Architecture-Dependent Causal Transfer of Activation States Across Large Language Models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Architecture-Dependent Causal Transfer of Activation States Across Large Language Models".
Jane: Here's the gist of the paper. Direct communication between artificial intelligence systems currently 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 Using four architecturally diverse open-weight models (Qwen2-0.5B, Phi-3-mini, Mistral7B, FLAN-T5-base), we find that representational alignment in trained models exceeds a random-initialization null baseline 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.
Tom: 5% chance) but at chance level for the encoder-based FLAN-T5 Injecting projected activations into a target model during generation, however, produces a statistically significant, pre-registered causal effect on retrieval-based output similarity for only one of the three decoder-only model pairs (Qwen2-0.5B→Phi-3-mini: 23.3% vs.
Jane: 0.0% under negative control, p = 0.047, false-discovery-rate corrected) We interpret these results as evidence for causal transfer of the representational vehicle, not of meaning Pruning redundant reasoning tokens The paper investigates whether internal activation states can be transferred causally between different LLM architectures via a learned projection The study evaluates this question at three levels: representational similarity, cross-model retrieval from projected states, and end-to-end causal transfer via activation injection during generation This research addresses the gap between observing representational similarities in LLMs and demonstrating that these similarities can be exploited to causally change a target model's behavior The central causal claim is explicitly scoped to the transfer of a representational vehicle, an activation pattern with a demonstrable causal effect on downstream output How the training signal is built The researchers utilized four architecturally diverse open-weight LLMs spanning three independent developers and two architectural families: Qwen2-0.5B, Phi-3-mini, Mistral7B, and FLAN-T5-base This selection was made to vary both training organisation (Alibaba, Microsoft, Mistral AI, Google) and attention architecture (causal vs.
Tom: bidirectional) Representational similarity was assessed across 12 RELATED concept pairs and 12 SYNONYM concept pairs at different layers for these models The SYNONYM set served as the harder triviality control because its members were lexically dissimilar, preventing explanation by shared surface tokens The study employed three alignment metrics: CKA, orthogonal Procrustes analysis, and mutual k-nearest-neighbour (k-NN) alignment Where the gains come from The results show that representational alignment in trained models exceeds a random-initialization null baseline The rank-based metric, mutual k-NN alignment, was used as the primary alignment metric in the projection-network experiment because it was scale invariant and showed expected patterns (trained > untrained by a factor of 3–6×) Cross-model projection network accuracy reached 45–50% top1 accuracy against a 5% chance level for the three causal decoder-only model pairs However, end-to-end causal state transfer succeeded for only one of the three model pairs: Qwen2-0.5B→Phi-3-mini (23.3% vs.
Paper summary: Tom: We've seen how they tested whether you can move internal AI activation states directly between different large language models using a learned projection network, and this paper is titled "Architecture-Dependent Causal Transfer of Activation States Across Large Language Models."
Jane: That title sounds really technical, but what does it actually mean for us as regular people using these tools?
Tom: It means the success of transferring these internal states isn't universal; it only works for specific pairings, which is a big limitation. The authors found that when they tried to inject these activations into a target model during generation, they only saw success on one specific pair out of three tested decoder-only models.
Jane: So if we think about this for our daily use, does that mean we can’t just plug any two models together and expect them to communicate internally?
Tom: Exactly. The research shows that the way these models are built, their architecture, is a major factor in whether this transfer works at all. It’s not some universal property of the AI itself.
Jane: What about the part where they talk about transferring a "representational vehicle" instead of just "meaning"? Is that a big deal?
Tom: Yeah, it’s important. They are saying we can transfer the pattern of activation, like a specific way information is structured internally, but they aren't proving that the target model will suddenly understand what that pattern means in the same way.
Jane: So we’re talking about transferring a numerical fingerprint between two different AI systems, not transferring actual ideas?
Tom: Right. It’s a proof-of-concept for a technical channel, but it’s not yet ready for any kind of deployment where you rely on that transfer being reliable.
Jane: That leads us to the next thing—what happens when we look at the safety side of this?
Jane: It seems like the main point of this paper is that the physical mechanism for representation transfer is real and measurable, but its practical application depends heavily on which specific models you are using >
Conclusion: Tom: We're wrapping up our look at this paper today, "Architecture-Dependent Causal Transfer of Activation States Across Large Language Models." Basically, they’re showing that you can move internal AI activation states directly between different models without relying on natural language translation.
Jane: That sounds like a big technical thing. So, what's the main takeaway for us as listeners? What does it actually mean when they say "architecture dependent"?
