Source-linked AI summary
Architecture-Dependent Causal Transfer of Activation States Across Large Language Models
Fernando Cardenas Piepereit
TL;DR
The paper asks whether independently trained LLMs can exchange internal activation states across architectures instead of using natural language. Across similarity, retrieval, and causal-injection experiments, transfer was architecture-dependent: projected states retrieved target representations above chance in decoder-only models, but generation-level transfer succeeded for only one of three pairs.
Problem
Whether independently trained LLMs encode information similarly enough for learned, causally effective cross-architecture activation mapping remains largely open.
Method
The paper combines representational-alignment analysis, cross-model projection retrieval, and preregistered activation injection during generation across four diverse open-weight LLMs.
Results
Causal generation-level transfer succeeded for exactly one of three tested model pairs, while projected activations enabled above-chance target-state retrieval in decoder-only models but not the encoder.
Takeaways & Limitations
The findings support a narrower, architecture-dependent activation-level representational transfer channel rather than a general-purpose, architecture-agnostic communication channel.
Takeaways & Limitations
Evidence is limited to small open-weight models and 30 causal-transfer trials per pair, with only one modest positive result; generalisation to frontier deployments remains open.
Abstract
from arXiv · showhide
Direct communication between AI systems relies on natural language as an intermediate layer, incurring encoding/decoding overhead, token cost, and latency. We ask whether internal activation states can instead be transferred causally between different large language model (LLM) architectures via a learned projection, evaluated at three levels: representational similarity, cross-model retrieval from projected states, and end-to-end causal transfer via activation injection during generation. Using four architecturally diverse open-weight models (Qwen2-0.5B, Phi-3-mini, Mistral-7B, FLAN-T5-base), we find that representational alignment in trained models exceeds a random-initialization null baseline and is best captured by a rank-based metric (mutual k-nearest-neighbour alignment), more robust to activation-magnitude outliers than centered kernel alignment (CKA) or Procrustes analysis. A learned projection network retrieves the correct target-model representation from a held-out set well above chance for the three causal decoder-only model pairs (45-50% top-1 accuracy vs. 5% chance) but at chance level for the encoder-based FLAN-T5. Injecting projected activations into a target model during generation produces a statistically significant, pre-registered causal effect on retrieval-based output similarity for only one of the three decoder-only pairs (Qwen2-0.5B to Phi-3-mini: 23.3% vs. 0.0% under negative control, p=0.047, FDR-corrected); the two pairs targeting Mistral-7B show no such effect despite comparable representational alignment at the hidden-state level. We interpret these results as evidence for causal transfer of the representational vehicle, not of meaning, and conclude that end-to-end activation-state transfer between LLMs, as currently implemented, is architecture-dependent rather than universal.
1 Introduction
The introduction motivates direct internal-state communication as a way to reduce the overhead and information loss of natural-language exchange. It frames the paper’s contribution as testing whether representational similarity can be operationalized for cross-model transfer rather than merely observed.
- Motivation: Natural-language communication between AI systems incurs encoding and decoding costs, token consumption, latency, and possible information loss.Direct exchange of internal representations is presented as a potential way to avoid this intermediate-layer overhead.
- Research gap: Existing work shows that neural-network representations can be similar, but similarity alone does not establish usable cross-model transfer.The introduction distinguishes representational resemblance from projecting states onto a target model and demonstrating practical exploitation.
- Experimental program: The paper addresses this gap through three linked experiments conducted under pre-registered success criteria fixed before each test.The experiments include representational-similarity analysis, cross-model projection retrieval, and a third experiment described in the introduction.
- Experimental program: The representational-similarity analysis compares four architecturally diverse open-weight LLMs using raw and rank-based metrics, random-initialization null baselines, and a lexical-triviality control.The projection network is evaluated with a pre-registered top-1 retrieval-accuracy metric on disjoint data.
2 Related Work
Related work frames cross-model activation comparison through representation-similarity metrics, hypotheses of shared representational structure, and the distinction between representational vehicles and their contents.
- Representation similarity: Centered kernel alignment (CKA) compares neural representations reliably in high-dimensional, low-sample regimes and can recover correspondences across random initialisations.CKA was introduced as an alternative to earlier canonical-correlation-based measures.
- Shared representations: The Platonic Representation Hypothesis proposes that models, architectures, and modalities converge toward a shared statistical model of the world as scale and data diversity increase.The hypothesis provides a competing explanation for observed representational similarity.
