Language agents solve complex tasks through plans and actions. A single step the world refuses puts the goal out of reach, and what the agent does next decides the task. Prompted planners fail at exactly this point, rewriting the refused step in new words, meeting the same refusal, and burning the attempt budget without moving. They fail because the plan was never tied to the world, so a refusal has nothing in the plan to attach to. A world is where a task runs, and it has its own rules, its own admissible actions, and its own constraints. We build synthetic worlds across 7 domains and extract training data from them. A program enforces each world's rules and grades its goal, and every world is admitted only if its goal is reachable from its initial state. Agents run inside and leave verified failures paired with repairs that carried the run to a state the world certified, a record of about 226K trajectories. On this record we train the World State Generator, a model that writes a plan as checkable states of the world and keeps that plan aligned with the world it runs in. That alignment is what a plan written in language lacks, since the world it runs in has physical limits, logical dependencies, and required orders the language never states, and the plan encounters these rules only when a state fails. WSG takes that failure as the rule the world has stated and rewrites the remaining states to obey it, so the plan bends to the world as the run goes on. Across 7 public benchmarks, WSG raises end-to-end success for two open models near 30B parameters over prompting and brings to the level of proprietary model.
Figures & tables
Figure 1: Left: worlds are built from entities, rules, and a goal, and a world is kept only if its goal is reachable from its initial state. Middle: agents run the SDP loop inside these worlds. When the world refutes a state, the certified prefix stays and the rest of the chain is rewritten. Right: each refutation becomes a training pair, the input (g,c,τ) and the rewritten chain. A pair is kept when the rewritten chain reaches a state the world certifies, whether or not the run later reaches the goal.
Domain
Component
Description
Physical
Entities
Objects with locations and properties.
Actions
Primitive actions such as open, move, and combine.
Enforced rule
An action whose precondition does not hold changes nothing.
Goal check
The goal predicate on the final configuration.
Admission
Exhaustive search over configurations.
Software
Entities
An index of Wikipedia documents.
Table 1: Entities, actions, enforced rules, goal checks, and admission criteria for the 7 domains.
Domain
What the program decides
Model
Episodes
Refuted (%)
Reached (%)
Repair calls
Kept (%)
Physical
the goal check on the engine state
Qwen3-32B
11,064
90.6
24.9
85,921
33.3
Qwen3.6-27B
10,614
91.5
19.1
87,231
39.2
Gemma4-31B
10,097
87.9
19.0
80,134
15.8
Software
the submitted answer against the gold answer
Qwen3-32B
11,040
86.2
43.4
37,870
27.0
Qwen3.6-27B
10,239
69.4
60.6
27,357
30.7
Gemma4-31B
10,380
67.3
53.0
28,200
33.3
Table 2: 7 domains, what the program grades in each, and what each extraction model produced inside them. Episodes counts runs of the SDP loop, Refuted the share of episodes in which the world refuted at least one state, and Reached the share that certified the goal. Repair calls counts the repair calls the proposer made, and Kept the share of those that survived the corpus filters. Intermediate states are judged by the validator, by a program in reasoning worlds, and the goal by the program in every domain.
Gemma4-31B
Qwen3.6-27B
Benchmark
Metric
Proprietary models (published)
SDP
WSG
SDP
WSG
τ2 -bench
Accuracy
46.3 Gemini-2.5-Pro
53.7 Sonnet-4
54.3 GPT-4.1
79.9 GPT-5-Think
42.5
56.3 (+13.8)
68.3
76.6 (+8.3)
ScienceWorld
Average score
36.4 GPT-4 ReAct
39.2 GPT-4 CoT
45.3 GPT-4 Reflexion
47.9 GPT-4 Plan-and-Act
35.8
44.4 (+8.6)
36.4
46.8 (+10.4)
ToolSandbox
Average score
64.3 GPT-4
65.6 GPT-3.5
69.2 Opus-3
73.0 GPT-4o
64.3
72.5 (+8.2)
64.7
69.6 (+4.9)
PlanBench-XL
Accuracy
51.9 GPT-5.4
52.2 Gemini-3.5-Flash
63.1 DeepSeek-V4-Flash
77.1 Gemini-3.1-Pro
38.5
53.9 (+15.4)
16.1
48.0 (+31.9)
REAL
Accuracy
25.0 o3-mini
35.0 Gemini-2.5-Pro
38.0 o3-pro
41.0 Sonnet-3.7
27.7
37.3 (+9.6)
22.6
40.2 (+17.6)
Table 3: Zero-shot performance on 7 agent benchmarks. SDP prompts the backbone for every proposer call, and WSG replaces the calls under τ with the trained model on the same backbone and runtime, with the gain over SDP in parentheses. Bold marks the best proprietary result and the best open result per row.
Gemma4-31B
Qwen3.6-27B
Benchmark
Recovery (%)
Steps
Replans
Time (s)
Recovery (%)
Steps
Replans
Time (s)
τ2 -bench
51.0 / 64.0
42.2 / 41.8
5.91 / 5.57
15.6 / 13.9
74.0 / 81.0
38.8 / 42.6
3.57 / 2.47
16.3 / 15.6
ScienceWorld
10.6 / 27.5
73.6 / 43.8
4.19 / 3.07
348 / 223
13.8 / 22.9
73.1 / 40.2
3.89 / 2.23
404 / 215
ToolSandbox
71.3 / 85.1
10.9 / 10.1
2.14 / 1.05
27 / 31
66.5 / 72.7
11.9 / 12.1
2.37 / 1.46
22 / 20
PlanBench-XL
40.2 / 63.1
32.0 / 30.7
1.90 / 4.55
307 / 335
20.8 / 48.4
32.8 / 36.2
1.30 / 4.26
259 / 353
REAL
17.2 / 25.0
33.6 / 27.6
3.94 / 2.43
832 / 376
18.9 / 23.0
41.9 / 28.6
3.92 / 2.01
1077 / 478
Table 4: Recovery after refutation and its cost, reported as SDP / WSG. Recovery is the share of episodes that reached the goal after at least one refutation. Steps counts messages per episode, Replans counts proposer calls under τ per episode, and Time is wall-clock seconds per episode.