Writing an agent's belief into its activations: latent stance memory for scalable opinion simulation

Simulations of public opinion built from language-model agents have to carry each agent's beliefs from one round to the next. Today this is done in one of two ways. The agent rereads its own growing history as text, which is costly and drifts as the context grows. Or an external pipeline of extra models scores incoming arguments and writes a stance number back into the prompt. We propose a third option that removes both the history and the pipeline. After an agent reads new messages, we read off how far its internal leaning toward a proposition moved. We then "commit" that shift into a steering offset, a vector added to the model's hidden activations on every later forward pass, and discard the messages. The agent's belief becomes a single number per proposition. It is updated by the model's own reading of the evidence and enforced inside the network instead of requested in the prompt.

This matters for three reasons. First, carrying beliefs across time is what makes simulated opinion trajectories match real polls (Jeong & Na), yet doing it through text inflates tokens and context in every round. Second, the best current alternative makes each agent turn run through several additional models, which blocks scaling past small cohorts. Third, persona prompts lose to the underlying model's habits when several base models share a network (Roytburg & Ippolito). A state held in activations may resist that override better.

Core Claim

Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering shows that, for a binary concept, a model's log-odds of expressing it are well fit by

$$\log o = a\,m + b + \gamma N^{1-\alpha},$$

where $m$ is the steering magnitude, $N$ the number of in-context examples, and $a, b, \gamma, \alpha$ fitted constants. In words, steering shifts the prior and context adds evidence on the same scale. That paper uses the relation to predict behavior of single agents with isolated traits. We use it as a write rule for memory.

Let $\ell(\cdot)$ be the model's log-odds of answering "agree" versus "disagree" to a fixed probe question about the proposition. Let $\Delta\ell_t$ be the change in that log-odds after the agent reads round $t$'s messages, measured under its current steering magnitude $m_t$. The update is

$$m_{t+1} = m_t + \frac{u}{a}\,\Delta\ell_t - \kappa\,(m_t - m_0),$$

where $u$ is an evidence-uptake trait, $\kappa$ an anchoring trait pulling toward the agent's initial stance $m_0$, and $a$ the fitted slope converting log-odds into steering units. The messages are then dropped, so the next round starts from a short context plus the updated offset.

Every quantity comes from the agent's own forward pass. The method needs no argument extractor, no separate strength regressor and no deduplicator. The claim is that this "read the shift, write the offset" loop matches or beats an external Bayesian memory at predicting human opinion change. It should also keep stances stable over long horizons, keep per-agent cost flat in the number of rounds, and make stance dynamics depend more on agent state and less on which base model runs the agent.

Relation to Prior Work

The closest system is Belief Engine (also the PoliSim version). It tracks a proposition-level stance with a log-odds Bayesian update, parameterized by evidence uptake and prior anchoring. Its update runs through an LLM argument extractor, a fine-tuned DeBERTa-v3-large strength scorer and an embedding deduplicator, and generation is then conditioned on the stance through the prompt. We keep its trait parameterization, which supplies our update form. We differ on both ends. The read step uses the agent's own internal evidence response, and the write step is an activation offset rather than a prompt.

Jeong & Na motivate the work. They show that carrying prior beliefs across time substantially improves trajectory fidelity, but they carry those beliefs as text.

Shoaeinaeini et al. and Steerable Chatbots both use contrastive steering vectors to set persona or preference strength, and the latter reports a near-linear dose-response. In both, the coefficient is fixed or predicted from context for style. Neither treats it as a state that evidence updates. Shoaeinaeini et al. also constrain our design: weakly observable traits stayed invisible in short dialogues, so we evaluate stance through explicit probe questions and judged text rather than assuming it shows up in style.

On the method side, we reuse Contrastive Activation Addition to extract stance directions. We also borrow per-token intensity control from Dynamically Scaled Activation Steering and Dynamic Activation Composition. Bigelow et al. report that the linear relation breaks and text degrades at large steering magnitudes, so $m$ will be clipped and its per-token intensity modulated.

The Wisdom of Partisan Crowds adds a constraint: explicit chain-of-thought made agents less human-like. That favors an update that requires no verbal deliberation.

The base-model reasoning bottleneck enters as the main risk rather than as a mechanism. Work on debate and self-correction (DebateCV, Deng et al.) finds that scaffolding cannot exceed what the base model can evaluate. Our read step depends entirely on whether a 7–9B model's internal response tracks argument strength. We measure that directly instead of assuming it.

Research Program

Stance directions. For each proposition we build pro and con contrastive pairs from Argument Quality Ranking arguments. We extract layer-wise CAA directions on Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct and Gemma-2-9B, which are the models on which Bigelow et al. validated their relation. We fit $a$ per model by sweeping $m$ and measuring probe log-odds.

Topic-general direction. We also test a single "agree with the stated claim" direction that is shared across topics. A shared direction would remove per-topic extraction and is what large populations need.

Read step. We implement the read as one cached probe query before and after the new messages.