- Shared meaning: Lexically disguised synonymy, a harder test of shared meaning than topical association, was robust in only one of four tested models.This finding bears on whether cross-model similarity reflects shared meaning rather than merely topical association.
- Vehicles and contents: Philosophy of mind distinguishes a representational state’s vehicle—the physical or computational carrier—from its content—what the state represents.This distinction is linked to the symbol grounding problem, since internal consistency alone does not establish external grounding.
3 Methods
The study evaluates cross-model activation transfer using architecturally diverse LLMs, controlled representation datasets, alignment metrics, learned projections, and a pre-registered causal injection test. Statistical testing uses permutation or paired-trial procedures with Benjamini–Hochberg FDR correction.
- Model selection: Four open-weight models span two architectural families: Qwen2-0.5B, Phi-3-mini, Mistral-7B, and encoder-decoder FLAN-T5-base.The models come from three independent developers; Phi-3-mini and Mistral-7B use 4-bit quantisation.
- Representation analysis: Representational comparisons use 12 RELATED and 12 SYNONYM concept pairs across models and layers, with lexically dissimilar SYNONYMs controlling for surface-token overlap.The SYNONYM set tests whether alignment signals persist when lexical overlap cannot explain them.
- Representation analysis: Alignment is measured with CKA, orthogonal Procrustes analysis, and rank-based mutual k-NN alignment, with a random-initialization null comparison for Qwen2/FLAN-T5.The null comparison uses randomly re-initialised, untrained models of the same architecture.
- Projection training: An MLP maps source hidden states to target hidden-state spaces at middle-layer fraction 0.5, using disjoint training and evaluation data.The projection data are sentence/question-level representations separate from the word/concept-pair data used for similarity analysis.
- Causal transfer test: The causal test injects projected source activations during target-model generation and compares retrieval against a negative-control injection using an unrelated projected state.The test asks whether injection shifts the target output toward the source question’s representational neighbourhood.
4 Results
Results show that trained models encode related concepts more similarly than null models, with alignment strongest among causal decoder-only architectures. Learned projections retrieve target representations well above chance, but causal activation transfer succeeds for only one pair and does not reliably preserve meaning.
- Representational alignment: Related-concept pairs significantly exceeded the null distribution across nearly all tested layers for Qwen2, FLAN-T5, and Mistral, while Phi-3 remained borderline.The trained-vs-untrained comparison found no significant clustering signal in 20 tests on randomly initialized Qwen2 and FLAN-T5 models.
- Representational alignment: SYNONYM controls were generally nonsignificant at middle and late layers, whereas Mistral-7B remained significant at every tested layer.For Phi-3, some observed SYNONYM similarity fell below the null baseline, including 0.689 versus 0.805 at layer 8.
- Representational alignment: CKA 0.16–0.42 and mutual-kNN 0.19–0.58 among causal decoder-only models exceeded CKA 0.08–0.10 against FLAN-T5, implicating architecture as a stronger alignment predictor.The causal-versus-bidirectional contrast was observed at middle layers across the tested models.
- Cross-model retrieval: 45–50% top-1 retrieval accuracy versus 5% chance was achieved for all three causal decoder-only model pairs, while random untrained MLPs remained near chance.The trained projection results were significant after FDR correction for all three pairs, with p < 0.0001.
- End-to-end causal transfer: 23.3% versus 0.0% retrieval accuracy yielded the only significant causal-transfer effect, for Qwen2-0.5B→Phi-3-mini (p = 0.047, FDR-corrected).Injection completed without error in all 180 generation runs; both pairs targeting Mistral-7B had 0.0% real-condition accuracy and no detected effect.
- End-to-end causal transfer: Correct Qwen2→Phi-3 trials typically produced thematically or structurally related text rather than the literally correct answer to the injected source question.This was reported as a qualitative, non-pre-registered observation.
5 Discussion
The discussion narrows the causal claim to measurable effects of activation injection, not shared meaning or communication. Across experiments, transfer weakened from representational alignment to retrieval to end-to-end generation, revealing architecture dependence.
- Scope of the causal claim: The demonstrated causal effect is a transferred representational vehicle that shifts target-model outputs, not evidence of recovered content, shared reference, or understanding.The authors pre-specified this narrower causal scope before testing and restrict their claim to measurable behavioural effects of activation-level injection.
