qwen3-0.6B while retaining its memory architecture, objective, and training procedure.Source: §A.4 · p. 16
  • At evaluation, methods write history into memory before processing the task input without history in the prompt; the w/o user condition uses empty memory.Source: §A.4 · p. 16

A.5 EFFICIENCY MEASUREMENT DETAILS

This appendix details ProTTT’s efficiency evaluation and supplementary experiments, including how memory readout scale affects behavior across tasks and backbones.

A.6 META-LEARNING SUPERVISION OBJECTIVES

  • The four CMV-trained meta-learning objectives are profile prediction, persuasion judgment, response selection, and response generation, which respectively describe user-relevant memory, classify replies, identify a persuasive reply, or generate one.The tasks differ in their post-memory-construction task while using the same CMV users.Source: §A.6 · pp. 17–18

A.5 EFFICIENCY MEASUREMENT DETAILS

  • PROTTT uses 748/170/394-token prompts and 26.2/15.1/14.2-ms query latency across CMV/LaMP-3/PersonaMem, versus Full History’s 18.9K/12.5K/10.5K tokens and 273.5/90.3/100.7 ms.Memory writing takes 2.82/1.18/0.89 s per user or group—less than user-specific LoRA and P2P, but slower than δ-Mem; P2P incurs additional adapter-generation costs on every query.Source: §A.5 · p. 17, Table 7 · p. 18, Table 8 · p. 18

A.7 CONTINUAL-MEMORY STREAM CONSTRUCTION AND EVALUATION

  • Continual-memory tests sequentially write new history segments without replay, score gold-profile statements after each write, and assess acquisition and retention using synthetic streams that exclude contradictory preference changes.The streams contain compatible, complementary user information; PROTTT’s Profile A score is tracked as later histories are written.Source: §A.7 · p. 20

A.8 GENERATION AND USER-INFORMATION RECALL DETAILS

  • Generation quality is evaluated on filtered WildChat prompts with memory reused per user, using non-termination, repetition-2, Distinct-2, and Self-BLEU; user descriptions are split into claims labeled supported, generic, unsupported, or contradicted.The evaluation prompts omit user history and textual profiles, testing whether user information can be expressed from memory.Source: §A.8 · pp. 20–21

A.9 PROTTT WITH LLAMA-3.2-1B-INSTRUCT

  • With Llama-3.2-1B-Instruct, history-built PROTTT improves over shared initialization and matches or beats every context-based baseline across all five personalization metrics.The result extends the effectiveness of profile-supervised memory construction beyond the Qwen backbone used in the main experiments.Source: §A.9 · p. 21, Table 11 · p. 22
  • At α = 0.5, Llama MMLU rises from .259 to .405 and non-termination falls from 5.9% to 1.4%; supported claims number 6.85, versus 5.15 at α = 1 and 3.93 for the base model.Across seven later history writes, Profile A’s mean token log-probability falls approximately .48 under PROTTT versus .68 under Full History; all four profiles remain above their initial values after the full sequence. Both memory settings also produce more unsupported claims than the base model.Source: §A.9 · p. 22, Table 12 · p. 23, Figure 6 · p. 23

A.10 EFFECT OF THE MEMORY READOUT SCALE

  • Readout optima vary by backbone and task: Qwen peaks at α = 1 on CMV and LaMP-3 but α = 0.75 on PersonaMem, while Llama peaks at α = 0.25, 0.5, and 1, respectively.For both backbones, stronger readout can improve personalization while degrading general behavior; at full readout, generation repetition and non-termination increase, and Llama Distinct-2 decreases.Source: §A.10 · pp. 23–24, Figure 7 · p. 24
  • The best memory readout scale depends on the task and backbone: smaller α tends to preserve frozen-backbone behavior, while larger α can strengthen personalization when the task benefits from a more influential memory.No single scale performs best across all evaluations.Source: §A.10 · p. 24
Loading 2609.32564v1…