SILICA benchmark exposes LLM agents as poor human behavior proxies
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have released SILICA, an open-source benchmark that directly tests whether LLM agents replicate human behavior in social experiments or simply memorize training data. The tool runs twelve open-weight models through five environments with published human baselines, plus rule-preserving perturbations and payoff variants designed to break memorized patterns. Early results show agent-human alignment only at initialization, suggesting that apparent social dynamics may reflect dataset reproduction rather than genuine interaction. This work addresses a critical validity gap in the growing use of agent populations as computational testbeds for social science.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
Skeptical readOur AI-generated reading of the wider context and the next developments to watch.
SILICA's real contribution isn't proving agents memorize social data (that's expected). It's the perturbation strategy: by tweaking payoffs and rule sets, the benchmark attempts to distinguish genuine behavioral learning from dataset reproduction. But the paper doesn't clarify whether agents trained on diverse social datasets would show better generalization, or whether the problem is fundamental to how LLMs encode social reasoning.
This connects directly to the confidence-divergence work from August and the persona-dialogue study. Both exposed gaps between what models claim to do and what they actually execute. SILICA extends that pattern: agents appear to engage in social dynamics, but the internal mechanism (memorization vs. reasoning) remains opaque. The PersonaForge and CultureConverse papers from the same week also highlight how multi-turn interaction exposes failures that single-shot evaluation misses, reinforcing SILICA's core finding that static benchmarks underestimate agent brittleness.
If the same twelve models show improved agent-human alignment when fine-tuned on SILICA's perturbation variants (rather than original environments), that suggests the gap is remediable and points toward social reasoning as learnable. If alignment stays flat across retraining, the memorization hypothesis holds and agent-based social simulation becomes fundamentally limited for social science.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
LLM confidence statements diverge from internal uncertainty across 30 models
A systematic study across 30 LLMs reveals that when models are asked to express confidence linguistically, their stated certainty often misaligns with internal probability signals derived from logits and semantic entropy. The research exposes a critical gap in model transparency: instruction-tuned variants tend to report higher confidence but don't necessarily track their actual uncertainty better.…
MentionsSILICA · arXiv
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Benchmarking large language model agent societies against human behavioural distributions”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.