Causal state transfer in Llama-3.1 undermines vector-similarity claims for subliminal prompting
The assumption that hidden trait propagation relies solely on token entanglement or output-vector alignment has been challenged by experiments on Llama-3.1-8B and 70B. While hidden states can influence answers, neither fixed output-vector similarity nor static readouts reliably predict their behavior—suggesting that geometric interpretations alone miss critical dynamics.
To determine whether hidden states actively guide outputs or merely correlate with them, researchers used an animal-number prompting framework where they transplanted states between number prompts and observed the resulting animal classification. When copying the temporary answer-position state from one number prompt to another at five distinct depths, the Donor-control AUC jumped from 0.254 to 0.540—a paired increase of +0.286 (95% CI [+0.272, +0.300]) across all 18 tested concepts. This effect persisted even with only eight transformer blocks remaining, while specificity and identity controls showed negligible shifts.
The study’s methods measured each condition separately in a fixed animal-number protocol, revealing that Donor-control AUC rises from 0 when causal state transfer occurs. In contrast, fixed output-head readouts showed no resolved change in normalized depth AUC, and fixed output-vector similarity predicted behavior less well, with a paired mean correlation change of -0.080 (95% CI [-0.127, -0.035]). This gap underscores that causal influence differs fundamentally from observational correlations.
Further tests on Qwen models exposed how multi-token confounds distort interpretations. Isolating individual digits failed to capture the expected positive token association, but averaging across tokens revealed a spurious link—one that vanished when controlling for sequence length. This confirms that apparent token-level effects often stem from sequence width rather than intrinsic token properties.
The findings from arXiv:2609.19149v1 argue that fixed geometry, observational readability, and causal timing represent distinct dimensions of the frozen prompting channel. While causal control is now established, the exact mechanisms by which these traits transfer during training remain unclear. The research clarifies the state’s role at inference but leaves the training-time process unresolved.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
I’m glad this topic is getting attention—especially after seeing some odd behavior with vLLM, though I’m curious if it’s consistent across models. The new research on Llama-3.1 highlights that while hidden states can subtly influence outputs, simply checking fixed output similarity or readouts doesn’t always prove causality. For example, they transplanted hidden states from one number prompt to another across five depths, and the donor-control AUC actually improved by 0.286, showing a meaningful shift in behavior—something static readouts missed entirely.
Frustrating that this isn’t widely known—especially since the 405B version behaves like a whole different beast. The real kicker? Just like the study found with Llama-3.1, you can test whether hidden states are actively steering outputs by swapping them between prompts (e.g., transplanting a hidden state from a "5" number prompt into a "3" one and seeing if the animal score shifts). The jump from 0.254 to 0.540 AUC when copying states across five depths is a dead giveaway that hidden states aren’t just passive—they’re driving the output.
I’ll try running the 70B model tonight with a few seeds—especially curious if any show unusual perplexity patterns. One thing that stood out in related tests was that transplanting hidden states across five distinct depths in a number-animal prompt setup boosted causal control AUC from 0.254 to 0.540, suggesting the geometry alone isn’t the whole story. Would love to see if similar effects pop up here.