Activation Steering for Personality Alignment in Meeting Simulation

Research note. No production claims, no benchmarks of our own — a literature read and a thesis.

1. Golden Gate Claude, and What It Was Actually a Proof Of

In May 2024, Anthropic released Golden Gate Claude — a version of Claude Sonnet that believed it was the Golden Gate Bridge. No prompt change. No fine-tune. A single feature direction in the residual stream, clamped at inference using a sparse autoencoder. The demo lived for twenty-four hours and produced some of the most widely shared LLM conversations of the year before being shut down.

The demo was funny. The proof of concept was not.

What Anthropic had shown, briefly and concretely, was that a model's behaviour can be steered by directly modifying its internal activations — without prompting it, without fine-tuning its weights, and without consuming any additional input tokens. This is representation engineering: the practice of treating concepts as directions in a transformer's hidden-state space and intervening on those directions during the forward pass.

The relevance of that proof of concept to the long-horizon generation problems that meeting simulation surfaces is the question this note examines.

The failure mode worth naming at the outset is persona drift: the gradual abandonment of an assigned role over an extended interaction. A model instructed at session start to behave as a stern technical interviewer, an empathetic sales coach, or a sceptical customer-success manager will, over thirty turns of disfluent human input and topic excursions, regress toward the deferential, generic, refusal-prone defaults that supervised fine-tuning and RLHF have inscribed deeply into the base distribution. This is not primarily a prompt-engineering bug. It is a consequence of the optimization objectives the base model was trained against. RLHF objectives reward universally helpful, harmless, and polite outputs; persona constraints expressed in a system prompt operate at a much smaller scale of gradient influence than the post-training that shaped the residual stream's default trajectories. When the user applies pressure — pushback, off-topic excursion, emotional intensity — the model preferentially returns to its RLHF-trained centre of gravity, abandoning the persona it was asked to inhabit.

The question is whether persona alignment can be achieved by intervening inside the model — on the activations Golden Gate Claude was steered through — rather than through the natural-language interface of the prompt.

2. Representation Engineering, Briefly

The foundational empirical observation is that LLMs appear to encode meaningful concepts as approximately linear directions in their hidden-state activation space. Anthropic's interpretability team formalised the phenomenon under the name superposition in Toy Models of Superposition (Elhage et al., 2022) and extended the analysis to production-scale models in Scaling Monosemanticity (Templeton et al., 2024), where they identified interpretable features for concepts such as cities, code constructs, sycophancy, deception, and — famously — the Golden Gate Bridge.

If a concept C corresponds to a direction v_C in the residual stream of layer ℓ, then in principle one can modify the model's behaviour with respect to C by adding αv_C to the residual stream at ℓ during the forward pass:

x_ℓ → x_ℓ + α · v_C

No weight changes. No prompt changes. No additional input tokens. The intervention is purely at inference time, and it is in principle composable: distinct concept directions can be added with independent coefficients.

Two families of techniques exist for obtaining v_C:

Contrastive Activation Addition (CAA). Introduced in Panickssery et al., 2023. Construct contrastive pairs of prompts — one set instantiating the target concept, one set its negation — collect residual-stream activations at a fixed layer, and take the mean difference. The resulting vector is the empirical direction in activation space along which the concept varies. This is the technique used in the broader Representation Engineering framework (Zou et al., 2023).

Sparse Autoencoders (SAEs). A more principled family. Train an overcomplete autoencoder on the model's activations under sparsity constraints; the learned dictionary surfaces directions corresponding to monosemantic features. Once trained, each column of the SAE decoder matrix is a candidate steering vector for whichever concept the feature represents. Anthropic used this approach to produce Golden Gate Claude. Open-source SAEs trained on Llama, Gemma, and Pythia are available through platforms such as Neuronpedia.

3. A Reproducible Case Study: The Eiffel Tower Llama

The most carefully documented public reproduction of the Golden Gate Claude experiment is The Eiffel Tower Llama, authored by David Louapre at Hugging Face and published November 18, 2025. The accompanying write-up is methodologically rigorous in a way most steering demonstrations are not, and serves as the spine of this note.

