PaperlayerView exact revision on arXiv

Source-linked AI summary

DPPM: Dual-Path Parametric Memory for Personalized Language Models

Yuhao Chen, Shuochen Liu, Jiayao Shi, Jian Hong, Chen Cheng, Xinyun Ding, Tao Wang, Ya Li, Quan Liu, Tong Xu

arXiv:2610.11776v1cs.CL

TL;DR

Long-term personalization needs cross-session memory, but repeated updates may attenuate earlier evidence and evidence pooling cannot represent order-dependent changes. DPPM combines direct evidence aggregation with chronological associative updates in a unified parametric memory. With Qwen3-4B, it outperforms evaluated baselines on PersonaMem-v2 and PrefEval.

  • Problem

    Long-term personalization needs cross-session memory, but textual context methods depend on accurate extraction and retrieval and incur inference-context costs; repeated updates may attenuate earlier evidence, and evidence pooling cannot represent order-dependent changes.

    Source: §1 · p. 1, §3.2 · p. 3

  • Method

    DPPM combines direct aggregation of interaction-history representations with chronological associative-state updates in a unified parametric memory.

    Source: §1 · p. 2

  • Results

    With Qwen3-4B, DPPM outperforms evaluated baselines, reaching 54.22% accuracy on PersonaMem-v2 and 86.79% accuracy on PrefEval.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors present preserving historical evidence while adapting its use as a practical design for long-term personalized parametric memory.

    Source: §6 · p. 8

    Experiments are limited to 2B–7B backbones and textual interaction histories; larger-scale validation and extensions to images and speech remain future work.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Long-term personalization requires language models to use interaction history to track users' preferences across sessions. Parametric memory encodes this interaction history into model parameters or adapters, reducing the need to include it in the inference context. However, independent context compilation leaves cross-session integration unspecified, while recurrent updates can attenuate earlier evidence. To address these challenges, we propose Dual-Path Parametric Memory (DPPM). Its Evidence path directly pools representations of the interaction history to preserve earlier evidence, while its Delta path sequentially updates an associative state to capture changes. Fusing both outputs produces history-conditioned LoRA adapters that combine evidence accumulation with ordered revision. Across multiple backbones, DPPM outperforms the evaluated baselines, achieving 54.22% on PersonaMem-v2 and 86.79% on PrefEval. These results suggest that DPPM provides a simple and effective design choice for cross-session personalized parametric memory.

1 Introduction

Long-term personalization needs memory that carries preferences and constraints across sessions without placing the full interaction history in each response context. DPPM combines direct evidence aggregation with sequential updates to address limitations of context compilation and recurrent memory.

1 Introduction

  • Textual-history systems rely on accurate extraction, retrieval, and contextual reasoning, while including more interaction history increases inference-time context costs.Such systems commonly construct response contexts through summarization, retrieval, or explicit memory editing.Source: §1 · p. 1
  • Existing parametric-memory methods either compile individual contexts into LoRA adapters or integrate information through sequential state updates.Doc-to-LoRA and SHINE are context-compilation examples; RPMem and Metis use state maintenance.Source: §1 · p. 1
  • Independent compilation does not itself support selective retention and revision, while recurrent consolidation can attenuate earlier contributions.These gaps motivate retaining direct access to earlier evidence while supporting revision as interaction history grows.Source: §1 · p. 1
  • DPPM separates evidence accumulation from sequential updates: its Evidence path aggregates history representations, and its Delta path updates an associative state chronologically.Learnable fusion combines both outputs into memory; a frozen decoder converts it into history-conditioned LoRA adapters.Source: §1 · p. 2
  • DPPM outperforms evaluated baselines, reaching 54.22% accuracy on PersonaMem-v2 and 86.79% accuracy on PrefEval with Qwen3-4B.The evaluation spans three backbones, architectures, and model sizes from 2B–7B parameters.Source: §1 · p. 2

2 Preliminary

The workflow divides interaction history into chronological segments, aggregates their compiled memories, and uses the resulting memory to condition responses without including the interaction history in the prompt. A context encoder and adapter decoder form the hypernetwork that generates LoRA factors for a segment.

