CoLink: Global Compression and Knowledge Linking for Multi-Turn Dialogue Modeling

Lian Yan; Zhenbo Xia; Haotian Wang; Chen Tang; Yi Guan; Jie Liu; Jingchi Jiang

cs.CLOct 25, 2026

CoLink: Global Compression and Knowledge Linking for Multi-Turn Dialogue Modeling

Authors: Lian Yan, Zhenbo Xia, Haotian Wang, Chen Tang, Yi Guan, Jie Liu, Jingchi Jiang

Organizations: Research Center of Language Technology, Faculty of Computing, Harbin Institute of Technology, Harbin, Heilongjiang 150001, China · Institute for Advanced Algorithms Research, Shanghai, 201306, China · Research Center of Artificial Intelligence of Things, Faculty of Computing, Harbin Institute of Technology, Harbin, Heilongjiang 150001, China, and also with the National Key Laboratory of Smart Farm Technologies and Systems, Harbin, Heilongjiang 150001, China

Abstract

Extracting salient information from long-horizon dialogue histories remains a significant challenge for Large Language Models (LLMs) due to the restricted context windows and efficient retrieval challenges. Current methods typically rely on history organization or local token-level optimization, suffering from information loss, retrieval redundancy, and myopic attention. Inspired by the view in cognitive science that dialogue constitutes sense-making and memory association, we propose CoLink, a framework that enhances LLMs' global context awareness in domain-specific multi-turn dialogues through information linking and compression. CoLink first clusters historical contexts into topic graphs to establish cross-semantic associations with external knowledge. Guided by the adaptive control of thought, it then retrieves and activates knowledge via synergizing semantic relevance and graph-structural topology. Crucially, to achieve effective compression of history and knowledge, we ground multi-turn dialogues in information theory and derive a principled optimization objective for extracting the core sense-making information. Coupled with a competitive slot mechanism, CoLink condenses the global context into dedicated memory and knowledge slots, efficiently distilling lengthy, redundant inputs into an extremely compact latent representation. Extensive empirical results on both public and our MMD datasets validate CoLink's superiority and generalizability in knowledge linking and sense modeling.

Figures & tables

Fig. 1: Rigid Retrieval & Local Tokens vs. History Compression in CoLink.

Fig. 2: Cognitive isomorphism between human memory mechanisms and the CoLink framework. The diagram illustrates two key biological parallels: (1) Associative Retrieval: Analogous to the hippocampus retrieving episodic memories to activate long-term cortical knowledge, CoLink constructs topic graphs to link historical contexts with external knowledge, dynamically activating relevant facts via the semantic relevance and structural linkage importance. (2) Abstract Encoding: Just as the human brain stores highly abstract semantic schemas rather than verbatim textual records, CoLink employs an global competitive compression strategy. This rigorously distills raw history and knowledge streams into compact Memory/Knowledge slots, mirroring the brain’s efficiency in retaining core semantics while filtering redundancy.

Fig. 4: Example of the data construction process.

Dataset	Domain	Lang.	# Samples	# Diseases	Avg. Turns	Debate Turns
MMD	Medical	Chinese	2,370	37	8.09	2.77
CMtMedQA	Medical	Chinese	5,993	14	8.85	–

TABLE I: Dataset statistics.

Hyperparameter	Description	Value
$E_{1}$	Stage 1 Epochs	$10$
$E_{2}$	Stage 2 Epochs	$10$
$\eta_{1}$	Stage 1 Learning Rate	$5e-5$
$\eta_{2}$	Stage 2 Learning Rate	$1e-5$
$w_{d}$	Weight Decay	$0.01$
$B$	Batch Size	$4$

TABLE II: Hyperparameter settings for CoLink.

Method	MMD (%)								CMtMedQA (%)
Method	B1	B2	B3	B4	R1	R2	RL	BS	B1	B2	B3	B4	R1	R2	RL	BS
MemoryBank	19.41	8.55	4.47	2.32	23.31	5.24	14.98	68.83	27.20	14.20	8.34	4.90	29.69	9.06	21.67	73.75
NV-embed	17.77	8.10	4.22	2.31	23.34	5.60	15.35	68.22	23.82	12.18	7.14	4.19	27.57	8.17	20.09	72.10
StreamLLM	0.38	0.06	0.03	0.02	1.75	0.02	1.69	45.97	0.60	0.11	0.06	0.04	3.70	0.21	0.99	46.08
GraphRAG	16.45	6.64	3.24	1.65	21.57	4.37	14.15	67.91	24.10	12.02	6.90	3.94	27.25	7.64	19.45	72.41
CCA	8.62	2.99	1.46	0.78	13.61	2.07	8.75	64.14	4.14	1.32	0.56	0.27	11.39	1.49	7.98	58.67

TABLE III: The comparative results. We highlight the best performance. B, R, and BS stand for BLEU, Rouge, and Bertscore, respectively.

