Long-term memory for LLM agents must answer two questions for every stored fact: whether the fact is still true and whether it should be surfaced for a given query. We refer to these two dimensions as validity and visibility, respectively. Existing memory systems often collapse these two decisions into a single scalar score. However, current and historical queries may require opposite retrieval decisions for the same fact: a superseded fact should be hidden for a current-intent query but surfaced for a historical-intent one. This creates a structural failure mode: superseded and expired facts are collapsed onto the same validity axis. To address this problem, we propose MemFSM, a training-free retrieval-side layer that separates validity from visibility through a memory-lifecycle finite-state machine. Along the validity axis, a four-state FSM tracks whether each fact has been superseded and expired. Along the visibility axis, an intent-aware router decides whether to surface or hide a fact based on its state and the query intent. By decoupling visibility from validity, MemFSM improves performance across heterogeneous memory benchmarks and diverse LLM backbones. To evaluate memory systems under fact conflicts and diverse query intents, we further introduce ValidMem, a benchmark that annotates each fact with both supersedure and expiry labels and pairs each memory store with both current-intent and historical-intent queries.

We propose a unified, multi-stage lay summarization pipeline for BioLaySumm 2025 (Subtask 1.1) that (1) selects and summarizes key article sections via BioBART, (2) retrieves K-shot demonstrations using BGE embeddings for in-context Llama 3 8B prompting, (3) applies LoRA adapters to Llama 3 8B for supervised fine-tuning, (4) merges section summaries with a second BioBART pass, and (5) refines outputs through reinforcement learning (PPO & GRPO) using a composite reward of factuality (AlignScore, SummaC), relevance (ROUGE-L, BERTScore), and readability (LENS, FKGL, DCRS, CLI). On PLOS and eLife validation sets, our complete systemreduces DCRS from 9.23 to 8.56 and reduces CLI from 12.98 to 12.65, ranking 3rd in readability. and outperforms llama3 finetune baseline in AlignScore 0.722 to 0.862, ranking 5th in factuality, demonstrating balanced gains across readability, relevance, and factuality.

Biomedical literature are crucial for disseminating new scientific findings. However, the complexity of these research articles often leads to misinterpretations by the public. To address this urgent issue, we participated in the BioLaySumm task at the 2024 ACL BioNLP workshop, which focuses on automatically simplifying technical biomedical articles for non-technical audiences. We conduct a systematic evaluation of the SOTA large language models (LLMs) in 2024 and found that LLMs can generally achieve better readability scores than smaller models like Bart. Then we iteratively developed techniques of title infusing, K-shot prompting , LLM rewriting and instruction finetuning to further boost readability while balancing factuality and relevance. Notably, our submission achieved the first place in readability at the workshop, and among the top-3 teams with the highest readability scores, we have the best overall rank. Here, we present our experiments and findings on how to effectively adapt LLMs for automatic lay summarization. Our code is available at https://github.com/zhoujieli/biolaysumm.