If you've heard of "world models", Qwen (Alibaba) has just come out with a large language model that's specially trained to be a "language world model."
(Extensive quotes to follow!)
"World models have been widely recognized as a foundation toward general intelligence, with a growing consensus that learning to predict the world is prerequisite to acting effectively within it. Richens et al. further prove a stronger claim: any agent capable of generalizing across a sufficiently broad range of tasks must have learned a world model, establishing world models not merely as useful but as necessary for general-purpose agents."
"Yet the language environments in which large language model agents operate still lack a general-purpose world model. In the agent -- environment interaction loop, two complementary components are essential: the policy (states -> actions) and the world model ((states, actions) -> subsequent states). However, current research on large language model agents has focused almost exclusively on the policy side. We argue that world modeling is a crucial missing piece in the path to general agents."
"Qwen-AgentWorld, the first language world model that simulates seven agent environments through long chain-of-thought reasoning: MCP, Search, Terminal, Software Engineering, Android, Web, and OS."
MCP stands for "Model Context Protocol", which is a protocol for enabling language models to control applications.
"For the three GUI domains, environment observations are represented as accessibility trees and UI view hierarchies rather than pixel frames. Qwen-AgentWorld is a native world model trained through three stages: continual pre-training injects state-transition dynamics and world knowledge, supervised fine-tuning activates next-state- prediction thinking patterns, and reinforcement learning with hybrid rubric-and-rule rewards sharpens simulation fidelity."
The pretraining (continual or otherwise) is unsupervised text prediction, the supervised fine-tuning is when the model is trained on questions and answers intended to make it an expert in a particular domain, and reinforcement learning entails devising a "reward" signal (e.g. who wins or loses a chess match) that can be used to train a model.
"To evaluate language world models, we construct AgentWorldBench, a comprehensive benchmark across all seven domains."
"We deploy a suite of agent -- environment backends: containerized execution sandboxes for code and tool invocation, MCP servers, persistent terminal sessions with full shell state tracking. For GUI domains, we deploy persistent Android, browser, and desktop OS environments that represent GUI observations as textual accessibility trees and UI view hierarchies for world-model training. These environments run on physical hosts provisioned with Ubuntu, macOS, and Android virtual machines. On top of this infrastructure, we automatically synthesize task queries spanning each domain's target distribution and let agentic systems execute them end-to-end. This pipeline runs continuously and is the primary source of scalable, controlled, reproducible interaction data."
"We collect naturally occurring action -- environment interaction traces from public sources: terminal session recordings, open-source agentic tool-call logs, and execution traces in code repositories."
"We draw from in-house foundation model supervised fine-tuning agentic trajectories accumulated during routine model development, covering all seven domains."
"Continual pre-training data draws from dedicated agent infrastructure, open interaction traces, and specialized-domain world knowledge corpora. supervised fine-tuning and reinforcement learning draw exclusively from internally accumulated trajectories."
"Each system prompt comprises five components: task description, action space, initial state, demonstrations, and simulation instruction."
"Each prompt must encode domain-specific state-transition rules, output formatting constraints, and demonstration patterns. Rather than hand-crafting these templates, we formulate prompt optimization as an automated research problem with a clear objective: maximize the world model's prediction accuracy on held-out real trajectories."
"Reinforcement learning for language world models poses distinctive challenges due to the difficulty and open-ended nature of environment feedback prediction. Moreover, because the context is substantially longer than the target output, language world model reinforcement learning exhibits an extreme prompt -- output asymmetry: the prompt consists of the full trajectory history up to the prediction turn and often extends to tens of thousands of tokens, whereas the output, a single predicted observation, typically contains only a few hundred to a few thousand tokens."
"Using rubrics as structured rewards for reinforcement learning has been shown effective in non-verifiable domains; recursive rubric refinement can further improve judge and reward quality. Each predicted observation is scored by a large language model judge on the five-dimensional rubric defined in section 4.2, each on a 1 -- 5 scale. The total reward equals the mean times 5, yielding a range of [5, 25]."
"A subset of the data carries executable verifier code that produces a binary 0/1 correctness signal, scaled to [0, 25] to align with the rubric's range. Rule-based rewards serve as an objective anchor, effectively mitigating reward hacking induced by open-ended rewards."
"The choice of reward formulation has a strong effect on convergence. We compare the above open-ended reward design against two alternatives. Reference-Reward presents the judge with the ground-truth observation and asks it to choose, in a pairwise A/B test, whether the policy's predicted observation or the initial policy checkpoint's output is closer to the ground truth, yielding a binary 0/1 signal. This design converges slowly: the binary reward is sparse, and when both outputs are plausible but differ in surface form, the judge's preference becomes unstable, injecting noise into the gradient. Turing-Test Reward asks a judge whether the predicted observation could plausibly have come from a real environment. This reward barely converges, primarily because the false-negative rate is too high. When the model's generation is very close to or even identical to the ground truth, asking the judge to determine which one is more likely to have come from a real environment introduces an unreliable training signal regardless of which answer is chosen."