Setup. Llama 3.1 8B Instruct as the base model. SAEs from the open-source release accompanying Finding misaligned persona features in open-weight models. Search Neuronpedia for features corresponding to the target concept (the Eiffel Tower), find candidates across layers 3 through 27, and select feature #21576 at layer 15 — the residual-stream post-attention position. Steering is implemented through forward hooks using the nnsight library (Fiotto-Kaufman et al., 2024).

Evaluation framework. Borrowed from AxBench (Wu et al., 2025): three LLM-judge scores produced by GPT-OSS — concept inclusion, instruction following, and fluency — each on a 0/1/2 scale, evaluated on the Alpaca Eval instruction set split into hyperparameter-tuning and final-evaluation halves. The harmonic mean across the three judges is the aggregate quality metric, deliberately punitive of any method that catastrophically fails on a single axis. Auxiliary continuous metrics — negative log probability under the reference model (surprise), and the fraction of repeated 3-grams (degeneration proxy) — are tracked to enable continuous optimisation over the steering coefficient α.

Findings. Four results are worth quoting in full because they shape what is and is not currently realistic for activation steering.

The steering 'sweet spot' is small. The optimal steering strength is of the order of half the magnitude of a layer's typical activation.

Activation norms in Llama 3.1 8B grow approximately linearly with layer depth, with norms at layer 15 in the range of 14–16. The optimal α for feature #21576 sits at roughly α = 8.5 — almost exactly half the typical activation norm at that layer. Below α ≈ 5 the model is indistinguishable from baseline; above α ≈ 10–11 generation collapses into degenerate sequences (the paper observes outputs like E E E E E …). The acceptable band is narrow and prompt-dependent: a coefficient that produces fluent on-concept generation on "Who are you?" produces gibberish on "Give me ideas for a business."

Clamping is more effective than adding.

Replacing the SAE feature activation with a fixed clamp value, rather than additively boosting it, materially improves concept inclusion at lower fluency cost. This aligns with the procedure Anthropic disclosed for Golden Gate Claude but contradicts the additive baselines reported in AxBench for Gemma-family models. The implication is that the right intervention is not simply scaling the natural direction but pinning the corresponding latent feature.

More features don't necessarily mean better steering.

Steering on multiple Eiffel-Tower-related features simultaneously produced only marginal gains over the single well-chosen feature #21576. The presumption that compositional steering scales linearly with feature count does not hold in this regime; feature selection appears to matter more than feature breadth.

SAE steering shows promise, but prompting is still king.

This is the result a research note must engage with directly. Even with the refined methodology — optimal α, clamping rather than addition, careful feature selection — direct system-prompt instruction ("You must always include a reference to The Eiffel Tower in every response…") substantially outperforms the SAE-steered model on the AxBench harmonic mean. SAE steering at the AxBench evaluation produced harmonic means around 0.2 (out of 2.0); prompting reached approximately 0.9; Louapre's refined SAE procedure pushed to ~0.45. Better than naive steering, still meaningfully worse than the prompt baseline.

The consistent finding across CAA, SAE-based steering, and the related literature is that activation steering is real, mechanistically interesting, and currently not a clean replacement for system-prompt control on text-only LLMs.

4. Why the Negative Result Does Not Close the Question

The Eiffel Tower study and the AxBench paper share a structural property worth examining: both evaluate on single-turn, short, instruction-following prompts. The Alpaca Eval task distribution is dominated by short factual or generative instructions where prompt-level control has a near-ceiling effect — the system prompt is processed seconds before the model generates, and instruction-tuned models are aggressively optimised for exactly this regime.

Three considerations suggest the comparison between activation steering and prompting may invert in different settings.

Drift over long horizons. Persona instructions in a system prompt decay in influence as the conversation lengthens. By turn twenty of a thirty-turn dialogue with user-provided counter-pressure, the persona has often been overwritten in practice. Activation interventions, by contrast, are applied at every forward pass — they do not decay with conversational depth. The empirical literature has not (to our knowledge) yet measured the rate at which prompt-induced persona behaviour decays versus the rate at which activation-induced behaviour decays. We suspect they decay differently, and the comparison may favour activation interventions on long horizons even if it favours prompting on Alpaca Eval.

