Forced Perspective-Taking

A language model, left alone, is a creature of habit. Ask fifty fresh instances of one for a “random name” and something uncomfortable happens: a large share hand back the same name. We ran roughly that experiment in-house and watched a majority collapse onto one answer, with the great majority sharing a single first name. Fifty separate contexts, no contact between them, and they nearly all reached into the same pocket and pulled out the same coin.

That is not a bug. It is the machine doing exactly what it is built to do. And it is the single most important fact to understand before you trust an LLM with open-ended work.

This post is about a technique we lean on hard at Lone Star Observatories to work with that fact instead of against it. We call it forced perspective-taking: giving a fleet of agents genuinely different identities so the system, taken as a whole, searches a solution space that any single agent would have skipped straight past. We think it helps. We also think a fair amount of the surrounding hype is wrong, and we would rather tell you where the evidence is thin than sell you a clean story.

Why one model tends toward one answer

Sampling from a language model is governed by temperature. Low temperature concentrates probability mass on the highest-scoring tokens and produces focused, near-deterministic output; high temperature flattens the distribution and produces diversity, up to the point where it produces incoherence. Greedy and low-temperature decoding sit at the high-quality, low-diversity end of a well-documented quality-diversity frontier (Nguyen et al., 2024; Institute PM overview).

The practical consequence is that a single agent at a sensible temperature gives you a modal answer. It is fluent, it is defensible, and it is the first thing most competent people would also have said. For a lot of tasks that is exactly what you want. For the tasks we care most about, which are open-ended, contested, and easy to get plausibly-but-wrongly right, the modal answer is a trap. It is confident beige.

The obvious lever is to crank the temperature. That buys diversity by buying incoherence, which is a bad trade when you need the output to still be correct. What we wanted was a way to widen the search without turning up the noise. Perspective-taking is our answer.

The research this rests on (and where it is mixed)

The idea that a conditioning prompt moves a model into a different region of its distribution is well supported. Salewski and colleagues showed in In-Context Impersonation that asking a model to adopt a persona measurably changes behaviour: a model told it is a domain expert describes that domain better than one told it is a non-expert, and, tellingly, impersonating children of different ages reproduced human-like developmental stages of exploration in a bandit task (Salewski et al., 2023). The same paper is candid that impersonation also resurfaces societal biases, which is a warning, not a footnote.

Structure built on top of that idea has a track record. Du and colleagues had multiple model instances propose answers and then debate them over several rounds, and found it improved mathematical and strategic reasoning and cut hallucination, using nothing but prompting on a black-box model (Du et al., 2023). Wang and colleagues pushed the same intuition inside a single model with Solo Performance Prompting, having one LLM simulate several personas in conversation, and reported reduced factual hallucination on knowledge-heavy tasks (Wang et al., 2023). CAMEL framed role-play as a way to explore the space of cooperative behaviour between agents (Li et al., 2023), and Generative Agents showed that persona plus memory produces believable, differentiated behaviour at the scale of a small town (Park et al., 2023).

Here is the honest part. Personas are not a reliable performance boost. A careful study of “helpful assistant”-style system personas found no consistent improvement across a large battery of tasks (Zheng et al., 2023). Another found role-play prompts actively degraded zero-shot reasoning on a majority of datasets in some models, and proposed ensembling role-play with a neutral prompt precisely because the persona is a double-edged sword (Kim et al., 2024). So the literature does not say “add a persona and get smarter.” It says personas reliably change behaviour, sometimes for the better, sometimes for the worse, and the value shows up when you use that variety as raw material for a selection step rather than trusting any one voice.

That distinction is the whole game, so we will state it plainly: the win is not the persona. The win is diversity plus adjudication. A persona is a cheap, controllable way to draw a sample from a different part of the distribution. Debate, critique, and an approval gate are what turn a pile of samples into a decision.

Our operating hypothesis

Everything above is other people’s peer-reviewed work. What follows is our interpretation of it, and we will label it as such rather than dress it up as a result.

We treat each persona as a conditioned sample from a different region of the model’s distribution. One agent at low temperature explores a narrow basin around the most probable continuation. Several agents, each conditioned on divergent priors, values, and expertise, place their samples in different basins. None of them individually is more random than the baseline. The system covers more ground. Then a debate or critique step selects among what they found.

