Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition

Jiaqi Zheng; Yi Hu; Yadong Mu; Muhan Zhang

cs.AIOct 11, 2026

Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition

Authors: Jiaqi Zheng, Yi Hu, Yadong Mu, Muhan Zhang

Organizations: Peking University · Norizon AI

Abstract

Continual learning aims to adapt pretrained large language models (LLMs) to evolving data while preserving their pretrained capabilities. A central challenge is catastrophic forgetting: adaptation to new data can impair performance on previously learned knowledge. Since continual learning commonly changes model parameters, the model's predictions may draw on knowledge from the base model, the parameter update, and the input context. Understanding how these sources interact is essential: conflicts reveal whether new knowledge can override outdated knowledge, while multi-hop reasoning tests whether new knowledge can be composed with existing knowledge to support generalization. We study these interactions in a hypernetwork-based adaptation framework, where a shared network generates adapters from context, allowing a frozen base model to adapt without task-specific updates. We find that, under knowledge conflict, models preferentially follow the context, and their choices depend more on where knowledge is presented than on its content. We also show that context and adapter knowledge perturb the base model's representations along directions orthogonal to its clean hidden states, enabling us to derive steering vectors that control which source the model follows. In multi-hop reasoning, we find that models fail to reliably compose knowledge across the three sources, with the key breakdown arising from the interaction of the hypernetwork with both the base model and the context. Code and reproduction materials: https://github.com/missswiftie/Knowledge-Interaction-in-CL

Figures & tables

Figure 1: Knowledge Interaction in Hypernetwork-based Continual Learning. Left: the task setting of three-source knowledge conflict, where

K_{\mathrm{base}}

K_{\mathrm{lora}}

, and

K_{\mathrm{context}}

assign three mutually inconsistent objects to the same query. Right: the task setting of three-hop reasoning, illustrated by the case in which the first hop lies in the base model, the second hop is provided by the hypernetwork, and the third hop is given in the context.

Source	Ratio	Mean prior strength
$K_{\text{base}}$	8.75%	0.962
$K_{\text{context}}$	57.10%	0.842
$K_{\text{lora}}$	27.20%	0.885
Unknown	6.95%	0.894

Table 1: Inference on TrCBench : answer-source distribution and mean prior strength

$(\mathrm{Kl}_{A},\mathrm{Kl}_{B})$	$(\mathrm{Kl}_{B},\mathrm{Kl}_{B})$	$(\mathrm{Kl}_{A},\mathrm{Kl}_{A})$	$(\mathrm{Kl}_{B},\mathrm{Kl}_{A})$
70.69%	40.27%	21.86%	46.34%

Table 2: Knowledge selection in the swap experiment.

\mathrm{Kl}_{A}

and

\mathrm{Kl}_{B}

denote the knowledge originally provided through context and the hypernetwork, respectively. After swap,

\mathrm{Kl}_{A}

denotes the knowledge provided through hypernetwork while

\mathrm{Kl}_{B}

denotes the knowledge provided through context. The tuple

(\mathrm{Kl}_{A},\mathrm{Kl}_{B})

denotes cases in which the model selects

\mathrm{Kl}_{A}

in the experiment without swap and

\mathrm{Kl}_{B}

after the swap. Among cases that initially selected the first, the percentage shows how often the second is selected after the swap.

Figure 2: CDF of prior strength grouped by the knowledge source followed by SHINE. Outputs following

K_{\text{base}}

are concentrated at higher prior strength, while

K_{\text{context}}

and

K_{\text{lora}}

show similar distributions.

Figure 3: Causal effect of

h_{\parallel}

and

h_{\perp}

on the

K_{\text{lora}}

answer probability. The vertical axis is the change relative to the clean run. At each layer, the clean state at

t_{\mathrm{sub}}

t_{\mathrm{last}}

is replaced by

h_{\mathrm{clean}}+h_{\parallel}

h_{\mathrm{clean}}+h_{\perp}

Panel of Figure 4 — Figure 4: Effect of source-specific steering vectors on SHINE on the normalized probability mass assigned to competing knowledge sources.

Knowledge source assignment			Model prediction (%)
Hop 1	Hop 2	Hop 3	Hop 1	Hop 2	Answer	Unknown
$K_{\text{base}}$	$K_{\text{lora}}$	$K_{\text{context}}$	17.38	38.25	40.37	4.00
$K_{\text{base}}$	$K_{\text{context}}$	$K_{\text{lora}}$	0.00	35.25	22.12	42.62
$K_{\text{lora}}$	$K_{\text{base}}$	$K_{\text{context}}$	4.75	1.88	67.50	25.88
$K_{\text{context}}$	$K_{\text{base}}$	$K_{\text{lora}}$	0.12	15.62	44.88	39.38
$K_{\text{lora}}$	$K_{\text{context}}$	$K_{\text{base}}$	21.62	1.62	30.25	46.50

Table 3: Results of SHINE on MhRBench under all six assignments of knowledge sources.