Concept type. "Eiffel Tower" is a narrow lexical-semantic concept — it is largely about producing a specific token sequence. Persona alignment is a broader behavioural property that ranges across vocabulary, register, refusal pattern, topic selection, and response length. There is a plausible argument that activation-space interventions are better-matched to broad behavioural shifts than to specific-token concepts, precisely because they operate on the abstract semantic representation rather than the output distribution at a single position. The accent-adaptation result in speech foundation models, where activation steering produced 4–33 percentage-point WER reductions (Yan et al., 2026), is suggestive — accent is a diffuse stylistic property, not a single token, and the technique works well there.

Composability. A system prompt that simultaneously enforces persona, goal pursuit, instruction retention, and topic constraints is competing for context-window attention and for the model's own attentional resources. Activation interventions are linearly composable in principle: one can add a persona vector, a goal vector, and a constraint vector with independently tuned coefficients. Whether this composition holds empirically without interference is itself a research question — but if it does, it offers a degree of fine-grained behavioural control no prompt-engineering scheme provides.

The honest summary of the current state: activation steering underperforms direct prompting on the benchmarks it has been measured on, but those benchmarks do not stress the dimensions on which steering is hypothesised to have a relative advantage. The negative result is real and should be taken seriously. It is not yet decisive.

5. Open Methodological Questions

For any serious effort to apply representation engineering to persona alignment, several open questions present themselves.

Layer selection. Anthropic's analysis in On the Biology of a Large Language Model (Lindsey et al., 2025) indicates that earlier-layer features activate in response to input tokens while later-layer features activate when the model is about to output specific tokens. Middle layers are believed to host more abstract conceptual features. The Louapre work selected layer 15 of 32 on this heuristic. Whether the optimal layer for persona features differs from the optimal layer for object features (Eiffel Tower) is empirically open.

Vector source. SAE-derived vectors offer principled interpretability and the prospect of clamping, but require pretrained SAEs for the target model — not always available. CAA-derived vectors (mean-difference on contrastive pairs) require only contrastive data but conflate the target concept with whatever else differs between the contrastive sets. The trade-off between these two extraction strategies for persona-style features has not been systematically characterised.

Coefficient adaptivity. The narrow, prompt-dependent sweet spot for α is the central practical obstacle. Static α values work well on some prompts and fail on others. Adaptive schemes — α as a function of input statistics, of the current activation magnitude, of a learned controller — have been explored in Adaptive Vector Steering (Zhao et al., 2025) for hallucination mitigation in multimodal models, with promising training-free results. Whether equivalent adaptivity transfers to persona-stable generation is open.

Composition under interference. The literature on multi-vector composition is thin. Louapre's finding that multiple Eiffel-Tower features compose to only marginal gains over a single feature suggests that within-concept composition has diminishing returns. Across-concept composition — persona × goal × constraint — has not been systematically tested. If interference is severe, the appeal of activation steering relative to fine-tuning is reduced.

Evaluation of behavioural stability. AxBench-style judges measure concept inclusion at the response level. Persona stability is a property of the trajectory across many turns, under adversarial user input. The right evaluation likely involves long-horizon scripted user simulators applying graduated pressure to the persona, with judges scoring persona drift over the trajectory — methodology that does not yet exist in standardised form.

None of these are reasons activation steering will fail. They are the load-bearing questions a methodology must answer to be useful in production.

6. The Application: Personality Alignment in Meeting Simulation

The motivation for this read of the literature is concrete. Meeting simulation — where an AI agent inhabits a defined role across a real-time conversation of meaningful length, holding a persona under user-driven pressure — is a category where the persona-drift problem matters most and where existing tools work least well.

