AuthorsXiaoyuan Yi, Maosong Sun, Ruoyu Li, Zonghan Yang
TL;DR
Working Memory model uses topic, local, and history memories with a topic trace mechanism to reach BLEU 1.315 on quatrains vs 0.425 for iPoet.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing poetry generators often pack all user inputs and history into a single small vector or attend over unlimited history, which is implausible for coherence.
On Chinese poetry generation, this causes weak topic expression and poor line-to-line coherence, leaving a large gap between computer-generated and human-authored poems.
HOW IT WORKS
Working Memory model introduces a topic memory, history memory, local memory, a genre embedding, and a Topic Trace mechanism around a GRU encoder–decoder.
You can think of Working Memory model as a writer’s working memory, where a few mental slots hold current topics and key past phrases, instead of one overloaded summary.
This design lets Working Memory model focus on relevant slots, flexibly reuse or skip topics, and maintain long-poem coherence beyond what a fixed context window can handle.
DIAGRAM
This diagram shows how Working Memory model generates each poem line, reads from memories, and writes salient history back into the history memory.
DIAGRAM
This diagram shows how Working Memory model is trained and evaluated on quatrains, iambics, and chinoiserie lyrics with BLEU and perplexity.
PROCESS
Working Memory Model Overview
Working Memory model first encodes user topic words into the topic memory and initializes the global trace vector and Topic Trace mechanism.
Working Memory Model
During each line generation, Working Memory model uses the Addressing Function to read from topic, local, and history memories into the decoder.
Memory Writing
After a line is generated, Working Memory model writes encoder hidden states of the previous line into local memory and salient states into history memory slots.
Topic Trace Mechanism
Across lines, Working Memory model updates the Topic Trace mechanism to track which topic slots were read and how often, guiding future memory addressing.
KEY CONTRIBUTIONS
Working Memory model for poetry generation
Working Memory model introduces explicit topic, history, and local memory modules with limited slots, dynamically read and written via an Addressing Function to improve coherence.
Genre embedding for structural control
Working Memory model uses a genre embedding that concatenates a 64 dimensional phonology embedding and a 32 dimensional length embedding to control tones and line lengths.
Topic Trace mechanism
Working Memory model adds a Topic Trace mechanism with a 24 dimensional topic trace vector to record topic usage, improving BLEU from 1.267 to 1.315 on quatrains.
RESULTS
BLEU
1.315
+0.890 over iPoet
PP
86
-52 perplexity vs iPoet
BLEU iambics
0.699
+0.379 over iambicGen
BLEU lyrics
0.568
+0.256 over lyricGen
On quatrains, iambics, and chinoiserie lyrics, Working Memory model is evaluated with BLEU and perplexity against iPoet, iambicGen, and lyricGen. These results show that Working Memory model substantially improves topic expression and coherence across three Chinese poetry genres.
BENCHMARK
On quatrains, iambics, and chinoiserie lyrics, Working Memory model is evaluated with BLEU and perplexity against iPoet, iambicGen, and lyricGen. These results show that Working Memory model substantially improves topic expression and coherence across three Chinese poetry genres.
BENCHMARK
BLEU scores on quatrains comparing Working Memory model to strong neural baselines.
KEY INSIGHT
Working Memory model without genre embedding already surpasses baselines, with BLEU 1.019 on quatrains versus iPoet’s 0.425 despite weaker structural control.
This is surprising because one might expect structural and phonological supervision to dominate, yet limited-slot working memory alone yields a +0.594 BLEU gain over iPoet.
WHY IT MATTERS
Working Memory model shows that small, dynamically written memories can maintain long-range coherence in generative language tasks like multi-line poetry.
Builders can now design generators that explicitly track topics and salient history across many segments, instead of relying on a single history vector or unbounded attention.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…