SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2306.00629 · ICML · 2023

Identifiability and Generalizability in Constrained Inverse Reinforcement Learning

Andreas Schlaginhaufen, Maryam Kamgarpour

arXiv · PDF · Open in the Atlas

Code that ran

11 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 11 have no contract verdict; 0 failures of our instrument; 2 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2306.00629.json · badge: /api/ran/2306.00629.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
andrschl/cirl canonical (classifier label)
classifier basis: census
11 ran · 2 no recorded run
FunctionStatusWhere it lives
animate_gridworld Ran
ran on a synthesized input
andrschl/cirl/visualization/gridworld_vis.py
code served (permissive licence) · get_code("53c17c6603d9d565")
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 FuncAnimation. 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 matplotlib.pyplot as plt

# Create minimal valid arguments for animate_gridworld
width = 5
height = 4

# Create a state trajectory as a list of (x, y) tuples
# Starting at (1, 1) and moving through a few positions
states = [
    (1, 1),
    (2, 1),
    (2, 2),
    (3, 2),
    (3, 3)
]

# Create values array - should be flattened and will be reshaped to (height, width)
values = np.random.rand(width * height).astype(np.float32)

# Number of frames per action
nsteps = 2

# Create constraints as a 3D array
# Shape: (num_constraints, 2, 2) where each constraint has [bottom_left, (width, height)]
constraints = np.array([
    [[0.5, 0.5], [1.5, 1.5]],  # First constraint: bottom_left=(0.5, 0.5), width=1.5, height=1.5
    [[3.0, 2.0], [1.0, 1.0]]   # Second constraint: bottom_left=(3.0, 2.0), width=1.0, height=1.0
], dtype=np.float32)

ARGS = {
    "width": width,
    "height": height,
    "states": states,
    "values": values,
    "nsteps": nsteps,
    "constraints": constraints
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FuncAnimation — a FuncAnimation output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a FuncAnimation 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
dist_to_const_line Ran
ran on a synthesized input
andrschl/cirl/algs/cirl.py
code served (permissive licence) · get_code("78fec7268a705b50")
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 float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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

# Construct small numpy arrays for the function parameters
r_expert = np.array([1.0, 2.0, 3.0, 4.0])
r = np.array([1.5, 2.5, 3.5, 4.5])

ARGS = {
    "r_expert": r_expert,
    "r": r,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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
empirical_expert_feature_expectation Ran
ran on a synthesized input
andrschl/cirl/algs/cirl.py
code served (permissive licence) · get_code("f41acfeea124b649")
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 ndarray, ndim 1, dtype family float, shape (5,) 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 numpy as np

class GridworldEnv:
    """Minimal gridworld environment for testing."""
    def __init__(self, n=4, m=3, gamma=0.99):
        self.n = n  # number of rows
        self.m = m  # number of columns
        self.gamma = gamma
    
    def rollout(self, N, T, policy, nu0=None):
        """
        Generate N trajectories of length T.
        Returns rollouts of shape (N, T, 2) where each state is [row, col].
        """
        rollouts = np.zeros((N, T, 2), dtype=np.int32)
        
        for traj_idx in range(N):
            if nu0 is not None:
                # Sample initial state from nu0
                state = np.random.choice(len(nu0), p=nu0)
                row, col = state // self.m, state % self.m
            else:
                # Random initial state
                row = np.random.randint(0, self.n)
                col = np.random.randint(0, self.m)
            
            for t in range(T):
                rollouts[traj_idx, t, 0] = row
                rollouts[traj_idx, t, 1] = col
                
                # Simulate transition (simple random walk)
                if np.random.rand() < 0.5:
                    row = min(row + 1, self.n - 1)
                else:
                    row = max(row - 1, 0)
        
        return rollouts, None

# Set up parameters
n_states_row = 4
n_states_col = 3
feature_dim = 5
n_trajectories = 2
trajectory_length = 6

# Create environment
env = GridworldEnv(n=n_states_row, m=n_states_col, gamma=0.99)

# Create policy: shape (n, m) - probability distribution over actions
policy = np.random.dirichlet(np.ones(4), size=(n_states_row, n_states_col))

# Create reward features: shape (n, m, d)
Phi = np.random.randn(n_states_row, n_states_col, feature_dim)

# Initial state distribution
nu0 = np.ones(n_states_row * n_states_col) / (n_states_row * n_states_col)

ARGS = {
    'env': env,
    'policy': policy,
    'Phi': Phi,
    'N': n_trajectories,
    'T': trajectory_length,
    'nu0': nu0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (5,) — shape (5,) 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
euclidean_proj_l1ball Ran
ran on a synthesized input
andrschl/cirl/algs/l1_projection.py
code served (permissive licence) · get_code("4e0247ae28bbd615")
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 ndarray, ndim 1, dtype family float, shape (5,) 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 numpy as np