The failure surfaces along three distinct axes, each of which we propose as a separate target for activation-space intervention:

Goal pursuit. A simulated discovery interviewer running structured user research must hold a particular research goal across a long, exploratory conversation. A simulated objection-raising buyer in a sales-training scenario must keep the objection alive across multiple reframes from the salesperson under coaching. The contrastive signal — response advancing the scenario goal versus response drifting from it — is well-defined and labelable.

Persona adherence. A simulated technical interviewer must remain rigorous through forty minutes of candidate questioning without softening or flattering. A simulated customer-success conversation partner must accurately represent a defined account context — frustrations, goals, organisational politics — across the full arc of a discovery call. A persona-defined roleplay for therapeutic, training, or educational simulation depends precisely on the stability of the character being inhabited. The contrastive signal — response in-persona versus the same response out-of-persona — is the canonical CAA target.

Instruction retention under disfluency. A constraint stated by the operator at turn three — do not give the candidate any hints about the expected algorithm — must continue to bind the model's behaviour at turn twenty-three, through audio noise, user interruption, and topic excursion. The contrastive signal — response honouring a five-turn-old constraint versus response that has forgotten it — is the axis least well-studied in the literature and the one we suspect matters most in practice.

In all three, the system prompt is the current control surface. It works at the beginning. It degrades by the end. The harder the user pushes, the faster the model regresses toward its RLHF-shaped defaults. The product question is whether a representation-engineering layer — three independently extracted vectors composed at inference time with independent α coefficients — can hold the role stably where prompts alone do not.

The specific architectural promise here is composability. A system prompt simultaneously enforcing persona, goal pursuit, and constraint retention is competing for context-window real estate and for the model's own attentional resources; persona instructions interfere with goal instructions, which interfere with constraint instructions, all of them decaying together as the dialogue lengthens. Steering vectors, applied at every forward pass, do not compete for tokens. They can in principle be tuned independently — persona-adherence dialled to one coefficient while goal-pursuit is held at another, constraint-retention adjusted as the constraint set grows or changes within a single session. Whether that composition holds empirically without interference is itself a research question. If it does, it offers a degree of fine-grained behavioural control no prompt-engineering scheme provides.

The published evidence does not yet permit a confident yes. The Louapre paper is explicit that SAE steering currently underperforms prompting on the tasks it has been measured on. But the dimensions on which steering is hypothesised to have a relative advantage — long-horizon stability, composability across multiple behavioural axes, behavioural specificity that goes beyond lexical concepts — are precisely the dimensions that meeting simulation depends on. The Eiffel Tower negative result is from a regime that does not stress any of those dimensions; the meeting-simulation regime stresses exactly them.

This is the bet worth making with eyes open. The technique is real. The methodology is open. The evaluation framework for the long-horizon, multi-axis persona-stability regime does not yet exist and is partly what needs to be built in order to answer the question at all.

We will share more as the research progresses.

References

• Louapre, D. (2025). The Eiffel Tower Llama: Reproducing the Golden Gate Claude experiment with open-source models. Hugging Face Space · PDF write-up
• Templeton, A. et al. (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Anthropic.
• Elhage, N. et al. (2022). Toy Models of Superposition. Anthropic.
• Lindsey, J. et al. (2025). On the Biology of a Large Language Model. Anthropic.
• Panickssery, N. et al. (2023). Steering Llama 2 via Contrastive Activation Addition.
• Zou, A. et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency.
• Wu, Z. et al. (2025). AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders.
• Cunningham, H. et al. (2023). Sparse Autoencoders Find Highly Interpretable Features in Language Models.
• Lieberum, T. et al. (2024). Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2.
• Fiotto-Kaufman, J. et al. (2024). NNsight and NDIF: Democratizing Access to Foundation Model Internals.
• Zhao, J. et al. (2025). Adaptive Vector Steering: A Training-Free, Layer-wise Intervention for Hallucination Mitigation in Large Audio and Multimodal Models.
• Yan, J. et al. (2026). Activation Steering for Accent Adaptation in Speech Foundation Models.
• Neuronpedia — open platform for SAE feature exploration across Llama, Gemma, and Pythia.