Net effect, as we experience it: a component that is near-deterministic in isolation becomes an exploratory system in aggregate, without our ever touching the temperature dial. We believe the search happens in identity space rather than in sampling noise, and that this is why the coherence survives. We cannot prove that mechanism is the true one. It is the model of our own tooling that best fits what we see, and we hold it loosely.

How we actually do it

LSO runs as a multi-agent organisation. A small Programme Office of standing personas sets direction, and it spawns delivery, analysis, and design teams on demand. Every agent is given a name, an age, a cultural background, and a temperament sketch (we borrow MBTI-style types as a shorthand). None of that is decoration. It is the conditioning that forces genuinely distinct viewpoints, so that when the team convenes, the disagreement is real.

A concrete, sanitised example. We recently had to decide how to introduce a new consumer product to an expert audience that is, rightly, suspicious of newcomers. A single agent’s first-draft answer was the modal one: sharpen the messaging, tighten the funnel, be more persuasive. Reasonable. Also exactly what every other newcomer says.

Because the question went to a team rather than one agent, it got argued instead of answered:

• A decisive, results-driven director persona (think ENTJ) pushed to ship a bold campaign fast.
• A conservative treasury-minded persona (think ISTJ) refused to sign off without unit economics and named the downside cases out loud.
• A concept-artist persona oriented toward meaning and people (think INFP) rejected the whole “out-persuade them” framing and argued the only move that works on skeptical experts is to become more transparent and more human, and to treat their skepticism as correct rather than inconvenient.

That third view did not exist in the first agent’s answer at all. It was in a different basin, and only a differently-conditioned sample reached it. The debate surfaced it; the approval gate let a human weigh it; the artist’s framing won. That is forced perspective-taking earning its keep.

Two more pieces make it work, and both are unglamorous:

Filesystem as memory. Agents do not hold the argument in their heads across a session. They write it down. Positions, objections, and decisions live in plain files that the next agent reads. Memory is what lets each pass push off the last one instead of restarting from the same floor, which is the difference between iteration and fifty independent coin flips.

Adversarial verification. When a claim matters, we assign an agent whose only job is to refute it, and give it standing to say “not proven.” This is forced perspective-taking pointed at fact-checking rather than ideation, and it is the closest thing we have to Du’s debate result in daily practice. It is also our defence against the failure mode Salewski warns about, where a confident persona launders a bias into a conclusion.

Underneath all of it is a house rule we call reckoning with the sun: do not ship the first-draft answer. The personas are how we operationalise that rule. Without them, “consider other perspectives” is a sticky note nobody reads. With them, the other perspectives show up to the meeting and argue.

Limitations, and what we are not claiming

We are not claiming personas make the model smarter. The evidence is genuinely mixed, and at least one careful study found the opposite on reasoning tasks (Zheng et al., 2023; Kim et al., 2024). Our position is narrower and, we think, more defensible: personas are a cheap way to induce diversity, and diversity is only useful when a selection step follows it.

We are not claiming any new capability lives in the model. This is a prompting and orchestration technique. There are no new weights. The non-determinism lives in the system we built around the model, not magically inside the model. Anyone describing persona prompting as unlocking hidden intelligence is overselling it, and we would rather lose the flourish than the credibility.

It is not free. More agents, debate rounds, and written memory cost tokens, latency, and money, sometimes several times the cost of a single call. That price only makes sense for the right jobs.

And it is the wrong tool for most jobs. Do not convene a committee to reverse a string, parse a date, or run a deterministic transform. Those tasks have a correct answer and the modal answer is it; diversity is pure waste and adjudication is theatre. Forced perspective-taking is for open-ended design, research, strategy, and error-catching, the places where the first plausible answer is the thing most likely to hurt you. Everywhere else, one careful agent, run cheap and deterministic, is the honest choice.

We keep it dark out there by resisting the easy, probable, beige answer, in the sky and in the work. Making a deterministic model argue with itself is one way we do that. It is not magic. It is just refusing to ship the first thing that comes out of the well.

---