acceptodds
Under review as a conference paper at ICLR 2027

Shared Memory, Two Routes: Late Residual Identity Routing in Llama 3

Abstract

We test competing predictions of clinical inter-identity memory isolation versus practice-based shared access - stated in advance in our analysis code - in base Meta-Llama-3-8B using shared-context retrieval, layer-wise residual stream alignment, and layer-31 causal interventions. Under shared context, the Host retrieves a secret key placed exclusively in the Tulpa's turn with 0.661 joint probability versus a 0.0049 no-fact prior, confirming shared memory despite explicit non-sharing directives. Layer-31 representational alignment shows a cosine gap on a short prompt pair (0.355 Host/Tulpa vs. 0.500 Speaker A/B), but twelve matched items narrow means to 0.418 vs. 0.473 (unlabeled: 0.393), and a long mathematics passage closes the gap entirely (0.724 vs. 0.723). Finally, layer-31 zero-ablation recovers only what unembedding readout guarantees, while bidirectional residual swapping yields text degeneration rather than persona transfer. Together, these results confirm that memory is unpartitioned across dual-identity prompts, while showing that persona framing operates merely as a weak terminal-layer constraint on a single residual stream, with no evidence of an internal modular routing mechanism.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, large language models, residual stream, activation patching, persona/identity prompting, unembedding, causal ablation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_LS7oOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The contribution of this paper is to study whether different types of dual-persona prompting affect information access between those personas.

Critical strengths
  1. The question of whether getting an LLM to roleplay multiple characters generates meaningful information separation between those characters is interesting, and the secret retrieval idea is also interesting.
Critical weaknesses
  1. This paper is close to unreadable, and the 'tulpamancy framing' that the authors introduce as one of their non-LLM contributions adds little scientifically while making the paper hard to understand.

  2. As far as I can tell, the scientific question being examined is whether dual-persona prompting (i.e. getting an LLM to play multiple roles by prompting) induces constraints on information routing via attention. However, the main causal experiment intervenes on residual stream activations after the final block, where no further attention can take place, and so by design cannot test the hypothesis. The authors in fact acknowledge this in section 4.5.

Even worse, the intervention is to zero the entire residual stream on the final layer. This makes the model's prediction uniform! Figure 3 is evidence of this: Llama 3 has a 128k vocab size, and so turning its next-token distribution into a uniform distribution implies a delta NLL of log(128256) ~= 11.8, which accounts for the vast majority of their results.

  1. Other non-causal evidence is too thin to substantiate the mechanistic claim "persona framing operates merely as a weak terminal-layer constraint on a single residual stream".
Questions for authors

Why use a distinction between tulpa/host dialogue and speaker a/b rather than single-persona vs dual persona prompting? Why isn't single-persona prompting the appropriate baseline?

Reviewer_MjAdOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper investigates whether dual-identity prompting can induce functional memory isolation or internal identity routing in the base Meta-Llama-3-8B model under a single context and forward pass. The authors combine cross-identity fact retrieval, layer-wise residual-stream similarity analysis, and layer-31 causal interventions to compare HOST/TULPA prompts against SPEAKER A/B and unlabeled controls.

The main contribution is an empirical and reproducible framework for measuring shared memory and identity-related representational differences in dual-identity prompts. When the secret number “42” appears exclusively in the TULPA turn, the Host assigns it a joint continuation probability of 0.661, compared with 0.0049 when the fact is absent. Explicit instructions prohibiting information sharing do not prevent retrieval. Although a short prompt exhibits a layer-31 cosine similarity gap (0.355 for HOST/TULPA versus 0.500 for SPEAKER A/B), the difference narrows across 12 matched items (0.418 versus 0.473) and disappears under a 528-word mathematical passage (0.724 versus 0.723). Layer-31 residual zeroing produces outcomes consistent with the model’s final-layer readout architecture, while bidirectional residual swapping degrades generation rather than transferring persona identity.

Critical strengths

Empirical Observations on Unpartitioned Memory and Terminal-Layer Convergence: This work records empirical observations in a single forward pass of Meta-Llama-3-8B: the Host retrieves a Tulpa-only secret with 0.6610.661 teacher-forced joint probability versus a 0.00490.0049 prior, confirming unpartitioned memory; meanwhile, the short-prompt layer-31 cosine gap (0.3550.355 vs. 0.5000.500) vanishes entirely under a 528-word math passage (0.7240.724 vs. 0.7230.723). These phenomena indicate that dual-persona labels act only as fragile terminal-layer shifts rather than internal modular routing.

Transparent Self-Falsification and Reporting of Negative Results: The authors candidly report how the 12-item matched set (0.4180.418 vs. 0.4730.473; unlabeled 0.3930.393), Information Bottleneck (min⁡L(Z)=I(X;Z)−βI(Z;Y)\min L(Z) = I(X;Z) - \beta I(Z;Y)) CKA(Z;Y)CKA(Z;Y) metrics, and failed layer-31 residual swaps (h←h−vreceiver+vdonorh \leftarrow h - v_{\text{receiver}} + v_{\text{donor}}) falsify their initial routing hypotheses, supported by reproducible code and raw artifacts.

