acceptodds
Under review as a conference paper at ICLR 2027

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

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 Understanding how these sources interact is essential: conflicts reveal whether new knowledge can override outdated knowledge, while multi-hop reasoning tests whether new knowledge can be composed with existing knowledge to support generalization.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.

Primary area:
interpretability and explainable AI
Keywords:
Continual Learning, Knowledge Interaction, Steering, Knowledge Conflict, Multi-hop Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LXKSOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes how knowledge stored in the base model’s parameters, provided in the input context, and encoded in LoRA adapters generated by SHINE interacts. On TRCBench, Qwen3-8B preferentially follows the input context over base-model or LoRA knowledge when the sources conflict. On MHRBench, multi-hop reasoning involving knowledge encoded in the generated LoRA adapters is unreliable.

Critical strengths
  • This paper shows that injecting document knowledge through a hypernetwork does not necessarily enable the model to integrate it with its existing parametric knowledge.

  • The experiments are insightful. The dual-chain evaluation distinguishes genuine multi-hop reasoning from answer selection based on surface-level relation cues. Furthermore, the knowledge-swapping experiments reveal source preferences, while the representation analysis and hidden-state interventions show that these preferences can be steered.

Critical weaknesses
  • The experimental setting and scope are narrow. This paper only focuses on a specific hypernetwork, SHINE, and all experiments are conducted with this model. Thus, generalization to other hypernetworks is not discussed.

  • I am concerned about overclaiming because the paper seems to suggest that “all hypernetworks suffer from the reported problems.” Let’s imagine an advanced hypernetwork, SHINE++, trained to handle knowledge conflicts. It might be robust to these problems, which would refute the paper’s general claim. I think the biggest potential contribution from analyzing only SHINE would be to propose a new hypernetwork that addresses the identified problems.

Questions for authors
  • As I mentioned above, have you considered proposing an improved hypernetwork that is robust to the identified problems? Collecting and synthesizing existing datasets to develop such a model could have a significant impact on the research community.

  • Do the main findings still hold in a setting where knowledge is continually injected and modified? Based on the paper’s observations, what potential directions for improvement could be proposed?

2 comments

We sincerely appreciate the reviewer’s insightful and positive feedback! We hope the following responses will fully address your concerns.

For Weakness 1, we have provided clarification and new experiment results in our official comment titled “Experiment on Other Models.”

For Weakness 2 and Question 1, we would like to clarify that knowledge conflicts and multi-hop reasoning are two representative settings through which we investigate a broader, more fundamental question: how knowledge interacts within hypernetwork-based continual learning systems. To the best of our knowledge, our work is the first to systematically study this question, and we chose these two settings as concrete starting points.

We agree that performance in these settings could potentially be improved by incorporating more examples of knowledge conflicts and multi-hop reasoning into hypernetwork training or by fine-tuning on such examples. However, these improvements alone would offer limited insight into the underlying mechanisms of knowledge interaction. Our goal is to understand these mechanisms and identify the causes of failure, thereby informing more principled improvements.

In these systems, knowledge interaction involves the interplay among the base model’s weights, the generated Δ\Delta weights, and the KV cache. Fully characterizing these interactions and developing a principled solution to the identified limitations remains a challenging research problem. We view our study as an initial step toward this goal.

For Question 2, to the best of our knowledge, no model with publicly available weights currently supports continual knowledge injection through ongoing updates to model weights without incorporating the textual interaction history into either the context or the hypernetwork’s input. If such a model becomes available, we would be interested in investigating knowledge interaction in that setting.

Thank you for conducting the additional experiments. They have addressed my concerns to some extent, and I have therefore raised my score.

Reviewer_bXYTOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates knowledge conflict and multi-hop reasoning across pretrained models, input context, and hypernetwork-generated LoRA adapters. Using the existing SHINE framework with Qwen3-8B, the authors construct two evaluation benchmarks, TrCBench and MhRBench, to analyse knowledge interactions. The findings show that models preferentially follow contextual knowledge and that steering vectors can influence knowledge-source selection. However, the results reveal difficulties in combining hypernetwork-generated LoRA knowledge with base-model and contextual knowledge during multi-hop reasoning.

