C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning

Yassir Lairgi; Ludovic Moncla; Khalid Benabdeslem; Rémy Cazabet; Pierre Cléau

cs.AIAug 27, 2026

C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning

Authors: Yassir Lairgi, Ludovic Moncla, Khalid Benabdeslem, Rémy Cazabet, Pierre Cléau

Organizations: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205, 69621 Villeurbanne, France · GAUC, Lyon, France

Abstract

Weak signals are early, low-visibility indicators that precede significant changes before those changes become established. Existing detection methods, based on keyword frequency, topic modeling, or untyped graph topology, fail to capture the semantic and relational structure through which such signals manifest. In this paper, we propose C-Unseen, a self-interpretable framework for weak signal detection in Dynamic Temporal Knowledge Graphs (DTKGs). We define a weak signal as a rare, semantically coherent subgraph that proliferates across consecutive TKG snapshots. The framework operates through two modules: a Rare Subgraphs Extractor, in which an LLM identifies subgraphs whose content is in tension with the dominant snapshot narrative via chain-of-thought reasoning, and a Weak Signal Alerter, in which the persistence of these rare subgraphs is tracked across time steps to isolate true weak signals. Experimental results demonstrate that C-Unseen outperforms keyword-, topic-, and graph-based baselines.

Figures & tables

Panel of Figure 1 — Figure 1: Weak signals precede a strong signal as fragmentary, low-visibility traces whose significance emerges only from their corroboration.

Figure 2: The architecture of our framework

ID	Description	$t^{\text{weak}}_{s}$	$t^{\text{strong}}_{s}$
SS_FORPROFIT_2019	Conversion to capped-profit (OpenAI LP)	Dec 2015	Mar 2019
SS_BOARD_COUP_2023	Altman removal and reinstatement	Mar 2019	Nov 2023
SS_NYT_LAWSUIT_2023	NYT v. OpenAI copyright suit	Jun 2021	Dec 2023
SS_DEFENCE_TURN_2025	Pivot to defence and national security	Jul 2019	Jun 2025
SS_FORPROFIT_PBC_2025	Conversion to for-profit PBC	Dec 2017	Oct 2025

Table 1: Strong signals annotated in Wiki-OpenAI. The SS date corresponds to

t_{\text{est}}

, the date on which the event was publicly confirmed.

Corpus
Property	Value
Source	Wikipedia (OpenAI page)
Period	Jan 2015 – Dec 2025
Number of yearly snapshots	11
Atomic facts
Total	773

Table 2: Summary statistics of the Wiki-OpenAI benchmark dataset.

$k$	Method	$\|\mathcal{D}_{M}\|$	$\|\mathcal{S}_{M}^{+}(k)\|$	$P_{M}(k)$	$R_{M}(k)$	$F_{1,M}(k)$
1	BEAM	19	3	0.789	0.600	0.682
	BERTrend	44.8 $\pm$ 5.2	4.20 $\pm$ 0.45	0.161 $\pm$ 0.021	0.840 $\pm$ 0.089	0.270 $\pm$ 0.034
	Yoon	311	3	0.032	0.600	0.061
	C-Unseen	50.2 $\pm$ 10.1	5.00	0.444 $\pm$ 0.066	1.000	0.613 $\pm$ 0.061
2	BEAM	19	1	0.053	0.200	0.083
	BERTrend	44.8 $\pm$ 5.2	2.80 $\pm$ 1.30	0.081 $\pm$ 0.026	0.560 $\pm$ 0.261	0.141 $\pm$ 0.047

Table 3: Detection accuracy on Wiki-OpenAI at shared-anchor-word match thresholds

k\in\{1,2,3\}

. C-Unseen and BERTrend are reported as mean

\pm

sample std over

n{=}5

runs; Yoon and BEAM are deterministic and reported from a single run. “–” denotes an undefined value.

$k$	Method	$\|\mathcal{D}_{M}\|$	$\|\mathcal{S}_{M}^{+}(k)\|$	$P_{M}(k)$	$R_{M}(k)$	$F_{1,M}(k)$
1	BEAM	20	3	0.750	0.600	0.667
	BERTrend	49.8 $\pm$ 2.0	4.00	0.120 $\pm$ 0.035	0.800	0.207 $\pm$ 0.053
	Yoon	273	3	0.022	0.600	0.042
	C-Unseen	58.4 $\pm$ 19.9	5.00	0.434 $\pm$ 0.076	1.000	0.603 $\pm$ 0.073
2	BEAM	20	1	0.050	0.200	0.080
	BERTrend	49.8 $\pm$ 2.0	3.20 $\pm$ 0.45	0.092 $\pm$ 0.028	0.640 $\pm$ 0.089	0.159 $\pm$ 0.044

Table 4: Detection accuracy on Wiki-OpenAI-Anon . Conventions as in Table 3 .