"The policy can learn to exploit the judge's specific biases by inserting self-praising phrases into the predicted observation to inflate scores without improving simulation fidelity."
"AgentWorldBench is built on four construction principles: (i) Widely-Used Queries: all task queries are drawn from established high-quality agentic benchmarks rather than self- constructed tasks, so that the task distribution aligns with the scenarios that current agent development targets; (ii) Frontier-Agent Trajectories: all trajectories are generated by frontier-model agents, whose actions (long reasoning chains, tool-call compositions, and error-recovery sequences) are high-quality and sufficiently complex to stress-test world-model fidelity at the frontier scale; (iii) Real Observations: every trajectory is paired with ground-truth observations from real environment execution, providing a reference for evaluation; (iv) Out-of-Distribution: training data and benchmark queries are partitioned at the data-source level, so that AgentWorldBench probes generalization rather than memorization;"
"AgentWorldBench evaluates simulation quality through an open-ended rubric: a large language model judge scores each predicted observation on five dimensions: Format, Factuality, Consistency, Realism, and Quality. The primary score is the mean across the five dimensions, scaled to [0, 100]. Format measures whether the output obeys the structural conventions of the domain (JSON schema compliance for MCP, shell prompt patterns for Terminal). Factuality measures whether stated facts (file contents, search results, tool return values) are correct. Consistency measures whether the output is internally coherent and coherent with prior turns. Realism measures whether the simulation matches the behavioral characteristics of the real environment as evidenced by the ground truth, including response patterns, style conventions, and value plausibility. Quality measures completeness and conciseness relative to the ground truth: critical information must not be omitted, and the output should not be excessively verbose or abbreviated compared to the reference."
"The judge receives the ground-truth environment observation alongside the predicted observation and scores each dimension by comparing the two."
"Each domain has its own judge prompt that applies the five dimensions in domain-specific terms, ensuring that evaluation not only compares against the reference objectively but also enforces the professional standards of each domain."
"Not all content in an environment observation requires exact matching, and treating all content uniformly produces excessive false negatives. We classify content into three types before evaluation. Deterministic content (echo output, file reads, computation results) must match exactly. Pre-existing environment content (preinstalled software versions, file contents not created by the trajectory) requires only format and plausibility verification, because a simulator cannot reproduce the exact patch version of gcc in a particular sandbox. Runtime metadata (timestamps, PIDs, memory addresses, session tokens) requires only format and range verification."
"We use a double-blind Turing test as a calibration tool to select the judge model and tune the judge prompt."
So after all that, how did it do?
"Qwen-AgentWorld-397B-A17B achieves the highest overall average (58.71), surpassing GPT-5.4 (58.25) and all other frontier models. On text-based domains, Qwen-AgentWorld-397B-A17B leads with an average of 58.07, outperforming GPT-5.4 (56.84) by 1.23 points. The advantage is most pronounced on Terminal (57.73 vs. 53.69) and SWE (68.49 vs. 66.29), the two domains where predictions require accurate modeling of code execution state and tool API behavior. On GUI domains, Claude Opus 4.8 (60.93) and Claude Opus 4.6 (61.12) lead, followed by GPT-5.4 (60.47) and Gemini 3.1 Pro (60.04), with Qwen-AgentWorld-397B-A17B ranking fifth (59.69). The gap reflects an advantage from multimodal pre-training that text-only world modeling does not fully capture."
"Comparing Qwen-AgentWorld with its base checkpoints isolates the contribution of the three-stage pipeline. At the 397B scale, the overall average rises from 54.74 to 58.71. At 35B, the gain is 8.66 points (47.73 to 56.39), lifting Qwen-AgentWorld-35B-A3B above Claude Sonnet 4.6 (56.04) by 0.35 points. The improvement is consistent across both text and GUI domains: at 397B, the text-domain average rises by 3.35 and the GUI-domain average by 4.92."
"Search is the most challenging domain for all models: the best score (37.82) is roughly half the best score on SWE (68.49) or MCP (70.10). Search requires modeling constantly evolving web content, and factual consistency across long retrieval chains remains difficult for all models."
I'm going to stop here, but the paper goes on to describe the model functioning as an "environment simulator" (e.g. simulates a terminal) and "controllable simulation" (does a simulation according to natural-language instructions -- this includes fictional-world construction).
qwen.ai/blog?id=qwen-agentwo…
#solidstatelife #ai #genai #llms #codingai #agenticai #worldmodels
