SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.21173 · 2026

Inverting the Bellman Equation: From Q-Values to World Models

Jakob Foerster, Alistair Letcher, Jonathan Richens, Mattie Fellows, Alexander Goldie, Oliver Richardson

arXiv · PDF · Open in the Atlas

Code that ran

10 of 11 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (3 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 3 have no contract verdict; 6 failures of our instrument; 1 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2606.21173.json · badge: /api/ran/2606.21173.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
aletcher/inverting-bellman community
Papers with Code archive
10 ran · 1 no recorded run
FunctionStatusWhere it lives
ContinuousGoal Ran
ran on a synthesized input · constructed an object; no result was computed
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("8b28b8605c55063e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ContinuousGoal: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp
import numpy as np

# Create small tensors for ContinuousGoal
target_state = jnp.array([[1.0, 2.0, 3.0, 4.0],
                          [5.0, 6.0, 7.0, 8.0]], dtype=jnp.float32)

reward_mask = jnp.array([[1.0, 0.0, 1.0, 0.0],
                         [0.0, 1.0, 0.0, 1.0]], dtype=jnp.float32)

ARGS = {
    "target_state": target_state,
    "reward_mask": reward_mask,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ContinuousGoal — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ContinuousGoal output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
QNetwork Ran
ran on a synthesized input · constructed an object; no result was computed
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("549021cbd6349871")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, QNetwork: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import numpy as np
import flax.linen as nn
from flax.linen.initializers import constant, orthogonal


class GoalRepr:
    """Stand-in class for goal representation with target_state attribute."""
    def __init__(self, target_state):
        self.target_state = target_state


ARGS = {
    'action_dim': 3,
    'num_goals': 2,
    'dense_hidden_size': 128,
    'dense_layers': 2,
    'norm_type': 'layer_norm',
    'sigmoid_output': False,
    'goal_input_dims': None,
    'obs_input_dims': None,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type QNetwork — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a QNetwork output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
WorldModel Ran
ran on a synthesized input · constructed an object; no result was computed
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("7f3428c1d86ea919")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, WorldModel: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import jax
import jax.numpy as jnp
import flax.linen as nn

# Create a small input tensor for testing
# The WorldModel expects input of shape (batch_size, input_dim)
# Let's use batch_size=2 and input_dim=4
input_data = jnp.ones((2, 4), dtype=jnp.float32)

# Initialize the model to get a sense of what it does
model = nn.Module()

# Define ARGS dict with all WorldModel parameters
ARGS = {
    'activation': 'tanh',
    'hidden_size': 64,
    'num_layers': 2,
    'output_dim': 2,
    'output_init_scale': 1.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type WorldModel — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a WorldModel output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_residual_loss Ran
ran; our contract draft was wrong, not the code
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("bb004af58345169b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp

# Create a small delta array (residual/difference values)
delta = jnp.array([1.5, -2.3, 0.8, -0.5])

# Create a config dict with WM_LOSS parameter
# Test with 'mse' to exercise the MSE branch
wm_config = {'WM_LOSS': 'mse'}

ARGS = {
    'delta': delta,
    'wm_config': wm_config,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_wrap_angle Ran
ran, honoured the contract we drafted
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("277e4f29cf1d948a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp
import numpy as np

def _wrap_angle(theta):
    """Wrap angle to [-π, π]."""
    return (theta + jnp.pi) % (2 * jnp.pi) - jnp.pi

# Create test arguments for _wrap_angle
# The function takes a single parameter: theta (an angle or array of angles)

# Test with a scalar
theta_scalar = jnp.array(3.5)

# Test with a small array
theta_array = jnp.array([0.0, jnp.pi, 2 * jnp.pi, -jnp.pi, 5.0])

# Test with a 2D array (batch of angles)
theta_2d = jnp.array([[0.5, 1.0, 1.5, 2.0],
                      [2.5, 3.0, 3.5, 4.0]])

ARGS = {
    "theta": theta_2d
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
apply_wm Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("2f12e5c52bf3aca4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 2, dtype family float, shape (2, 6) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
from jax import random
import flax.linen as nn