Method	Wiki-OpenAI	Wiki-OpenAI-Anon
BEAM	–	–
BERTrend	1.70 $\pm$ 0.45	2.00
Yoon	–	–
C-Unseen	1.00	1.13 $\pm$ 0.30

Table 5: Mean lead time

\bar{L}_{M}^{\text{years}}(k{=}3)

over covered signals. “–” denotes no covered signal (mean undefined).

Figure 3: Qualitative comparison of the baseline methods’ output over the detected SS_DEFENCE_TURN_2025 by Yoon, BERTrend, and C-Unseen . It was not detected by BEAM.

Arm	$\|\mathcal{D}_{M}\|$	$\|\mathcal{S}_{M}^{+}\|$	$P_{M}$	$F_{1,M}$	$\Delta F_{1}$
local+flat	144.8 $\pm$ 18.5	5.00	0.059 $\pm$ 0.013	0.112 $\pm$ 0.023	$-$ 0.159
local+CoT	135.2 $\pm$ 19.9	4.80 $\pm$ 0.45	0.072 $\pm$ 0.010	0.134 $\pm$ 0.018	$-$ 0.137 $\pm$ 0.089
full DTKG+flat	195.0 $\pm$ 14.1	5.00	0.061 $\pm$ 0.009	0.115 $\pm$ 0.016	$-$ 0.156
full DTKG+CoT	189.6 $\pm$ 10.5	5.00	0.069 $\pm$ 0.013	0.130 $\pm$ 0.023	$-$ 0.141 $\pm$ 0.093
C-Unseen	50.2 $\pm$ 10.1	4.20 $\pm$ 1.10	0.162 $\pm$ 0.053	0.271 $\pm$ 0.084	—

Table 6: Ablation on Wiki-OpenAI at

k{=}3

, mean

\pm

sample std over

n{=}5

runs.

Temporal KGs

Temporal knowledge graphs represent time-stamped facts and events to support reasoning about how relationships change and what may happen next.

See all 69 papers

AdaTKG: Adaptive Memory for Temporal Knowledge Graph Reasoning
May 8, 2026Seunghan Lee, Jun Seo, Jaehoon Lee +7Graph Representation Learning Temporal KGs
LG AI Research Seoul, South Korea
Learning Across Temporal Graph Representation Silos
Oct 14, 2026Minwoo Yu, Young-guk HaDynamic Graph Learning Temporal KGs
M. Yu and Y.-g. Ha are with Smart Computing Laboratory, Department of Computer Science & Engineering, Konkuk University, Seoul 05029, Republic of Korea.
LLM-Guided Knowledge Distillation for Temporal Knowledge Graph Reasoning
Feb 16, 2026Wang Xing, Wei Song, Siyu Lin +2Knowledge Distillation Temporal KGs
School of Computer Science and Technology, Xidian University · School of Computing and Artificial Intelligence, Southwest Jiaotong University · School of Information Science and Engineering, Chongqing Jiaotong University

May 8, 2026J/K move · Enter open · S save

AdaTKG: Adaptive Memory for Temporal Knowledge Graph Reasoning

Seunghan Lee, Jun Seo, Jaehoon Lee, Sungdong Yoo+6

LG AI Research Seoul, South Korea

Graph Representation Learning Temporal KGs

Open paper AskPDF arXiv

Temporal knowledge graphs (TKGs) represent time-stamped relational facts and support a wide range of reasoning tasks over evolving events. However, existing methods produce entity representations that are static at the entity level, in that each representation is a function of learned parameters only and retains no trace of the interactions in which the entity has participated. In this paper, we depart from this static view and propose that each entity be modeled as an adaptive process whose representation is refined every time the entity participates in a fact. To this end, we propose AdaTKG, which maintains a per-entity memory that is updated with every observed interaction, with the memory accumulating online and predictions improving as more interactions arrive. Specifically, we instantiate the memory update as a learnable exponential moving average governed by a single shared scalar instead of using learnable parameters for each entity, enabling AdaTKG to handle entities unseen during training. Extensive experiments confirm consistent gains over TKG baselines, demonstrating the effectiveness of adaptive memory. Code is available at: https://github.com/seunghan96/AdaTKG

Abstract

Temporal KGs

Explore similar work

AdaTKG: Adaptive Memory for Temporal Knowledge Graph Reasoning

Learning Across Temporal Graph Representation Silos

LLM-Guided Knowledge Distillation for Temporal Knowledge Graph Reasoning