AuthorsKai Zhang, Yangyang Kang, Fubang Zhao, Xiaozhong Liu
TL;DR
MaLP uses a Dual-Process enhanced Memory plus PEFT to personalize medical assistants, reaching 91.53% win rate on response generation with LLaMA-7B (+13.12 over LoRA).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory-based assistants rely on rigid dictionary-based memory, which MaLP shows is inflexible and heavily dependent on retriever quality.
Without retraining, these systems barely provide personalized and engaging experiences, so patients with different dialogue preferences receive mismatched advice and degraded user satisfaction.
HOW IT WORKS
MaLP centers on Dual-Process enhanced Memory (DPeM) coordinating Working Memory, Short-Term Memory (STM), Long-Term Memory (LTM), Coordinator C, and Retriever R around a PEFT-tuned LLM.
DPeM mirrors human cognition: Working Memory acts like RAM, STM like a scratchpad, and LTM like disk, with rehearsal and executive processes moving knowledge between them.
This dual-process structure lets MaLP store user-specific and common-sense knowledge beyond a fixed context window, enabling retrieval-aware personalization that static prompts and single-layer memories cannot match.
DIAGRAM
This diagram shows how MaLP’s DPeM runs Learning, Summarizing, and Memorizing via Rehearsal and Executive processes over dialogue iterations.
DIAGRAM
This diagram shows how MaLP injects medical knowledge, builds DPeM memory, applies LoRA, and evaluates on QA, preference classification, and response generation.
PROCESS
Medical Knowledge Adaptation
MaLP first trains a domain adapter on HealthCareMagic-100k and iCliniq, constraining changes with knowledge loss and sample loss to preserve the base LLM’s capabilities.
DPeM Mechanism
MaLP applies the Dual-Process enhanced Memory, where Coordinator C writes notes into Working Memory, filters them into STM, and promotes frequent items into LTM.
Memory Generation
Using user dialogue D, MaLP forms Mworking, MST M, and MLT M tables, storing user-specific and common-sense knowledge with flag tables tracking access frequency.
Memory Utilization
At query time, Retriever R uses Rc on STM and Rs on LTM to build prompt p, then feeds x and p into the LoRA-tuned LLM ˆΦ to generate personalized response y.
KEY CONTRIBUTIONS
Dual-Process enhanced Memory mechanism
MaLP introduces DPeM with Working Memory, STM, LTM, Coordinator C, and Retriever R, yielding about 7% relative improvement over dictionary-based memory on key metrics.
MaLP unified frame with PEFT
MaLP combines DPeM with LoRA-based PEFT, letting the LLM internalize user preferences while DPeM supplies structured user-specific and common-sense memory.
Personalized medical dialogue dataset
MaLP constructs a self-chat medical dataset with user profiles and preferences, achieving an average quality score of 5.27 and a 94% safety ratio on 100 samples.
RESULTS
ROUGE-L Profile QA
33.91
+4.25 over LLaMA-7B w LoRA
ROUGE-L Knowledge QA
36.37
+2.77 over LLaMA-7B w LoRA
Preference Classification Accuracy %
69.95
+8.90 over LLaMA-7B w LoRA
Response Generation Win Rate %
91.53
+19.52 over LLaMA-7B w LoRA
On the MaLP medical dialogue benchmark with LLaMA-7B, MaLP is evaluated on Profile QA, Knowledge QA, Preference Classification, and Response Generation. The gains over the LoRA baseline show that combining DPeM with PEFT substantially improves both factual correctness and alignment with user preferences.
BENCHMARK
On the MaLP medical dialogue benchmark with LLaMA-7B, MaLP is evaluated on Profile QA, Knowledge QA, Preference Classification, and Response Generation. The gains over the LoRA baseline show that combining DPeM with PEFT substantially improves both factual correctness and alignment with user preferences.
BENCHMARK
Win Rate % for response generation with different configurations of LLaMA-7B on the MaLP medical dialogue benchmark.
BENCHMARK
Preference Classification Accuracy % for LLaMA-7B with different MaLP modules enabled.
KEY INSIGHT
MaLP with LoRA alone reaches 61.05% preference classification accuracy, but still trails MaLP’s full 69.95% despite strong user preference modeling.
This is surprising because many assume PEFT is sufficient for personalization, yet MaLP shows that structured DPeM memory still adds an extra 8.90 percentage points.
WHY IT MATTERS
MaLP unlocks a way to coordinate short- and long-term personalization, combining DPeM memory with PEFT so assistants remember both preferences and stable medical facts.
Builders can now create medical agents that adapt to each user over many sessions without retraining full LLMs, using MaLP’s memory tables and LoRA layers as lightweight personalization handles.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…