2.1 Parametric Memory Workflow

  • The workflow divides interaction history into T chronological segments and aggregates their compiled memories in order to form consolidated memory.A decoder produces memory parameters, which are applied to the frozen backbone; inference uses the compiled memory without the interaction history in the prompt.Source: §2.1 · p. 2

2.2 Hypernetwork-Generated LoRA

  • The hypernetwork combines a context encoder that maps text to latent memory with an adapter decoder that generates LoRA factors.For target matrix Wℓ, the factors Aℓ and Bℓ have dimensions r_L×d_in and d_out×r_L, respectively.Source: §2.2 · p. 2
  • Once trained, the shared hypernetwork generates adapters in a forward pass without per-segment optimization.Source: §2.2 · p. 2

3 Method

DPPM encodes interaction-history segments and consolidates their representations through direct Evidence pooling and sequential Delta updates. It fuses the paths’ outputs and decodes them into history-conditioned LoRA adapters.

3.1 Overview and Setup

  • A frozen compiler encodes each chronological segment into q_t; consolidation parameters are shared across positions, and queries and answers are excluded from history.Source: §3.1 · pp. 2–3

3.2 Motivation for Dual-Path Memory

  • Under contracting recurrent updates, earlier-segment influence may attenuate; direct pooling mitigates age effects but cannot distinguish reordered histories.Source: §3.2 · p. 3

3.3 Evidence Accumulation

  • Evidence applies coordinate-wise temporal weights to normalized segment representations, pools them directly, and RMS-calibrates the resulting representation.Source: §3.3 · pp. 3–4

3.4 Sequential Delta Updates

  • Delta computes content-derived keys and coordinate-wise gates, reading predictions and writing gated residuals in chronological order; key similarity controls interference.A parallel normalizer follows the state updates before the Delta readout is RMS-calibrated.Source: §3.4 · p. 4

3.5 Fusion and Adapter Construction

  • Learned logits fuse calibrated Evidence and Delta representations, then a frozen decoder produces history-conditioned LoRA factors and adds the shared-adapter update once.Source: §3.5 · p. 4

3.6 Answer Supervision

  • Answer cross-entropy trains both memory paths and their fusion, while the compiler, decoder, shared adapter, and backbone remain frozen.Source: §3.6 · p. 4

4 Experiments

Experiments find that DPPM improves over evaluated baselines across two benchmarks and three backbones. Ablations and further analyses examine the two memory paths, parameter count, category-level behavior, answer efficiency, and latent retention.

4.1 Experimental Setup

  • DPPM was evaluated using Qwen3-4B-Instruct-2507, Gemma-2-2B-IT, and Mistral-7B-Instructv0.2, trained separately per dataset for five epochs with AdamW, and each run’s final checkpoint was evaluated.PersonaMem-v2 has 18,527 training, 2,059 validation, and 5,000 test examples, excluding test personas from training and validation; PrefEval uses 16 training and four held-out topics.Source: §4.1 · pp. 4–5

4.2 Main Results

  • 54.22% on PersonaMem-v2 and 86.79% on PrefEval; DPPM improves over RPMem by 8.50 and 4.32 percentage points, respectively.The gains extend across seven reported metrics and remain positive at every PrefEval dialogue interval, including 3.67 points after 300 intervening turns.Source: Table 1 · p. 5, §4.2 · p. 5
  • DPPM exceeds the strongest textual baseline by 21.26 points on PersonaMem-v2 and 11.30 points on PrefEval, although the leading textual baseline differs by benchmark.Full Context leads on PersonaMem-v2, while Mem0 leads on PrefEval; direct D2L also trails trained consolidation methods.Source: §4.2 · p. 5

4.3 Ablation Study

  • DPPM outperforms RPMem across Gemma, Qwen, and Mistral; with matched Qwen parameter counts, it gains 8.73 and 4.23 points on the two benchmarks.Original RPMem uses 524,800 trainable parameters, and DPPM uses 2,564 more (0.49%).Source: §4.3 · pp. 5–6, Table 3 · p. 6
  • Combining Evidence and Delta improves PersonaMem-v2 Overall over either single path on all three backbones, and PrefEval Overall on Qwen and Mistral.On Qwen, DPPM exceeds the stronger standalone path by 4.45 points on PersonaMem-v2 and 0.81 points on PrefEval.Source: §4.3 · p. 5