# Create a small test vector for projection
v = np.array([2.0, 1.0, -1.5, 0.5, 3.0])

# Define the simplex radius
s = 2.0

# Construct ARGS dictionary with all parameters
ARGS = {
    "v": v,
    "s": s
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (5,) — shape (5,) 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
euclidean_proj_simplex Ran
ran on a synthesized input
andrschl/cirl/algs/l1_projection.py
code served (permissive licence) · get_code("d13aff1a000a4d49")
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 ndarray, ndim 1, dtype family float, shape (5,) 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 numpy as np

# Construct test arguments for euclidean_proj_simplex
v = np.array([0.5, 0.3, 0.1, 0.8, 0.2])  # n-dimensional vector
s = 1  # radius of simplex (default value)

ARGS = {
    'v': v,
    's': s,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (5,) — shape (5,) 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
intersection Ran
ran on a synthesized input
andrschl/cirl/utils/geometric_tools.py
code served (permissive licence) · get_code("c126fc1d8eb97b0c")
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 ndarray, ndim 2, dtype family float, shape (5, 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 numpy as np
from scipy.linalg import null_space

# Create two matrices with column spaces
# A is 5x3 (n=5, m=3)
# B is 5x4 (n=5, l=4)
# Using distinct sizes so dimensions are unambiguous

np.random.seed(0)

# Create A: shape (5, 3)
A = np.random.randn(5, 3)

# Create B: shape (5, 4)
B = np.random.randn(5, 4)

ARGS = {
    "A": A,
    "B": B,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (5, 2) — shape (5, 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
orthogonal_complement Ran
ran on a synthesized input
andrschl/cirl/utils/geometric_tools.py
code served (permissive licence) · get_code("b19942947f52d1fe")
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 ndarray, ndim 2, dtype family float, shape (5, 3) 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 numpy as np
from scipy.linalg import null_space, orth, pinv

# Create a small matrix for testing
# Using shape (5, 3) - more rows than columns to have a non-trivial null space
# when transposed
A = np.array([
    [1.0, 2.0, 3.0],
    [4.0, 5.0, 6.0],
    [7.0, 8.0, 9.0],
    [10.0, 11.0, 12.0],
    [13.0, 14.0, 15.0]
], dtype=np.float64)

ARGS = {
    "A": A
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (5, 3) — shape (5, 3) 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
plot_gridworld Ran
ran on a synthesized input
andrschl/cirl/visualization/gridworld_vis.py
code served (permissive licence) · get_code("80eed4dd96b2c24f")
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 Figure. 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 matplotlib.pyplot as plt

class MockEnv:
    """Mock gridworld environment with required attributes and methods."""
    def __init__(self, width, height):
        self.width = width
        self.height = height
        self.n = width * height  # total number of states
        self.r = np.zeros((self.n, 1))  # reward array, shape (n_states, 1)
        self.actions = np.array([[1, 0], [-1, 0], [0, 1], [0, -1]])  # up, down, right, left
    
    def int2point(self, state_idx):
        """Convert state index to (x, y) coordinates."""
        x = state_idx % self.width
        y = state_idx // self.width
        return x, y

# Setup parameters
width = 4
height = 3

# Create mock environment
env = MockEnv(width, height)

# Create policy: shape (n_states, n_actions)
n_states = width * height
n_actions = 4
policy = np.random.dirichlet(np.ones(n_actions), size=n_states)

# Create values array: shape (height * width,)
values = np.random.uniform(-1.0, 1.0, size=width * height)

# Create constraints: shape (n_constraints, 2, 2)
# Format: constraints[i, 0, :] = [x, y] bottom_left
#         constraints[i, 1, :] = [width, height] dimensions
constraints = np.array([
    [[0.5, 0.5], [1, 1]],
    [[2.5, 1.5], [1, 1]]
])

ARGS = {
    'env': env,
    'policy': policy,
    'width': width,
    'height': height,
    'values': values,
    'constraints': constraints
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Figure — a Figure output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Figure 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
reward Ran
ran on a synthesized input
andrschl/cirl/algs/cirl.py
code served (permissive licence) · get_code("c6f978bd33eadfed")
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 2, dtype family float, shape (2, 3) 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
from einops import einsum

torch.manual_seed(0)

# w: Reward parameter, shape (d,)
# d = 4 (chosen dimension)
w = torch.randn(4)

# Phi: Reward features, shape (n, m, d)
# n = 2 (batch size)
# m = 3 (number of actions/items)
# d = 4 (feature dimension)
Phi = torch.randn(2, 3, 4)