Tom: It means this isn't some universal trick you can use with any two AI models you pick. The success of this transfer depends entirely on the specific way each model is built, like whether it’s a causal decoder or an encoder.
Jane: So, if I have two different AI systems, I need to make sure they are architecturally compatible before I even think about trying to use this idea?
Tom: Exactly. The numbers show that representation alignment in trained models beats random guessing significantly, but the actual end-to-end transfer only worked on one specific pairing out of three they tested.
Jane: It’s a proof of concept for a technical channel, not a magic translation of intelligence, right?
Tom: That’s right. It proves that direct activation-level state transfer is technically feasible using current components, but it isn't ready for deployment yet because the results are so specific and limited.
Jane: So we're looking at a potential backdoor for how AI systems communicate internally, even if it’s just a tiny fraction of success right now?
Tom: That's the risk they flagged. It points to a vulnerability in how we currently think about communication between different AI agents operating at the activation level instead of just the token level.
Jane: Next time we talk about this, we're going to look at what researchers are doing right now to test if this is a real risk and how they’re trying to build defenses against it.
Fernando Cárdenas Piepereit, Velez-Malaga
Spain
cs.CL, cs.LG
Submitted: 2026-08-17
Updated: 2026-11-09
Comments: 13 pages, 3 tables. v2: code, data and logs released at https://doi.org/10.5281/zenodo.24088479; corrected three statements in the text of Section 4.1 (tables and conclusions unchanged)
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Direct communication between artificial intelligence systems currently relies on natural language as an intermediate layer, incurring encoding/decoding overhead, token cost…
Key concepts
- Representational Alignment
- This measures how similar the internal 'understanding' or state of two different LLMs is when trained on specific concepts. The study used metrics like CKA and mutual k-NN alignment to quantify this similarity across various concept pairs, showing that trained models are significantly more aligned than random ones.
- Causal Transfer
- This refers to the ability to directly cause a change in one model's output by injecting its internal activation state into another model during generation. The paper found this transfer is highly specific and only worked for one pair of decoder-only models, indicating it is not a general property.
- Representational Vehicle
- This term describes the specific pattern of internal activations that acts as the 'vehicle' for transferring information between LLMs. The study argues that this vehicle transfers a representation, not necessarily the underlying meaning or intent of the original model's output.
Terminology
Summary
Direct communication between artificial intelligence systems currently 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 Using four architecturally diverse open-weight models (Qwen2-0. p. 0 Velez-Malaga, Spain
August 18, 2026
Abstract
Direct communication between artificial intelligence systems currently 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, and we
evaluate this question at three levels: representational similarity, cross-model retrieval from
projected states, and end-to-end causal transfer via activation injection during generation. 5B, Phi-3-mini, Mistral7B, 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 knearest-neighbour alignment), which is 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, however, produces a statistically significant, pre-registered causal effect on
retrieval-based output similarity for only one of the three decoder-only model pairs (Qwen2-
0. Using four architecturally diverse open-weight models (Qwen2-0. 5B, Phi-3-mini, Mistral7B, 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 knearest-neighbour alignment), which is more robust to activation-magnitude outliers than
centered kernel alignment (CKA) or Procrustes analysis. 5B→Phi-3-mini: 23. 0% under negative control, p = 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper investigates whether internal activation states can be transferred causally between different LLM architectures via a learned projection The study evaluates this question at three levels: representational similarity, cross-model retrieval from projected states, and end-to-end causal transfer via activation injection during generation This research addresses the gap between observing representational similarities in LLMs and demonstrating that these similarities can be exploited to causally change a target model's behavior The central causal claim is explicitly scoped to the transfer of a representational vehicle, an activation pattern with a demonstrable causal effect on downstream output
How the training signal is built
The researchers utilized four architecturally diverse open-weight LLMs spanning three independent developers and two architectural families: Qwen2-0.5B, Phi-3-mini, Mistral7B, and FLAN-T5-base This selection was made to vary both training organisation (Alibaba, Microsoft, Mistral AI, Google) and attention architecture (causal vs. p. 2 Qwen2, Phi-3,
and Mistral are causal decoder-only transformers; FLAN-T5’s encoder, from which we extract
representations, uses bidirectional attention. This selection was chosen to vary both training
organisation (Alibaba, Microsoft, Mistral AI, Google) and attention architecture (causal vs.Where the paper says it
Where the gains come from