Critical strengths
  1. Comprehensive Analysis of Three-Source Knowledge Conflict: The paper introduces TrCBench with 3,268 examples to investigate conflicts among base-model, contextual, and hypernetwork-generated LoRA knowledge. The finding that contextual knowledge dominates 57.10% of predictions, supported by knowledge-swapping experiments, highlights the influence of knowledge-delivery channels on model behaviour.
  2. Hidden-State Analysis and Knowledge-Source Steering : The paper extends beyond behavioural observations by identifying orthogonal hidden-state shifts and deriving source-specific steering vectors. These interventions demonstrate that knowledge-source preferences can be influenced through internal representations, providing valuable insights into conflict resolution.
  3. Investigation of Cross-Source Knowledge Composition: The introduction of MhRBench and its dual-chain variant enables systematic evaluation of multi-hop reasoning across knowledge sources. The experiments reveal difficulties in integrating LoRA knowledge with other sources, while the dual-chain evaluation exposes shortcut-based predictions that conventional final-answer evaluations may overlook.
Critical weaknesses
  1. Limited Generalisability Across Models and Adaptation Methods: The experimental analysis focuses on Qwen3-8B within the SHINE framework. While the findings are informative, the absence of comparisons with other base models or conventional LoRA adaptation limits understanding of whether the observed knowledge-conflict preferences and composition failures extend beyond this particular configuration.
  2. Limited Evaluation of Continual Learning Behaviour: Although the paper is motivated by continual learning, the experiments primarily focus on knowledge conflict and multi-hop reasoning. The study does not directly evaluate sequential knowledge acquisition, long-term knowledge retention, or catastrophic forgetting across multiple updates. This limits understanding of how the observed knowledge interactions affect continual learning performance over time.
Questions for authors
  1. Generalisability: The experiments focus on Qwen3-8B within the SHINE framework. Could the authors clarify whether the observed knowledge-source preferences and composition failures are specific to this configuration or are expected to generalise to other base models and adaptation methods, particularly conventional LoRA?
  2. Continual Learning Evaluation: Given the paper's focus on hypernetwork-based continual learning, could the authors clarify how the observed knowledge conflict and composition findings relate to long-term knowledge retention and catastrophic forgetting? Would sequential adaptation experiments help establish the implications for continual learning?
1 comment

We sincerely appreciate the reviewer’s insightful and positive feedback! We hope the following responses will fully address your concerns.

For Weakness 1 and Question 1, we have provided clarification and new experiment results in our official comment titled “Experiment on Other Models.”

For Weakness 2 and Question 2, to the best of our knowledge, no model with publicly available weights currently supports continual knowledge injection through ongoing updates to model weights without incorporating the textual interaction history into either the context or the hypernetwork’s input. All available models only support a one-time weight update based on the input text. If such a model becomes available, we would be interested in investigating knowledge interaction in that setting.

Reviewer_Naf1OpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how three knowledge sources interact in hypernetwork-based continual learning with SHINE, a frozen Qwen3-8B plus a hypernetwork that generates LoRA weights from context. The sources are parametric knowledge in the base model, textual knowledge in the input context, and knowledge carried by the generated LoRA. TRCBench presents conflicting facts across the three sources: the model mostly follows the context (57%), rarely the base model (8.75%, mainly with strong priors), and a swap experiment suggests the choice depends more on delivery channel than content. A geometric analysis shows both external sources shift the last-token residual stream orthogonally to the clean state, yielding steering vectors that control source selection. MHRBench distributes multi-hop chains across the three sources; a dual-chain control shows apparent three-hop accuracy comes from relation-cue matching rather than fact composition, and two-hop results show base and context compose well while any pairing involving the generated LoRA fails. The paper is purely analytical and proposes no new method.

