Adversarial Reasoning: Multiagent World Models for Closing the Simulation Gap
Feb 8, 2026
ai-researchreasoningworld-modelsmultiagentgame-theory
Guest essay by Ankit on Latent Space, introduced by swyx. Argues that current LLMs — including reasoning models — have a fundamental blind spot: they can’t model being modeled.
The Core Problem
LLMs are optimized to produce completions that a human rater approves of in isolation. RLHF pushes models toward being helpful, polite, balanced, and cooperative. This is a bad default in adversarial settings because it systematically under-weights second-order effects:
- How the counterparty will interpret your message
- What it signals about your leverage
- How they’ll update their strategy after reading it
LLMs produce artifacts that look expert. They don’t yet produce moves that survive experts.
Three Kinds of World Models
The essay situates adversarial reasoning as the third frontier of world models:
- 3D video world models — Fei-Fei Li’s Marble, Google’s Genie 3
- Latent representation models — Meta’s JEPA/V-JEPA/EchoJEPA school pursuing Platonic representation
- Multiagent world models — AI that tracks theory of mind, anticipates reactions, and reveals/mines for information in adversarial situations
Why Reasoning Models Don’t Fix This
Reasoning models are solipsistic. They aren’t thinking about the counterparty’s hidden incentives or emotional deficits. Even if prompted to “consider the opposition’s perspective,” their learning is text-based. They treat social dynamics as a causal chain of words (if "sorry" → then "forgiven") rather than a collision of incentives.
The training signal mismatch is key: domain experts get trained by the environment — if your argument is predictable, it gets countered; if your concession leaks weakness, it gets exploited. LLMs mostly learn from descriptions of those dynamics and from static preference judgments, not from repeatedly taking actions in environments where other agents adapt and punish predictability.
The Pluribus Lesson
Pluribus (Meta’s poker AI) cracked what current LLMs can’t: in an adversarial environment, your opponent is watching you and updating. Pluribus calculated how it would act with every possible hand, then balanced its strategy so opponents couldn’t extract information from its behavior. The LLM can’t model being modeled — that gap is exploitable, and no amount of “think strategically” prompting fixes it.
What Adversarial Robustness Requires
To behave adversarially robust by default, a model must reliably:
- Detect that the situation is strategic (hardest step — no default ontology distinguishes “cooperative task” from “task that looks cooperative but will be evaluated adversarially”)
- Identify the relevant agents and what each is optimizing
- Simulate how those agents interpret signals and adapt after your move
- Choose an action that remains good across plausible reactions
Step 1 is the real bottleneck. The model has no way to distinguish context types without explicit framing.
Scaling Won’t Solve It
More raw IQ doesn’t fix the missing training loop. An LLM given an “aggressive negotiator” prompt will execute that strategy consistently — which means a human can probe, identify the pattern, and exploit its predictability. The LLM doesn’t observe that it’s being tested.
The deeper handicap: the thing you’re trying to learn is not fully contained in the text. LLMs can catch up by brute force, but are far more inefficient than humans at learning strategic dynamics from experience.
The Fix: Outcome-Based Training
The article argues we need models trained on the question humans actually optimize: what happens after my move? Grade the model on outcomes (did you win the negotiation, did you concede leverage, did you get exploited), not on whether the message sounded reasonable. This requires multi-agent environments where other self-interested agents react, probe, and adapt.
Why It Matters Now
As LLMs get deployed as agents in procurement, sales, negotiation, policy, security, and competitive strategy, exploitability becomes practical. Google DeepMind is already expanding AI benchmarks beyond chess to poker and Werewolf — games that test social deduction and calculated risk. Both DeepMind and ARC-AGI/Code Clash are modeling these as games, but solving adversarial reasoning is serious business.
The essay frames this as evidence that the age of pure scaling is flipping back to the age of research.