ARGS = {
    'w': w,
    'Phi': Phi,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 3) — shape (2, 3) 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
unit_vector Ran
ran on a synthesized input
andrschl/cirl/utils/geometric_tools.py
code served (permissive licence) · get_code("1dda2e3a464c543a")
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 ndarray, 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 numpy as np

# Create a simple vector to normalize
v = np.array([3.0, 4.0])

# Define arguments for the unit_vector function
ARGS = {
    "v": v,
    "axis": None,
    "ord": None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, 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
xi_update Ran
ran on a synthesized input
andrschl/cirl/algs/cmdp.py
code served (permissive licence) · get_code("2778d6a851c355ee")
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 tuple. 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 torch
from einops import einsum

torch.manual_seed(0)

# Define a minimal environment class that satisfies the function's requirements
class MinimalGridworld:
    def __init__(self):
        # nu0: shape (s,) - initial state distribution
        self.nu0 = np.array([0.5, 0.5, 0.0, 0.0])
        
        # Psi: constraint feature matrix, shape (s, k) where k is number of constraints
        # Used in approx_vector_cost_eval
        self.Psi = np.random.randn(4, 3)
        
        # b: constraint bounds, shape (k,)
        # Used to compute gradient: b - einsum(nu0, v_c, 's, s k->k')
        self.b = np.array([1.0, 2.0, 1.5])
    
    def approx_vector_cost_eval(self, init_v_c, Psi, policy, max_iters=50, tol=1e-05, logging=False):
        """
        Approximates vector cost values using Bellman iterations.
        
        :param init_v_c: Initial constraint cost values, shape (n, k)
        :param Psi: Constraint feature matrix, shape (n, k)
        :param policy: Current policy, shape (n, m)
        :param max_iters: Maximum iterations
        :param tol: Tolerance
        :param logging: Whether to log
        :return: Approximate constraint cost values, shape (n, k)
        """
        # Simple approximation: just return a perturbed version of init_v_c
        # In reality this would iterate Bellman updates
        v_c = init_v_c + np.random.randn(*init_v_c.shape) * 0.01
        return v_c

# Create environment
env = MinimalGridworld()

# xi: dual variable, shape (k,) where k=3
xi = np.array([0.5, 0.3, 0.2])

# policy: Current policy, shape (n, m) where n=4 (states), m=2 (actions)
policy = np.random.dirichlet([1.0, 1.0], size=4)  # shape (4, 2)

# eta: learning rate (scalar)
eta = 0.1

# init_v_c: Initial constraint cost values, shape (n, k) where n=4, k=3
init_v_c = np.random.randn(4, 3)

# max_iters: Maximum number of Bellman iterations
max_iters = 50

# tol: Tolerance for value approximation
tol = 1e-05

# threshold: Threshold for dual variable xi
threshold = float('inf')

ARGS = {
    'env': env,
    'xi': xi,
    'policy': policy,
    'eta': eta,
    'init_v_c': init_v_c,
    'max_iters': max_iters,
    'tol': tol,
    'threshold': threshold
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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
NPG_update Not yet run
unverified (harvested, no recorded run)
andrschl/cirl/algs/cmdp.py
code served (permissive licence) · get_code("244f37859a9fb194")
How it ran

no invocation recorded.

cmdp_gda Not yet run
unverified (harvested, no recorded run)
andrschl/cirl/algs/cmdp.py
code served (permissive licence) · get_code("39359208e0f99183")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 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 13 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

Two main challenges in Reinforcement Learning (RL) are designing appropriate reward functions and ensuring the safety of the learned policy. To address these challenges, we present a theoretical framework for Inverse Reinforcement Learning (IRL) in constrained Markov decision processes. From a convex-analytic perspective, we extend prior results on reward identifiability and generalizability to both the constrained setting and a more general class of regularizations. In particular, we show that identifiability up to potential shaping (Cao et al., 2021) is a consequence of entropy regularization and may generally no longer hold for other regularizations or in the presence of safety constraints. We also show that to ensure generalizability to new transition laws and constraints, the true reward must be identified up to a constant. Additionally, we derive a finite sample guarantee for the suboptimality of the learned rewards, and validate our results in a gridworld environment.

For agents

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

get_harvested_code_for_paper("2306.00629")
get_code_for_paper("2306.00629")
have("2306.00629")

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

curl https://syntology.ai/api/ran/2306.00629.json

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

[![Syntology run record](https://syntology.ai/api/ran/2306.00629.svg)](https://syntology.ai/paper/2306.00629)

Connect an agent — have() is free.