# Define a minimal world model as a simple MLP
class SimpleWM(nn.Module):
    output_dim: int
    
    def setup(self):
        self.dense1 = nn.Dense(8)
        self.dense2 = nn.Dense(self.output_dim)
    
    def __call__(self, x):
        x = nn.relu(self.dense1(x))
        x = self.dense2(x)
        return x

# Initialize the model and get parameters
key = random.PRNGKey(0)
wm_model = SimpleWM(output_dim=4)
dummy_input = jnp.ones((2, 8))  # batch_size=2, input_dim=8 (4 state + 4 action)
p_params = wm_model.init(key, dummy_input)

# Create sample tensors
batch_size = 2
state_dim = 4
action_dim = 4
s = jnp.ones((batch_size, state_dim))  # current state
a_oh = jnp.ones((batch_size, action_dim))  # action (one-hot or continuous)

# Define projection functions
def state_to_eff_fn(s):
    """Project full state to effective state (identity for this example)"""
    return s

def eff_to_obs_fn(s_eff):
    """Lift effective state back to observation space"""
    # Simple example: append zeros to go from 4D to 6D
    return jnp.concatenate([s_eff, jnp.zeros((s_eff.shape[0], 2))], axis=-1)

# Build the arguments dictionary
ARGS = {
    'p_model': wm_model,
    'p_params': p_params,
    's': s,
    'a_oh': a_oh,
    'residual': True,
    'state_to_eff_fn': state_to_eff_fn,
    'angle_dims': [0, 1],  # Wrap first two dimensions as angles
    'eff_to_obs_fn': eff_to_obs_fn,
    'wm_input_dims': [0, 1, 2, 3]  # Use all state dimensions
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (2, 6) — shape (2, 6) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_reward Ran
ran; our contract draft was wrong, not the code
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("558feeda58f8f8b0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArrayImpl, ndim 1, dtype family float, shape (2,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax.numpy as jnp
import numpy as np

# Set seed for reproducibility
np.random.seed(0)

# Define small tensors for testing
# obs: observation, shape (batch=2, dim=4)
obs = jnp.array([
    [1.0, 2.0, 3.0, 4.0],
    [5.0, 6.0, 7.0, 8.0]
])

# target_state: target state, shape (batch=2, dim=4)
target_state = jnp.array([
    [1.1, 2.1, 3.1, 4.1],
    [5.5, 6.5, 7.5, 8.5]
])

# reward_mask: mask to select relevant dimensions, shape (dim=4,)
reward_mask = jnp.array([1.0, 1.0, 0.0, 1.0])

# reward_type: one of "sparse", "sparse_negative", "gaussian", "l2"
reward_type = "gaussian"

# sigma: standard deviation parameter for gaussian reward
sigma = 1.0

# a: threshold parameter for sparse reward types
a = 0.5

ARGS = {
    "obs": obs,
    "target_state": target_state,
    "reward_mask": reward_mask,
    "reward_type": reward_type,
    "sigma": sigma,
    "a": a
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
goal_achieved Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("3363b6be1ccdd4d6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 1, dtype family float, shape (2,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# The function performs element-wise operations on tensors and uses:
# - obs: tensor that supports subtraction, multiplication, exponentiation, sum
# - target_state: tensor same shape as obs
# - reward_mask: tensor same shape as obs for masking
# - a: scalar threshold value
# - invert: boolean flag

# Create small tensors with batch size 2 and feature dimension 4
obs = torch.randn(2, 4)
target_state = torch.randn(2, 4)
reward_mask = torch.ones(2, 4)  # mask to include all dimensions
a = 0.5  # threshold value
invert = False

ARGS = {
    "obs": obs,
    "target_state": target_state,
    "reward_mask": reward_mask,
    "a": a,
    "invert": invert
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family float, shape (2,) — shape (2,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_q_network Ran
ran; our contract draft was wrong, not the code
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("dcca8d69ef2ddfe1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned QNetwork. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import numpy as np
import flax.linen as nn
from flax.linen.initializers import constant, orthogonal

# Create a simple goal representation class
class GoalRepr:
    def __init__(self, target_state):
        self.target_state = target_state