Method	MMD			CMtMedQA
Method	History/Knowledge Processing Time (ms)	GPU Memory (MB)	Compression Ratio	History/Knowledge Processing Time (ms)	GPU Memory (MB)	Compression Ratio
MemoryBank	2113.57	/	0.6388	4521.04	/	0.4884
NV-embed	218.41	27674.82	0.8095	202.69	27612.20	0.7732
StreamLLM	/	3050.95	0.9665	/	3064.59	1.0000
GraphRAG	41511.27	/	2.8946	47679.49	/	2.4866
CCA	/	3679.53	0.8903	/	3724.54	1.0000

TABLE IV: Comparison of online resource consumption and compression ratios.

Input Encode	BLEU (%)				ROUGE (%)			BS (%)
Input Encode	B1	B2	B3	B4	R1	R2	RL	BS (%)
Base(Full History)	18.21	11.02	7.84	4.66	30.03	10.13	15.31	69.38
Base(Summary History)	21.89	13.11	7.65	5.57	32.36	11.21	17.09	70.46
Memory Slot(Ours)	24.21	14.04	8.60	5.02	31.82	10.44	17.70	70.25

TABLE V: Transfer performance of CoLink on Llama-3.2-3B, involving only first-stage memory slot training.

Input Encode	BLEU (%)				ROUGE (%)			BS (%)
Input Encode	B1	B2	B3	B4	R1	R2	RL	BS (%)
Base(Full History)	23.28	12.67	7.97	5.12	28.72	8.62	25.60	74.96
Base(Summary History)	18.18	9.78	6.03	3.75	30.38	9.02	27.03	74.16
Memory Slot(Ours)	38.49	23.16	15.59	10.48	39.42	14.52	35.33	78.68

TABLE VI: Transfer performance of CoLink on English open-domain multi-turn dialogue dataset.

Method	BLEU (%)				ROUGE (%)			BS (%)
Method	B1	B2	B3	B4	R1	R2	RL	BS (%)
CoLink (stage1)	26.40	13.40	7.88	4.54	29.09	8.63	20.52	73.30

TABLE VII: Generalization across different datasets.

Fig. 5: Attention and loss saliency of summary history and memory slots towards the query and next-token prediction.

Method	Human				LLM-as-Judge
Method	Accuracy & Professionalism	Informativeness	Contextual Coherence	Overall	Accuracy & Professionalism	Informativeness	Contextual Coherence	Overall
MemoryBank	3.77	3.08	4.28	3.71	3.94	3.49	4.46	3.96
NV-embed	3.56	2.72	4.14	3.47	3.82	3.24	4.31	3.79
GraphRAG	3.54	2.44	4.65	3.54	3.21	2.56	4.42	3.40
CCA	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00
MemTree	2.80	1.84	4.40	3.01	3.78	3.03	4.70	3.84

TABLE VIII: Human and LLM-as-a-Judge evaluation on Accuracy & Professionalism, Informativeness, and Contextual Coherence.

History Len.	MMD (%)								CMtMedQA (%)
History Len.	B1	B2	B3	B4	R1	R2	RL	BS	B1	B2	B3	B4	R1	R2	RL	BS
CoLink(Ours)	27.52	13.87	7.66	4.10	31.63	8.64	19.26	72.84	31.62	18.02	11.74	7.62	34.28	12.31	25.50	75.36
w/o Memory Slots	24.13	11.42	6.26	3.35	28.35	7.08	17.87	70.94	28.73	15.83	9.96	6.18	31.89	10.81	23.74	74.33
w/o Knowledge Slots	27.61	13.82	7.71	3.96	31.88	8.54	19.33	72.78	30.08	17.00	10.91	6.97	32.73	11.62	24.01	74.81

TABLE IX: Performance after ablating memory slots and knowledge slots.

Method	Accuracy (%)
Base	79.08
+Full Knowledge	82.00
+Summary Knowledge	59.83
CoLink (Knowledge Slots)	90.37

TABLE X: Performance Comparison on the GMD Dataset with Precisely Aligned Knowledge Graphs.

Method	BLEU (%)				ROUGE (%)			BS (%)
Method	B1	B2	B3	B4	R1	R2	RL	BS (%)
CoLink (stage1)	23.21	10.96	5.66	3.01	27.90	6.84	17.60	70.88
w/o $\mathcal{L}_{rec}$	16.74	6.60	3.24	1.73	21.36	3.88	14.36	67.84
w/o $\mathcal{L}_{KL}$	25.65	11.81	5.99	3.15	29.27	6.81	17.33	71.36
w/o $\mathcal{L}_{pred}$	17.13	6.74	3.23	1.60	22.54	4.19	15.25	67.73

TABLE XI: Ablation study on different loss terms.

