A Stanford-led team built 85 language models from scratch — with compute budgets spanning from 10¹⁸ to 10²⁰ FLOPs — and found that scaling reliably improves most social simulation tasks, but fails to fix how models reproduce human cognitive biases like risk aversion [S1]. That gap matters because the entire premise of AI-driven social simulation rests on a bet: that bigger models will eventually simulate people faithfully enough to replace surveys, focus groups, and behavioural studies. Whether that bet pays off depends on a question this preprint now forces into the open — and the answer isn't what the scaling faithful would hope.

The experiment nobody had run

Social simulation with LLMs — using AI agents to stand in for human populations in opinion polls, behavioural experiments, and forecasting — is, in the authors' words, "promising but not yet faithful enough" for wide adoption [S1]. The technique has drawn attention from researchers who see it as a cheaper, faster alternative to running real surveys or controlled behavioural studies. But nobody had systematically tested whether the field's central assumption — that bigger models will close the fidelity gap — actually holds.

The team, spanning Stanford University and Open Athena [P2], did something methodically unusual. They pre-trained 85 transformer models from scratch using the Qwen3 architecture — the model series from Alibaba Cloud's Qwen team [P5] — on the DCLM web text corpus, with compute budgets spanning two orders of magnitude [S1]. They then evaluated 35 larger open-weight models up to 70 billion parameters [S1]. The goal was to trace, model by model, how social simulation accuracy changes as you add compute — and whether the relationship is clean enough to predict.

It is, mostly. The authors report "strong compute scaling" across all three sub-domains they tested: opinion modelling, behavioural simulation, and longitudinal forecasting [S1]. They claim their approach can predict downstream accuracy from training loss [S1] — meaning you could estimate how well a model will simulate a population without running the full simulation first.

Where scale stops working

The exceptions are where this gets interesting.

Most behavioural and opinion tasks improve rapidly with scale — but, the authors stress, "particularly when they involve populations that are well-represented in English web corpora" [S1]. That qualifier is doing heavy lifting. If your simulated population is English-speaking, Western, and heavily discussed online, bigger models get better at representing them. If not, the gains slow.

Longitudinal forecasting — predicting how opinions or behaviours shift over time — scales more slowly than static opinion modelling [S1]. So does simulation of underrepresented opinions [S1]. The authors note these slower-scaling tasks tend to be less correlated with standard AI benchmarks like MMLU, a widely used test of general knowledge and reasoning [S1]. In other words, the tasks that don't overlap with what makes a model "smart" in the conventional sense are the ones where more compute buys you less.

Then there's the hard wall. In behavioural simulation, scaling "fails to improve model calibration with human cognitive biases like risk aversion" [S1]. Models also fail to pick up human heuristics — mental shortcuts such as learning correlated rewards from related tasks [S1]. Even fine-tuned models showed no noticeable improvement from 0.5 billion to 8 billion parameters on these specific tasks [S1]. Bigger doesn't help. The bias stays.

The pattern is stark: scaling improves the things benchmarks measure, but human cognitive biases fall outside that measurement frame entirely.

What it means

The core finding is a split verdict on the scaling thesis as applied to social simulation. For most use cases — predicting what a population thinks, how it might respond to a policy change, whether a product feature will land — bigger models will get better, and the improvement is predictable enough to plan around [S1]. That's genuinely useful. It means a research team can estimate, before spending the compute, how accurate their simulation is likely to be.

But the failure on cognitive biases is a structural limitation, not a temporary one. Risk aversion — the tendency for people to prefer avoiding losses over acquiring equivalent gains — is one of the most documented findings in behavioural economics. If a 70-billion-parameter model can't reproduce it, and scaling doesn't help, the problem isn't a lack of data or parameters. It's that the model is learning a different thing than what humans do. Web text describes risk aversion; it doesn't embody it.

Think of it this way: imagine training a brilliant reader of psychology textbooks to act as a stand-in for a real person. They can tell you what people tend to think. They can't feel the loss aversion that makes someone refuse a favourable bet. More textbooks won't fix that.

What it means for business

For a two-person market research firm, the findings cut both ways. On one hand, LLM-based opinion simulation is getting cheaper and more reliable — if your target audience is well-represented in English web data, scaling up your model will predictably improve results [S1]. A suburban agency testing ad copy or product concepts against a mainstream demographic could see real value within this quarter.

On the other hand, any simulation involving underrepresented populations, long-horizon forecasting, or behavioural economics is not ready to replace human studies. A campaign strategist simulating how a niche demographic's views will evolve over six months is working in the slow-scaling zone [S1]. A product team that needs to understand how loss-averse consumers will respond to a pricing change is working in the no-scaling zone [S1].

The practical takeaway: use LLM social simulation as a fast first pass for well-represented populations and static opinion questions. Treat it as a hypothesis generator, not a replacement for targeted research, when the task involves time dynamics, minority viewpoints, or cognitive biases.

What we don't know yet

This is a single preprint, not peer-reviewed [S1]. Several limitations are explicit:

  • All models use the Qwen3 architecture and DCLM corpus — generalisation to other model families like Llama, Mistral, or proprietary systems is unverified [S1].
  • Predictions about scaling rely on extrapolation from scaling laws, not direct observation at every scale [S1].
  • The "rapid improvement" claim is conditional on English web representation; non-English and non-Western populations may scale differently [S1].
  • The cognitive-bias failure is observed but not explained — the authors don't propose a fix, and it's unclear whether architectural changes, different training data, or new fine-tuning methods could address it.

The next concrete signal to watch: whether independent teams replicate these findings on other model families, and whether the cognitive-bias gap narrows with new training methods rather than raw scale. Until peer review and replication, treat the scaling optimism as provisional — and the bias failure as the more durable finding.

If this kind of evidence-first analysis is what you want in your inbox, subscribe — there's more where this came from.

Sources

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.