# Create a minimal PQN config dict
pqn_config = {
    'ACTION_DIM': 4,
    'NUM_GOALS': 2,
    'NETWORK_DENSE_HIDDEN_SIZE': 64,
    'NETWORK_DENSE_LAYERS': 2,
    'NORM_TYPE': 'layer_norm',
    'NETWORK_SIGMOID_OUTPUTS': False,
    'GOAL_INPUT_DIMS': [0, 1, 2],
    'OBS_INPUT_DIMS': [0, 1]
}

ARGS = {
    'pqn_config': pqn_config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type QNetwork — a QNetwork output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a QNetwork output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_world_model Ran
ran; our contract draft was wrong, not the code
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("e6b68123583448f4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned WorldModel. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import flax.linen as nn
from flax.linen.initializers import constant, orthogonal

# Define the WorldModel and make_world_model as provided
class WorldModel(nn.Module):
    activation: str = 'tanh'
    hidden_size: int = 64
    num_layers: int = 2
    output_dim: int = 2
    output_init_scale: float = 1.0

    @nn.compact
    def __call__(self, x):
        act_fn = nn.relu if self.activation == 'relu' else nn.tanh
        for _ in range(self.num_layers):
            x = nn.Dense(self.hidden_size, kernel_init=orthogonal(np.sqrt(2)), bias_init=constant(0.0))(x)
            x = act_fn(x)
        x = nn.Dense(self.output_dim, kernel_init=orthogonal(self.output_init_scale), bias_init=constant(0.0))(x)
        return x

def make_world_model(wm_config, state_dim):
    return WorldModel(activation=wm_config['ACTIVATION'], hidden_size=wm_config['DENSE_HIDDEN_SIZE'], num_layers=wm_config['DENSE_LAYERS'], output_dim=state_dim, output_init_scale=wm_config.get('OUTPUT_INIT_SCALE', 1.0))

# Create a minimal wm_config dict with all required keys
wm_config = {
    'ACTIVATION': 'tanh',
    'DENSE_HIDDEN_SIZE': 32,
    'DENSE_LAYERS': 2,
    'OUTPUT_INIT_SCALE': 1.0
}

# Define state_dim as a distinct integer
state_dim = 5

# Construct ARGS
ARGS = {
    'wm_config': wm_config,
    'state_dim': state_dim
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type WorldModel — a WorldModel output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a WorldModel output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
world_model_loss_sampled Not yet run
unverified (harvested, no recorded run)
aletcher/inverting-bellman/training/wm.py
code served (permissive licence) · get_code("3e078645516d88c3")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 11 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Model-based and model-free reinforcement learning are traditionally viewed as separate paradigms: instead of learning a model of the transition kernel P , modelfree agents typically estimate value functions tied to a specific policy and reward. In this paper, we challenge this dichotomy by proving that value-based agents trained on a sufficiently rich set of reward functions, e.g. using goal-conditioned RL, implicitly encode a unique and accurate world model. To extract this model in practice, we introduce P -learning, an inverse analogue to Q-learning that samples from an agent's Q-values, policies and rewards to decode its internal model of the environment. We then provide sufficient conditions on the type and number of goals for which agents encode the true kernel P , covering both stochastic and deterministic MDPs over finite or continuous state spaces. Even when our assumptions are violated, we empirically demonstrate that agents trained on a handful of reward functions encode accurate dynamics in Reacher, MountainCar and stochastic variants of FourRooms. Surprisingly, we find that policies trained exclusively on a Reacher agent's implicit world model are quasi-optimal on outof-distribution, velocity-based goals despite position-only training -suggesting that agents contain hidden generalisation capabilities and providing a new lens into the connection between model-based, model-free, and goal-conditioned RL. Figure 1: Illustration of P -learning: extracting the world model contained in an agent's Q-values. Code available at github.com/aletcher/inverting-bellman.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2606.21173")
get_code_for_paper("2606.21173")
have("2606.21173")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2606.21173.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2606.21173.svg)](https://syntology.ai/paper/2606.21173)

Connect an agent — have() is free.