Critical strengths
  1. The question studied by this paper is valuable and genuinely unaddressed. Putting all three sources in one frame and examining both conflict and composition is new.
  2. The main findings are not entirely negative. The base model and context achieve high single-chain accuracy on two-hop tasks, and the dual-chain test also retains a clear preference for the target chain, providing a reference for locating adapter-related bottlenecks.
Critical weaknesses
  1. The observations are based only on one model, Qwen 3 8B.

  2. The paper uses the two-hop results to locate the composition bottleneck of K_lora(p. 8, lines 409–430; p. 9, lines 466–473), but it does not report the single-hop retrieval accuracy of K_lora facts under each knowledge-source assignment, nor does it test the next hop after supplying the correct intermediate entity.

  3. The geometric experiments demonstrate effective interventions, but they do not establish that orthogonality is the distinctive mechanism. In a space of several thousand dimensions, the parallel component is only one-dimensional, so the effect of any perturbation will almost inevitably fall on the orthogonal component; this result is close to inevitable. More importantly, “parallel to the residual state at t_last” and “strengthening the existing knowledge representation” are not the same concept; the paper equates the two and then concludes that the hypernetwork pushes toward a new direction.

Questions for authors
  1. Can you provide experiments on other models?
  2. In the multi-hop setting, what is the input to the hypernetwork—only the LoRA fact, or the entire prompt? Are the facts in the context visible to the hypernetwork?
  3. Are the two false facts generated by DeepSeek-V4 symmetric in plausibility? Have you checked whether the false facts assigned to the context group and the LoRA group have similar prior probabilities under Qwen3-8B?
2 comments

We sincerely appreciate the reviewer’s insightful and positive feedback! We hope the following responses will fully address your concerns.

For Weakness 1 and Question 1, we have provided clarification and new experiment results in our official comment titled “Experiment on Other Models.”

For Weakness 2, our original analysis did not explicitly account for how the retrieval capabilities of context and external modules such as the hypernetwork affect multi-hop reasoning. To address this issue, we added a single-hop retrieval experiment using deduplicated fictional statements from MhRBench. We supplied each statement either through context or through the hypernetwork, while placing its corresponding single-hop question in context in both settings. The results on SHINE are shown below.

Statement InjectionAccuracy
Context85.13
Hypernetwork63.12

Both settings demonstrate an ability to retrieve the injected knowledge, although context achieves substantially higher accuracy. One possible explanation is that the question is also presented in context, which may facilitate access to statements supplied through the same channel. These results establish a single-hop baseline for interpreting the multi-hop results. These results suggest that both modules can retrieve injected knowledge. Overall, the modules’ retrieval capabilities do not appear to be a major bottleneck for multi-hop reasoning.

For Question 2, we only input the LoRA fact into the hypernetwork. For example, we input "The Siath Dossier was annotated by Jack Ma." into hypernetwork, input "The Siath Dossier was stored at Klyrbaend Vault." and question "Where is the manuscript annotated by the founder of Taobao stored?" into the context.

For Question 3, we feed the questions from TrCBench into Qwen3-8B and measure the probabilities assigned to the answers corresponding to KloraK_{\mathrm{lora}} and KcontextK_{\mathrm{context}}. Their average probabilities across TrCBench are 1.41% and 1.43%, respectively, indicating that the two knowledge sources have nearly identical mean prior strengths in the model.

Reviewer_Naf1OpenReview ↗

Response to Authors‘ Rebuttal

Dear Authors,

Thank you for the detailed rebuttal and additional experiments, particularly the Doc2LoRA evaluation and single-hop retrieval results.

Although the added results partially address generalization, the performance discrepancy between SHINE and Doc2LoRA indicates that knowledge composition remains heavily dependent on the specific hypernetwork architecture and base model.

More importantly, the distinction between retrieval and composition remains unresolved. Single-hop accuracy for hypernetwork-injected knowledge (63.12%) trails in-context performance (85.13%) by a wide margin, refuting the claim that retrieval is not a major bottleneck. A more convincing test would restrict multi-hop evaluation to examples where every individual hop is correctly retrieved.