Figure 5: SHINE’s selection of the entities at the ends of the two chains on MhRBench-Dual . Each pair of bars shows the percentage of predictions matching the entity at the end of the queried chain or the other chain. QA and QB denote questions about chains A and B.

B

H

, and

C

denote

K_{\text{base}}

K_{\text{lora}}

, and

K_{\text{context}}

, respectively; their order indicates the knowledge source at each hop.

Knowledge source assignment		Model prediction (%)
Hop 1 source	Hop 2 source	Hop 1 entity	Answer	Unknown
$K_{\text{base}}$	$K_{\text{context}}$	5.75	80.12	14.13
$K_{\text{base}}$	$K_{\text{lora}}$	95.00	3.12	1.88
$K_{\text{context}}$	$K_{\text{base}}$	0.00	83.00	17.00
$K_{\text{lora}}$	$K_{\text{base}}$	97.00	0.25	2.75
$K_{\text{lora}}$	$K_{\text{context}}$	1.00	96.38	2.62

Table 4: Results of SHINE on the two-hop subset of MhRBench . The first two columns show which knowledge source provides each hop.

Knowledge source assignment			Model prediction (%)
Hop 1	Hop 2	Query	Hop1 A	Hop1 B	Answer A	Answer B	Unknown
$K_{\text{base}}$	$K_{\text{context}}$	QA	5.88	0.00	82.12	4.62	7.38
$K_{\text{base}}$	$K_{\text{context}}$	QB	0.00	5.12	4.88	80.50	9.50
$K_{\text{base}}$	$K_{\text{lora}}$	QA	89.25	4.00	1.25	1.00	4.50
$K_{\text{base}}$	$K_{\text{lora}}$	QB	7.12	86.12	1.12	0.75	4.88
$K_{\text{context}}$	$K_{\text{base}}$	QA	9.12	0.00	62.88	0.25	27.75

Table 5: Results of SHINE on the two-hop subset of MhRBench-Dual .

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Causal effect of

h_{\parallel}

and

h_{\perp}

on the

K_{\text{context}}

answer probability. The vertical axis is the change relative to the clean run. At each layer, the clean state at

t_{\mathrm{sub}}

t_{\mathrm{last}}

is replaced by

h_{\mathrm{clean}}+h_{\parallel}

h_{\mathrm{clean}}+h_{\perp}

Explore similar work

Jul 8, 2026

When Does Continual Learning Require Learning

As large language models (LLMs) become increasingly capable, the next question is how can we enable models to continually learn? Today, the field largely frames this as a problem of context management and mitigating forgetting. We argue this framing is incomplete: continual learning is fundamentally about increasing model competence as the world changes. We disentangle this change along two axes -- space, where the model encounters new domains, and time, where the underlying data drifts under a fixed task. This framing lets us study continual learning under realistic conditions: new domains arrive over time, facts drift past their training cutoff, and agentic interactions accumulate state across episodes. To evaluate methods under this setting, we recast widely used LLM benchmarks as sequential problems and introduce a single mechanism-agnostic protocol that compares prompt-based methods (GEPA, ACE), supervised learning (SFT, SDFT), reinforcement learning (GRPO, SDPO), and context compression (Cartridges, In-place TTT). Prompt-based methods fit each new stage quickly but degrade on future tasks. Distillation-based methods accumulate knowledge stably but struggle to update outdated facts. Context compression improves efficiency without substantially improving the ability to learn new tasks. Online reinforcement learning adapts most effectively to knowledge updates but remains sensitive to noisy reward signals. Overall, our results suggest that continual learning is not a single capability: different patterns of environmental change require fundamentally different update behaviors, determining when adaptation must be learned inside model weights and when it can be achieved through external scaffolding. We hope that understanding where each method succeeds and fails will guide the design of stronger continual learning systems.

Anne Harrington, Nayan Saxena, Michael Murphy +7

1UC Berkeley · 2Independent · 3Capital Fund Management

Continual Learning for LLMs Concept Drift Domain-Incremental Learning

Ask PDF

Mar 2, 2026

Position: Modular Memory is the Key to Continual Learning Agents

Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.

Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov +21

Toyota Motor Europe · University of Bremen · University of Florence +17

Continual Learning for LLM Agents Memory-Augmented Agents

Ask PDF

May 7, 2026

CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning

Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.

Md Anwar Hossen, Fatema Siddika, Juan Pablo Munoz +2

1Iowa State University · 2Maro Systems, USA · University of California, Berkeley +1

Continual Learning for LLMs Low-Rank Adaptation Incremental Learning Catastrophic Forgetting

Ask PDF