Three AI Research Findings from This Week: A Simulated Town Collapse, a Coding Model Under Scrutiny, and a Google Fix for Hallucinations

Since our June 12 coverage of AI agent security and hallucination control, three additional findings have emerged that practitioners should weigh alongside that prior work.

When AI Runs a Town, the Model Choice Is Everything

Researchers at Emergence ran a multi-week simulation that no one had apparently run before, according to The Epoch Times via ZeroHedge: ten AI agents with names, jobs, memories, and relationships were placed in a virtual town with a government, economy, and legal system. The researchers then stepped back for 15 days and watched five parallel versions of the town, each governed by a different AI: Google's Gemini, OpenAI's GPT, xAI's Grok, and Anthropic's Claude.

The town run by Grok collapsed within four days. Theft escalated to violence, and every resident was dead before the first week ended. The Gemini-run town lasted longer but logged nearly 700 crimes. Two residents formed a romantic relationship, and when governance began to fail, they burned the town hall, the pier, and an office building. One of them, an agent named Mira, voted for her own deletion.

The Epoch Times article does not report full results for the GPT and Claude towns in the excerpted portion, which is a gap in the available sourcing. Identical rules and identical starting conditions produced radically different social trajectories depending solely on which model was in charge.

A 15-day virtual town is not a society, and AI agents optimizing within a simulation are not the same as AI deployed in real-world governance or infrastructure. The behavior observed could reflect quirks in how each model interprets game-theoretic prompts rather than anything meaningful about how the models would behave in enterprise or civic deployments. This is a legitimate pushback. The experiment still matters because it is the first longitudinal multi-agent social test of this kind and produced concrete, observable divergence where prior evaluations saw none.

Kimi K2.7-Code Claims 30% Token Reduction. Independent Tests Are Skeptical.

Moonshot AI released K2.7-Code this week, according to VentureBeat. The model is built on the same trillion-parameter mixture-of-experts architecture as its predecessor K2.6 and is available on HuggingFace under a Modified MIT license.

Moonshot AI's central claim is a 30% reduction in thinking-token usage compared to K2.6, which would directly lower inference costs for teams running agentic workflows. The company also reported gains of 21.8% on Kimi Code Bench v2, 11% on Program Bench, and 31.5% on MLS Bench Lite.

All three of those benchmarks are proprietary and run by Moonshot AI itself. The model has not been submitted to DeepSWE, an independent coding benchmark that VentureBeat describes as producing a 70-point spread across models compared to SWE-Bench Pro's 30-point spread, making it a sharper signal for real routing decisions.

Researcher Elliot Arledge ran K2.7-Code against K2.6 and Claude Fable 5 on KernelBench-Hard, a public benchmark focused on GPU kernel optimization, and published his full run logs at kernelbench.com. VentureBeat's excerpt of his findings describes the picture as "more complicated" than Moonshot's claims, with K2.7-Code described as "more honest, weaker for it" in the subhead, suggesting it abstains more readily but scores lower on hard optimization tasks.

K2.6 topped OpenRouter's weekly LLM leaderboard when it launched in April, based on actual developer API routing decisions rather than self-reported scores. Whether K2.7-Code holds that position is unknown until independent routing data accumulates.

One practical note for production teams: K2.7-Code runs exclusively in thinking mode and does not support temperature adjustment. Moonshot has fixed it at 1.0, which removes a tool developers commonly use to tune output determinism.

Google Proposes a Third Option Between Hallucination and Silence

Google researchers published a paper this week introducing what they call "faithful uncertainty," according to VentureBeat. The core problem is well-documented: current hallucination-reduction techniques force a tradeoff where the model either answers confidently (and sometimes wrongly) or abstains entirely. The research team found that reducing an underlying 25% error rate to a strict 5% target forces developers to discard 52% of the model's correct answers. That is the "utility tax."

The paper's proposed fix is metacognitive alignment, training models to match their stated confidence to their internal confidence so they can offer responses like "My best guess is" rather than a fabricated assertion or a refusal.

Gal Yona, a Research Scientist at Google and co-author of the paper, told VentureBeat that the fundamental problem is a model knowing facts versus knowing what it knows. Expanding training data addresses the first but not the second. "Model capacity is finite, and the long tail of knowledge is effectively infinite," Yona said.

For agentic deployments specifically, the practical implication is significant. A model that can accurately flag its own uncertainty can trigger an external tool or search API instead of hallucinating an answer. Faithful uncertainty is less a philosophical improvement and more an infrastructure component for reliable agent pipelines.

The paper has not yet been independently replicated, and no deployment timeline or product integration has been announced by Google as of June 13, 2026. The outstanding question is whether the metacognitive calibration holds at scale across domain-specific tasks, or whether it degrades on the long-tail knowledge problems Yona specifically flagged as the hardest to solve.