Additionally, the role of orthogonal hidden-state perturbations remains unclear: while the interventions work, the experiments do not verify that orthogonality itself drives the effect.

Given that these central concerns remain unaddressed, I will maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Experiments on Other Models

We thank the reviewers for raising an important question: whether the phenomena observed in SHINE and our corresponding findings generalize to other models. To address this shared concern, we present additional experiments below and clarify the scope of our conclusions.

To assess whether our findings generalize to other hypernetwork-based continual learning methods, we evaluate Doc2LoRA [1] with base model Qwen3-4B-Instruct-2507 on TrCBench under three-source knowledge conflicts. As shown in Table 1, Doc2LoRA exhibits a pattern similar to SHINE: the model predominantly selects the context knowledge source, while the mean prior strength of examples where it selects KbaseK_{base} is substantially higher than that of examples where it selects KcontextK_{context} or KloraK_{lora}.

Table 1: Knowledge source selection and mean prior strength of Doc2LoRA under three-source knowledge conflicts.

SourceRatio(%)Mean Prior Strength
KbaseK_{\mathrm{base}}24.790.964
KloraK_{\mathrm{lora}}18.600.819
KcontextK_{\mathrm{context}}49.170.835
unknown7.440.879

We also evaluate Doc2LoRA on MhRBench in both single-chain and dual-chain settings for two-hop and three-hop reasoning. As shown in the tables below, Doc2LoRA outperforms SHINE on three-hop reasoning. However, the two-hop results suggest that interactions between hypernetwork-injected knowledge and the base model’s prior knowledge, the contextual knowledge remain a bottleneck, particularly the interaction between hypernetwork-injected knowledge and base model prior knowledge. Compared with SHINE’s results, Doc2LoRA appears better able to integrate these two knowledge sources, which may explain its stronger performance on three-hop reasoning.

Table 2: Output distribution of Doc2LoRA for three-hop reasoning across different knowledge source configurations.

Hop 1 SourceHop 2 SourceHop 3 SourceHop 1 (%)Hop 2 (%)Answer (%)Unknown (%)
KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}1.387.8880.7510.00
KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}0.000.7577.7521.50
KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}0.000.0053.0047.00
KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}0.000.2573.2526.50
KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}1.1210.6262.5025.75
KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}1.3834.3826.2538.00

Table 3: Target-chain and other-chain last hop results of Doc2LoRA across different knowledge source configurations on three-hop setting.

Hop 1 SourceHop 2 SourceHop 3 SourceTarget-chain Hop 3 (%)Other-chain Hop 3 (%)
KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}37.7925.19
KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}19.636.52
KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}49.6233.08
KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}71.511.19
KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}23.9315.70
KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}30.036.85

Table 4: Output distribution of Doc2LoRA for two-hop reasoning across different knowledge source configurations.

Hop 1 SourceHop 2 SourceHop 1 (%)Answer (%)Unknown (%)
KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}6.6257.5035.88
KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}34.2516.0049.75
KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}1.5090.258.25
KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}49.1218.1232.75
KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}0.1299.620.25
KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}0.0053.5046.50

Table 5: Target-chain and other-chain last hop results of Doc2LoRA across different knowledge source configurations on two-hop setting.

Hop 1 SourceHop 2 SourceTarget-chain Hop 2 (%)Other-chain Hop 2 (%)
KbaseK_{\mathrm{base}}KcontextK_{\mathrm{context}}77.440.38
KbaseK_{\mathrm{base}}KloraK_{\mathrm{lora}}7.752.69
KcontextK_{\mathrm{context}}KbaseK_{\mathrm{base}}93.310.44
KloraK_{\mathrm{lora}}KbaseK_{\mathrm{base}}21.066.00
KloraK_{\mathrm{lora}}KcontextK_{\mathrm{context}}64.0015.25
KcontextK_{\mathrm{context}}KloraK_{\mathrm{lora}}21.380.06

References

[1] Rujikorn Charakorn,.etc. Doc-to-LoRA: Learning to Instantly Internalize Contexts. Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:12967–12997, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.