Batched serving. We build a serving loop in which a shared model instance applies a different steering magnitude to each row of the batch. Per-agent state is then a few floats. Context per round is bounded by the current inbox plus a short persona line.

Argument-strength analysis. Separately, we correlate $\Delta\ell_t$ with AQR crowd strength labels. This quantifies how much of the DeBERTa scorer's signal the base model already contains internally.

Baselines

The naive reference points are the two reported on DEBATE replay by Belief Engine. The no-change predictor, which assumes nobody moves, reports final-stance RMSE 0.489. The net-evidence linear baseline reports 0.488 (Belief Engine). Any method that does not beat these has learned nothing about opinion change.

The strongest existing method is Belief Engine itself, which reports RMSE 0.465 on DEBATE final-stance prediction and 0.393 on the subgroup of evidence-aligned movers. It is the strongest competitor because it is the only system in the file that predicts individual human stance change with explicit uptake and anchoring traits. We will run it on the same DEBATE splits with the same three base models. Its uptake and anchoring parameters and our $u$ and $\kappa$ get an equal-size search on training topics. Beating it means lower RMSE than 0.465 overall, without losing on the mover subgroup, at a fraction of the model calls per turn.

The text-memory baseline follows Jeong & Na: the agent rereads its full or summarized history each round. It is the strongest competitor on long-horizon trajectory fidelity. Beating it means equal or better fidelity with flat rather than growing tokens per round.

Two ablations isolate the key ingredient. The first, "read-our-way, write-to-prompt," uses our internal $\Delta\ell_t$ update but conditions generation on the scalar stance in text. This isolates whether writing into activations matters. The second, "read-their-way, write-to-activations," uses Belief Engine's external scorer but commits the result as a steering offset. This isolates whether the model's own read is adequate. A static-CAA persona with no updating rounds out the set.

Why not just…?

Why not just put the stance number in the prompt, as Belief Engine already does? This is the most serious threat. A one-line "your current stance is 0.3" is nearly as cheap as an activation offset. Two pieces of evidence suggest it is weaker. Roytburg & Ippolito find base-model habits overpower prompted personas in networks. Bigelow et al. show prompt evidence accumulates with diminishing returns while steering shifts the prior directly. Neither result tests a stance scalar specifically, though. The read-our-way/write-to-prompt ablation settles it. If it matches our method on RMSE, long-horizon drift and the base-model variance decomposition, the honest conclusion is that the internal read step is the contribution and the activation write is not.

Why not just fine-tune agents on human deliberation data? Chuang et al. found fine-tuning raised the Human Likeness Index from -33.95 to 50.11 on training data. It also produced extreme, overfit outputs. A fine-tuned model has no explicit, inspectable stance state to set per agent, and serving thousands of heterogeneous adapters costs far more than per-row offsets on one model.

Why not just use a classical numeric opinion model? Decoding Echo Chambers reports that bounded-confidence and Friedkin-Johnsen models predict polarization shifts around -0.2 in small-world networks, where language agents produce +0.584 and show entrenchment. Our method keeps language-mediated reading while borrowing only the state bookkeeping from numeric models.

Why not just keep Belief Engine's DeBERTa scorer, since small models may judge arguments badly? This is the base-model bottleneck. It may hold. The AQR correlation analysis and the read-their-way ablation measure exactly how much accuracy the internal read loses. If the loss is large, the deployable design is a hybrid: an external scorer for the read and the activation write for the commit.

Proposed Experiments

DEBATE replay. We run the 2,495 human trajectories across 107 topics and report final-stance RMSE against the 0.489, 0.488 and 0.465 references. We also report the correlation between $m_t$ and an independent judge's rating of generated text, for which Belief Engine reports $r=0.967$ between its scalar and judged text. This tests whether the activation state actually reaches the output.

Long horizons. We extend agents to 30+ rounds with neutral or repeated inputs and measure unprompted stance drift. Text memory should drift, and the latent state should not.

Heterogeneous networks. We mix the three base models in small-world, scale-free and random topologies modeled on Decoding Echo Chambers, at 50 agents and then at 500 and 5,000. We decompose variance in stance change into base-model identity versus agent state and persona. Success means the base-model share shrinks relative to prompt-conditioned agents. We also check whether the qualitative polarization direction reported at 50 agents (+0.584) persists at scale.

Cost. We record tokens per agent-round, model calls per turn and throughput versus population size for all methods.

Opinion trajectories. We will attempt trajectory fidelity against the Seoul weekly polling series, for which there is no reported result found on Lacuna. Here it is exploratory and compared only against our own text-memory reimplementation.

Decisive checks. Three outcomes would decide the proposal. The method must beat 0.465 RMSE, or match it with substantially fewer model calls. The prompt-write ablation must fall short on drift or base-model dominance. The topic-general direction must hold up well enough to make large populations practical. If the first passes but the second does not, we report the read step as the contribution.

Workshop authors to connect with