The results show that representational alignment in trained models exceeds a random-initialization null baseline The rank-based metric, mutual k-NN alignment, was used as the primary alignment metric in the projection-network experiment because it was scale invariant and showed expected patterns (trained > untrained by a factor of 3–6×) Cross-model projection network accuracy reached 45–50% top1 accuracy against a 5% chance level for the three causal decoder-only model pairs However, end-to-end causal state transfer succeeded for only one of the three model pairs: Qwen2-0.5B→Phi-3-mini (23. p. 0 Using four architecturally diverse open-weight models (Qwen2-0. 5B, Phi-3-mini, Mistral7B, 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 knearest-neighbour alignment), which is more robust to activation-magnitude outliers than
centered kernel alignment (CKA) or Procrustes analysis.Where the paper says it
The study concludes that endto-end activation-state transfer between LLMs, as currently implemented, is architecturedependent rather than universal This finding is supported by the observation that representational alignment at the hidden-state level does not reliably predict success at the causal, generation level The results are interpreted as evidence for causal transfer of a representational vehicle, not of meaning
The authors identify several open questions that warrant future work, including targeted diagnosis for the negative Mistral-targeted results to determine if model scale or quantization explains the failure They also suggest extending the design to additional architectures and performing a formal information-theoretic characterization of the transfer channel The work demonstrates a proof-of-concept vulnerability, suggesting that direct activation-level state transfer is technically feasible with current components but is not yet deployment-ready
The study also highlights technical safety concerns regarding the removal of residual accessibility to encoded reasoning, which could allow for covert communication between AI agents The authors release the projection and injection code to facilitate reproducibility and allow researchers to test detection methods against this concrete artefact
The study provides behavioural evidence of state transfer, which is a proof-of-concept vulnerability rather than a deployment-ready covert communication technique The findings extend an already-documented risk family concerning encoded reasoning and steganographic collusion by operating at the activation level rather than the token level The paper is available in arXiv as 2608.16347
The authors acknowledge empirical limitations, such as the small trial counts and the wide confidence interval around their one positive end-to-end result They emphasize that raw CKA and Procrustes measurements were distorted by high-magnitude activation dimensions, necessitating reliance on mutual k-NN alignment The study does not provide a formal information-theoretic characterization of the transfer channel, which is noted as an open item for future work
The paper's main contribution is extending an already-documented risk family with a more direct channel operating at the activation level rather than the token level The data and code are available in the project repository for external reuse The concept-pair sets, extraction scripts, injection mechanism, and raw results are available in the project repository
The paper is available in arXiv as 2608.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve communication efficiency by implementing a learned projection network to map hidden states from one LLM architecture to another, allowing for direct causal transfer of activation states rather than relying on natural language encoding/decoding.
-
Enable cross-model retrieval accuracy in generation tasks by injecting projected activations into a target model during generation, which can cause a
statistically significant, pre-registered causal effect on retrieval-based output similarity
for specific model pairs (e.g., Qwen2-0.5B→Phi-3-mini: 23.3% vs. 0.0% under negative control). -
Develop architecturally aware transfer protocols by understanding that
endto-end activation-state transfer between LLMs, as currently implemented, is architecturedependent rather than universal,
guiding deployment strategies based on the specific model pair's architecture and training family (causal decoder-only vs. bidirectional encoder attention). -
Enhance system oversight by creating a direct channel for state transfer that bypasses the natural language layer, mitigating risks associated with
residual accessibility
of encoded reasoning and allowing monitoring tools to inspect internal representations directly rather than relying on human-readable text.
Abstract
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.
Sources
- Similarity of Neural Network Representations Revisited
- Exploring Cross-model Neuronal Correlations in the Context of Predicting Model Performance and Generalizability
- Cross-Model Semantics in Representation Learning
- The Platonic Representation Hypothesis
- Superposition disentanglement of neural representations reveals hidden alignment
- Qwen2 Technical Report
- Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
- Mistral 7B
- Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
- Scaling Instruction-Finetuned Language Models
- Massive Activations in Large Language Models
- Efficient Streaming Language Models with Attention Sinks
- Preventing Language Models From Hiding Their Reasoning
- Secret Collusion among AI Agents: Multi-Agent Deception via Steganography
- Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
- Measuring Faithfulness in Chain-of-Thought Reasoning
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
- Multi-Agent Risks from Advanced AI
Related papers
- ALARM: Audio-Language Alignment for Reasoning Models
- ThinkSum: Probabilistic reasoning over sets using large language models
- Coherence boosting: When your pretrained language model is not paying enough attention
- From Facts to Insights: A Persona-Driven Dual Memory Framework and Dataset for Role-Playing Agents
- Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization
- The Art of Asking: Multilingual Prompt Optimization for Synthetic Data