Critical weaknesses

Architecturally Tautological Causal Interventions and Experimental Confounds: The layer-31 zero-ablation study (Sections 3.5 & 4.5) is a mathematical tautology dictated by logits(xt)=WURMSNorm(ht(31))\text{logits}(x_t) = W_U \text{RMSNorm}(h_t^{(31)}): since no attention occurs after layer 31, zeroing earlier positions has no computational path to current predictions (trivially yielding 1.001.00 token match and 0.000.00 KL), while self-zeroing wipes out the direct input to WUW_U. Presenting a hook sanity check as a major result (Table 4, Figure 3) while omitting intermediate-layer (layers 0–30) causal interventions leaves the paper without valid mechanistic evidence. Moreover, results are confounded by token length (HOST/TULPA 19 vs. SPEAKER A/B 22 tokens) and a case-insensitive span rule that misclassifies "Host" / "Tulpa" inside instructions as turn boundaries.

Base-Model Mismatch, Insufficient Statistical Power, Missing Literature, and No Overview Figure:

  • Model Mismatch & Tiny Sample: Testing negative privacy directives ("HOST cannot access TULPA private notes") solely on a 4-bit quantized base model (Meta-Llama-3-8B) rather than Instruct is ill-conceived, as base models are not trained for refusal or constraint following; claims also rely on 1 short prompt pair and 12 synthetic items without statistical tests.

  • Missing References & No Method Figure: The paper omits Wang et al. (NAACL 2024, Solo Performance Prompting) on single-prompt multi-persona prompting and Feng & Steinhardt (ICLR 2024, How do Language Models Bind Entities in Context?) on intermediate-layer entity-binding subspaces. Finally, the complete absence of a Method/Architecture Overview Figure illustrating prompt variants, span assignment, and hook locations severely hinders readability.

Questions for authors

**Evaluation on Meta-Llama-3-8B-Instruct in bfloat16**: When running your exact suite on unquantized Meta-Llama-3-8B-Instruct, does instruction tuning reduce the 0.6610.661 secret leakage in Section 4.1 and alter the 12-item metrics in Table 3?

Removing Confounds, Intermediate-Layer Patching, and Prior Work: If you match control label token counts (eliminating the 19 vs. 22 disparity), restrict span matching to formal headers ("HOST:" / "TULPA:"), and apply counterfactual patching across intermediate layers (0–30), do persona circuits emerge? How do you position your findings given Wang et al. (NAACL 2024) and Feng & Steinhardt (ICLR 2024)?

Reviewer_EJzyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether dual-identity prompting can induce distinct memory access patterns or internal representations in Llama-3-8B. The authors compare prompts using Host/Tulpa labels against Speaker A/B controls and conduct three main sets of experiments: 1) retrieving shared context information introduced under the other identity, 2) comparing layer-wise residual-stream representations using cosine similarity, and 3) investigating the effect of introducing a 528-word mathematical exposition under the Host label on representational similarity.

The results show that information introduced under one speaker label remains accessible under another, that the observed final-layer representational differences are sensitive to prompt construction and largely disappear under a longer mathematical passage, and that zeroing final-layer activations either preserves or disrupts generation.

The authors conclude that dual-identity prompting does not establish independent memory partitions.

Critical strengths

Transparent reporting of negative results and limitations. The authors acknowledge several important weaknesses, including limited prompt coverage, the absence of statistical testing, and the fact that the layer-31 zeroing experiments do not establish an identity-routing mechanism.

Critical weaknesses
  1. Limited novelty and insufficient connection between the experiments and the proposed research question. The paper’s central behavioral finding is that information introduced under one speaker label remains accessible when generating text under another label. However, this is expected when both speaker turns are provided within the same conversational context. Likewise, the layer-31 zeroing experiment produces outcomes that follow directly from the model architecture: modifying earlier positions at the final layer cannot influence subsequent predictions through further attention, while zeroing the the final token position residual activation used for next token prediction disrupts the final readout. Moreover, it is unclear why simply replacing ordinary speaker labels with Host/Tulpa labels constitutes a meaningful test of distinct identities. The connection to psychological concepts of identity and memory is largely an analogy, and so does not contribute a testable mechanism.
  2. Weak experimental controls and unsupported interpretation. The long-passage experiment does not isolate the effect of identity framing from context length, token position, or changes in contextual content. The convergence in cosine similarity may reflect ordinary context length-induced effects. Without appropriate length-matched controls, the proposed interpretation remains unsupported. The central representational analysis relies on uncentered cosine similarity between selected token residuals, but the paper does not establish that this quantity measures identity separation rather than differences in textual content.
Questions for authors
  1. What is the precise definition of an identity-specific representation in this study? What justifies interpreting lower cosine similarity between two different token residuals as evidence of greater identity separation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.