4.4 Further Analysis

  • DPPM improves Forget, Sensitive, and Therapy over RPMem by 21.68, 17.57, and 15.34 points; Others accuracy rises from 13.68% to 50.23%.PrefEval gains span all four topics, from 1.90 points on Technology to 6.50 on Education.Source: §4.4 · pp. 6–7, Figure 3 · p. 7
  • At 46.14 ms answer-forward latency, DPPM gains 8.50 points over RPMem; it is also about 3.1× and 62.7× faster than RAG and Full Context.For the longer history, DPPM uses 7,762.32 MiB, 57.46% less peak GPU memory than Full Context.Source: §4.4 · p. 6, Figure 4 · p. 7
  • At age 32, DPPM retains 0.970/0.667 on PersonaMem-v2/PrefEval, versus RPMem’s 0.268/0.116; 64-token OPD scores 54.56% and 87.04%.The 64-token OPD configuration has the highest two-dataset average at 70.80%; short OPD and answer FKL score below CE on both datasets, while GRPO trades off PrefEval accuracy.Source: §4.4 · pp. 7–8, Figure 5 · p. 8, Table 4 · p. 8

5 Related Work

Prior work maintains personalized memory through records, retrieval, and revision, while parametric methods encode histories into model components. DPPM adds direct evidence accumulation alongside sequential revision.

5.1 Personalized Memory

  • Persistent-memory systems store user information and support its use through retrieval, selection, organization, and updating.The evaluation landscape also covers conversational recall and reasoning, preference following, implicit personalization, and evolving preferences.Source: §5.1 · p. 8

5.2 Parametric Memory

  • Parametric-memory approaches range from factual editing and prompt optimization to hypernetworks that generate LoRA parameters in a forward pass.Doc-to-LoRA and SHINE map contexts to adapters without repeated optimization.Source: §5.2 · p. 8
  • Sequential memory methods use associative updates or learned adaptation, while Metis integrates updated memory into the backbone and RPMem connects recurrent consolidation with adapter generation.Source: §5.2 · p. 8
  • DPPM extends recurrent parametric memory with direct evidence accumulation alongside sequential revision, reducing reliance on a single recurrent state.Source: §5.2 · p. 8

6 Conclusion

The conclusion presents DPPM as a cross-session personalization method combining evidence accumulation and sequential updates. Experiments and analysis support its effectiveness and the complementary roles of its two paths.

6 Conclusion

  • DPPM combines evidence accumulation and sequential updates for retaining and revising interaction history without extra compiler training.Source: §6 · p. 8
  • Experiments across two benchmarks and backbones of different architectures and sizes demonstrate DPPM's effectiveness.Source: §6 · p. 8
  • Theoretical analysis and controlled experiments support complementary roles for the two paths, motivating DPPM as a practical design for long-term personalized parametric memory.Source: §6 · p. 8

Limitations

The study's stated limitations are its 2B–7B backbone range and text-only interaction-history interface; it also discloses generative-AI assistance and author review.

Limitations

  • Computational constraints limit experiments to 2B–7B backbones, leaving validation at larger scales for future work.Source: Limitations · p. 9
  • The current memory interface processes textual interaction histories; extending it to images and speech is identified as a next step.Source: Limitations · p. 9

A AI Use Statement

  • Generative AI assisted with implementation, proof drafting, manuscript writing and revision, figure creation, and result-interpretation discussions; authors reviewed the work and take responsibility for its final content.Source: A AI Use Statement · p. 10

B Experimental Configuration

The configuration trains separate memory modules for each dataset and backbone while keeping the compiler, decoder, shared adapter, and backbone frozen. It specifies auxiliary-objective and latency-measurement settings.