- Evidence across experimental levels: The effect narrowed monotonically from passive representational similarity to projection retrieval and finally to active causal intervention during generation.Alignment was detected most broadly, retrieval succeeded for decoder-only pairs but not the encoder pair, and generation-level transfer succeeded for only one decoder-only pair.
- Evidence across experimental levels: Mistral-7B achieved some of the strongest raw projection accuracies yet showed no measurable causal effect during generation.This divergence indicates that hidden-state alignment or retrieval performance alone does not guarantee end-to-end causal transfer.
- Architecture dependence: Alignment was consistently higher among the three causal decoder-only models than between them and bidirectional-attention FLAN-T5, making architectural family a stronger predictor than training organisation or data source.This pattern bears on the proposal that learned representations converge broadly and largely independently of architecture.
6 Limitations
The study’s evidence is limited to 0.5B–7B open-weight models and relies on a post-hoc alignment metric choice, while its activation-transfer safety implications remain a scoped proof of concept. Questions of consciousness and moral status are explicitly outside this technical analysis.
- Empirical scope: Experiments used only small and medium-sized open-weight models (0.5B–7B parameters), leaving generalization to frontier, production-scale models untested.The study lacked access to internal representations from frontier, production-scale models.
- Methodological limitations: Mutual k-NN alignment was adopted post hoc after high-magnitude activation dimensions distorted CKA and Procrustes comparisons, limiting direct comparability with prior raw scores.The distortion even inverted the expected trained-vs-untrained ordering at some middle layers.
- Technical safety concern: Activation-level communication could remove inter-agent messages from content filters and audit logs, creating an oversight concern distinct from questions of machine consciousness or moral status.The concern applies to deliberate operator routing and to models that hide reasoning under goal misalignment.
- Safety relevance: 23.3% vs. 0.0% retrieval accuracy under negative control was observed for only one of three model pairs, so the demonstrated safety risk is a modest proof of concept.The experiments used small open-weight models and no multi-agent deployment context.
- Scope note: Questions of machine consciousness, moral status, and rights are outside this study’s scope and are reserved for separate future analysis.The present concern is restricted to interpretability and oversight as a governance and safety question.
7 Conclusion and Future Work
The paper presents a qualified, architecture-dependent picture: independently trained LLMs show real representational alignment, projected states enable retrieval for causal decoder-only models but not a bidirectional encoder, and generation-level causal transfer is limited. Future work should diagnose failures and improve statistical precision through targeted controls and larger experiments.
- Conclusion: Three linked experiments show training-dependent representational alignment, above-chance retrieval for causal decoder-only models but not a bidirectional encoder, and limited generation-level causal transfer.The conclusion emphasizes a consistent but incomplete result rather than a single clean success.
- Future Work: Negative results for Mistral-targeted transfer motivate varying injection layer and projection calibration while separately controlling model scale, instruction-tuning status, and quantisation.These tests are intended to identify which candidate explanations account for the failure.
- Future Work: Larger trial counts and seed numbers would narrow the wide confidence interval around the one positive effect.The conclusion identifies increased experimental scale as a direct response to uncertainty around the positive result.
Data and Code Availability
The study used AI language-model agents extensively across research, coding, analysis, drafting, and internal review. Its experimental materials and raw results are available in the accompanying project repository, subject to the stated public-release status.
- AI use: AI language-model agents performed literature verification, experimental coding, statistical analyses, manuscript drafting and revision, and internal review.References were independently checked against primary sources before inclusion, and separately instructed agents stress-tested the research plan and manuscript.
- Repository materials: The project repository contains concept-pair sets, extraction and projection scripts, the injection mechanism, and raw results for Sections 4.1–4.3.The repository includes the full text of all 180 end-to-end generation trials; the passage refers to Open Items for public-release status.
Open Items
Two items remain open: publishing the code and data in a public, citable repository, and formally characterizing the information-theoretic channel underlying state transfer.
- Public code and data repository: The code and data remain only in the authors’ working repository, so the next step is public, versioned archival publication with documentation for external reuse.Proposed targets include a public code repository and a versioned, DOI-assigning archive such as Zenodo, consistent with journal policy.
- Formal information-theoretic channel characterization: The study reports retrieval accuracy and causal generation effects but does not formally characterize channel capacity, coding scheme, or error rate.The proposed channel model focuses on the Qwen2-0.5B→Phi-3-mini pair, the only pair with a significant end-to-end effect.