$\alpha$	BLEU (%)				ROUGE (%)			BS (%)
$\alpha$	B1	B2	B3	B4	R1	R2	RL	BS (%)
$10.0$	20.30	8.74	4.40	2.18	24.30	5.11	15.74	69.28
$0.1$	21.32	9.40	4.96	2.63	26.46	5.95	16.73	69.57

TABLE XII: The performance after scaling

\mathbb{L}_{KL}

Fig. 6: Training loss trajectories under different ablation settings.

Input Encode	BLEU (%)				ROUGE (%)			BS (%)
Input Encode	B1	B2	B3	B4	R1	R2	RL	BS (%)
Layer 0 Hidden State	15.86	6.00	2.87	1.41	21.07	3.68	14.79	67.42
nn.embedding	17.14	6.74	3.21	1.60	21.64	3.67	14.43	67.93
Last Layer Hidden State	23.21	10.96	5.66	3.01	27.90	6.84	17.60	70.88

TABLE XIII: Impact of Input Encoding Methods on Model Performance

Fig. 7: Comparison of generations for a case across different models.

Method	BLEU (%)				ROUGE (%)			BS (%)
Method	B1	B2	B3	B4	R1	R2	RL	BS (%)
History Cluster	25.18	11.56	5.86	3.09	28.98	6.76	17.75	71.18
Full History	25.58	11.86	6.15	3.20	28.88	6.86	17.61	71.34

TABLE XIV: Performance comparison between compressing Retrieved Topic Clusters and Full History.

Method	BLEU (%)				ROUGE (%)			BS (%)
Method	B1	B2	B3	B4	R1	R2	RL	BS (%)
AdaptSlot (w/o Sparse Loss)	6.65	2.51	1.43	0.88	15.45	2.36	10.72	60.53
AdaptSlot	17.09	6.49	3.28	1.70	22.38	3.76	14.19	67.33
CoLink	23.21	10.96	5.66	3.01	27.90	6.84	17.60	70.88

TABLE XV: AdaptSlot vs. Fixed-Slot.

Language Modeling

Language modeling develops systems that learn patterns in text and other sequences to understand, predict, and generate language.

See all 522 papers

Improving Multi-turn Dialogue Consistency with Self-Recall Thinking
May 14, 2026Renning Pang, Tian Lan, Leyuan Liu +3Memory-Augmented Language Models Long-Term Conversational Memory
University of Electronic Science and Technology of China
Context-Driven Incremental Compression for Multi-Turn Dialogue Generation
Jun 10, 2026Yeongseo Jung, Jaehyeok Kim, Eunseo Jung +5Truncated Backpropagation Through Time Memory-Augmented Language Models
Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong · NVIDIA AI Technology Center, NVIDIA · University of Tsukuba, Japan +1
StateTree: Enhancing Long-Term Dialogue Reasoning via Reinforcement Learning
Sep 30, 2026Naen Xu, Wanqing Cui, Yibo Hu +5Temporal Reasoning in Language Models RL for Language Model Reasoning
Zhejiang University · Taobao & Tmall Group of Alibaba

May 14, 2026J/K move · Enter open · S save

Improving Multi-turn Dialogue Consistency with Self-Recall Thinking

Renning Pang, Tian Lan, Leyuan Liu, Xiaoming Huang+2

University of Electronic Science and Technology of China

Memory-Augmented Language Models Long-Term Conversational Memory Multi-Turn Dialogue Management

Open paper AskPDF arXiv

Large language model (LLM) based multi-turn dialogue systems often struggle to track dependencies across non-adjacent turns, undermining both consistency and scalability. As conversations lengthen, essential information becomes sparse and is buried in irrelevant context, while processing the entire dialogue history incurs severe efficiency bottlenecks. Existing solutions either rely on high latency external memory or lose fine-grained details through iterative summarization. In this paper, we propose Self-Recall Thinking (SRT), a framework designed to address long-range contextual dependency and sparse informative signals in multi-turn dialogue. SRT identifies helpful historical turns and uses them to generate contextually appropriate responses, enabling the model to selectively recall and reason over context during inference. This process yields an endogenous reasoning process that integrates interpretable recall steps without external modules. SRT incorporates: (1) Dependency Construction: Generating and converting it into self-recall chains; (2)Capability Initialization: Training to enable reasoning chains with recall tokens capability; (3)Reasoning Improvement: Refining accuracy via verifiable rewards to optimize recall and reasoning for correct answers. Experiments on multiple datasets demonstrate that SRT improves F1 score by 4.7% and reduces end-to-end latency by 14.7% over prior methods, achieving a balance between reasoning latency and accuracy, and outperforming state-of-the-art baselines.

Abstract

Language Modeling

Explore similar work

Improving Multi-turn Dialogue Consistency with Self-Recall Thinking

Context-Driven Incremental Compression for Multi-Turn Dialogue Generation

StateTree: Enhancing Long-Term Dialogue Reasoning via Reinforcement Learning