B Experimental Configuration

  • Each dataset and backbone receives a separately trained memory module, using chronological segments from the released D2L tokenizer while the compiler, decoder, shared adapter, and backbone remain frozen.Evaluation uses final-epoch checkpoints.Source: B Experimental Configuration · p. 10
  • Table 5 is identified as the source of DPPM cross-entropy training settings.Source: Table 5 · p. 10
  • Auxiliary settings specify answer CE, OPD with teacher-scored reverse KL, and GRPO with correctness rewards; latency is measured on one A800 at batch size 1.OPD uses a one-epoch warmup and weight 0.5; GRPO samples eight answer actions.Source: B Experimental Configuration · p. 10

C Significance Analysis

On Qwen3-4B-Instruct-2507, the analysis compares DPPM with RPMem across seven component metrics using paired training-seed runs. All seven mean differences favor DPPM and remain significant after Holm correction.

C Significance Analysis

  • The comparison pairs DPPM and RPMem runs on Qwen3-4B-Instruct-2507 across seven component metrics, excluding the derived Avg. column.The analysis uses matching seeds 42–46 and two-sided paired t-tests, with Holm correction across the seven-test family.Source: C Significance Analysis · p. 10
  • The tests assume independent seeds and approximately normal seed-level accuracy differences on fixed evaluation sets.The family-wise error rate is controlled at 0.05.Source: C Significance Analysis · p. 10
  • All seven mean differences favor DPPM, with every Holm-adjusted p-value below 0.01.The gains remain significant for both Overall metrics and all five diagnostic or interval-specific measures.Source: C Significance Analysis · p. 10

D Implementation Details

The implementation uses compiled segment representations and specifies evaluation conventions and a latent-retention diagnostic. The analysis gives conditions under which recurrent updates contract perturbations to earlier states.

D Implementation Details

  • Each compiled memory contains vectors indexed by layer, module, and latent slot, with the equations describing one vector across T ≥ 1 segments.Learned projections are shared across positions while memory states remain separate.Source: D Implementation Details · pp. 10–11
  • Evaluation uses seeds 42–46 for main comparisons, while auxiliary-objective variants use seed 42 and restrict privileged teacher state to training.PersonaMem-v2 Overall averages 5,000 questions, and PrefEval Overall averages 1,620 instances; the two-dataset Avg. is the arithmetic mean of the displayed Overall scores.Source: D Implementation Details · p. 11
  • The latent-retention diagnostic moves one preference through five ages in fixed-length histories containing 32 distractors.It uses 24 profiles and frozen seed-42 checkpoints without additional training; sensitivity is normalized by each profile’s age-zero sensitivity.Source: D Implementation Details · p. 11

E.1 Conditional Recurrent Attenuation

  • When κt = γt + RtLt is below one at each update, the gated recurrent state contracts perturbations, accounting for both state retention and gate sensitivity.The bound uses uniform limits on the gate, state-to-input difference, and gate sensitivity; differentiability near the connecting state segment supports the iterated bound.Source: §E.1 · pp. 11–12

E.2 Direct Evidence Sensitivity

Direct Evidence pooling provides a position-independent sensitivity bound under bounded score differences, though that bound decreases with history length. Because pooling is permutation-invariant, ordered revision requires a complementary mechanism.

E.2 Direct Evidence Sensitivity

  • Under bounded-score pooling, if each segment’s scores vary by at most B across t, with B ≥ 0 independent of T, then its weight satisfies w_i,j ≥ [1 + (T − 1)e^B]^-1.Source: §E.2 · p. 12
  • For fixed score range B, the pooling weight lower bound bT = Θ(T −1) is independent of segment position, and the raw Evidence lower bound exceeds the recurrent upper bound when bT > LinκT−i.This comparison concerns changes that leave scores unchanged; Lin bounds input sensitivity and κ bounds contraction.Source: §E.2 · p. 12

E.3 The Need for Ordered Revision

  • Evidence pooling is permutation-invariant, so a predictor using only the query and pooled evidence cannot answer correctly for both orders when those orders require different answers.Oppositely ordered conflicting updates under a last-update-wins rule illustrate why order-sensitive Delta updates are needed.Source: §E.3 · p. 12
  • Direct evidence accumulation and sequential state updates are complementary: the former mitigates recurrent attenuation, while the latter supports ordered revision.The design implication is stated under the analysis’s conditions.Source: §E.3 · p. 12
Loading 2610.11776v1…