diff --git a/tests/test_runners.py b/tests/test_runners.py
index 8a817c1f6..7e03d7e54 100644
--- a/tests/test_runners.py
+++ b/tests/test_runners.py
@@ -108,6 +108,67 @@ def test_integrate_double_nvt(
     assert not torch.isnan(final_state.energy).any()
 
 
+def test_integrate_converts_init_kwarg(
+    ar_supercell_sim_state: SimState, lj_model: LennardJonesModel
+) -> None:
+    """integrate scales Nose-Hoover `tau` to internal units like `timestep`.
+
+    Otherwise `tau` is ~98x too small for metal units, giving an over-stiff
+    thermostat that diverges on force spikes
+
+    See https://github.com/TorchSim/torch-sim/issues/579 for more info
+    """
+    tau = 0.1  # ps, same convention as `timestep`
+    final = ts.integrate(
+        system=ar_supercell_sim_state,
+        model=lj_model,
+        integrator=ts.Integrator.nvt_nose_hoover,
+        n_steps=1,
+        temperature=100.0,
+        timestep=0.002,
+        init_kwargs={"tau": tau},
+    )
+    expected = tau * ts.units.MetalUnits.time
+    assert torch.allclose(final.chain.tau, torch.full_like(final.chain.tau, expected))
+
+
+def test_integrate_converts_step_kwarg(
+    ar_supercell_sim_state: SimState,
+    lj_model: LennardJonesModel,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """integrate divides inverse-time kwargs (e.g. Langevin `gamma`) by the unit factor.
+
+    The opposite direction from relaxation times: `gamma` is a rate, so it scales as
+    1/time, not time. It is a step kwarg, so it travels through `**integrator_kwargs`
+    to the step function (issue #579 / InverseTimeArg).
+    """
+    gamma = 10.0  # 1/ps, same time convention as `timestep`
+    # gamma is consumed inside the step function and never stored on the state, so
+    # unlike the tau test in test_integrate_converts_persistent_init_kwarg, we must
+    # spy on the step call to see the converted value.
+    received: dict[str, object] = {}
+    init, real_step = ts.integrators.INTEGRATOR_REGISTRY[ts.Integrator.nvt_langevin]
+
+    def spy_step(**kwargs: object) -> MDState:
+        received["gamma"] = kwargs["gamma"]
+        return real_step(**kwargs)
+
+    monkeypatch.setitem(
+        ts.integrators.INTEGRATOR_REGISTRY, ts.Integrator.nvt_langevin, (init, spy_step)
+    )
+    ts.integrate(
+        system=ar_supercell_sim_state,
+        model=lj_model,
+        integrator=ts.Integrator.nvt_langevin,
+        n_steps=1,
+        temperature=100.0,
+        timestep=0.002,
+        gamma=gamma,
+    )
+    assert received["gamma"] == gamma / ts.units.MetalUnits.time
+
+
 def test_integrate_double_nvt_multiple_temperatures(
     ar_double_sim_state: SimState, lj_model: LennardJonesModel
 ) -> None:
diff --git a/torch_sim/integrators/__init__.py b/torch_sim/integrators/__init__.py
index e6f09f534..3af9ac8fe 100644
--- a/torch_sim/integrators/__init__.py
+++ b/torch_sim/integrators/__init__.py
@@ -74,13 +74,25 @@
 """
 
 # ruff: noqa: F401
+import typing
 from collections.abc import Callable
+from dataclasses import dataclass
 from enum import StrEnum
-from typing import Any, Final
+from typing import Any, Final, Literal
 
 import torch_sim as ts
 
-from .md import MDState, initialize_momenta, momentum_step, position_step, velocity_verlet
+from .md import (
+    InversePressureArg,
+    InverseTimeArg,
+    MDState,
+    PressureArg,
+    TimeArg,
+    initialize_momenta,
+    momentum_step,
+    position_step,
+    velocity_verlet,
+)
 from .npt import (
     NPTLangevinAnisotropicState,
     NPTLangevinIsotropicState,
@@ -202,3 +214,52 @@ class Integrator(StrEnum):
         npt_crescale_triclinic_step,
     ),
 }
+
+
+@dataclass(frozen=True, slots=True)
+class UnitKwarg:
+    """Metadata for integrator parameters carrying physical units.
+
+    ``factor`` converts the parameter to internal units (multiply by it).
+    ``channel`` is the function the parameter is used in (either ``"init"``
+    or ``"step"``).
+    """
+
+    factor: float
+    channel: Literal["init", "step"]
+
+
+_UNIT_FACTORS: Final = frozenset(
+    arg.__metadata__[0]
+    for arg in (TimeArg, InverseTimeArg, PressureArg, InversePressureArg)
+)
+
+
+def _collect_unit_kwargs(integrator: Integrator) -> dict[str, UnitKwarg]:
+    """Read the unit-annotated kwargs off an integrator's init/step signatures."""
+    init_fn, step_fn = INTEGRATOR_REGISTRY[integrator]
+    unit_kwargs: dict[str, UnitKwarg] = {}
+    for channel, fn in (("init", init_fn), ("step", step_fn)):
+        hints = typing.get_type_hints(fn, include_extras=True)
+        for name, hint in hints.items():
+            for meta in getattr(hint, "__metadata__", ()):
+                if isinstance(meta, float) and meta in _UNIT_FACTORS:
+                    if name in unit_kwargs:
+                        raise TypeError(
+                            f"{integrator}: unit kwarg {name!r} is annotated in both "
+                            f"the init and step functions, making its `integrate` "
+                            f"channel ambiguous"
+                        )
+                    unit_kwargs[name] = UnitKwarg(float(meta), channel)
+    return unit_kwargs
+
+
+#: Unit-carrying kwargs of each integrator, derived at import time from the
+#: ``TimeArg``/``InverseTimeArg``/``PressureArg``/``InversePressureArg`` annotations
+#: on the registered init/step functions. :func:`torch_sim.runners.integrate` reads
+#: this to unit-convert kwargs.
+INTEGRATOR_UNIT_KWARGS: Final[dict[Integrator, dict[str, UnitKwarg]]] = {
+    integrator: unit_kwargs
+    for integrator in Integrator
+    if (unit_kwargs := _collect_unit_kwargs(integrator))
+}
diff --git a/torch_sim/integrators/md.py b/torch_sim/integrators/md.py
index a88685746..1b82d660e 100644
--- a/torch_sim/integrators/md.py
+++ b/torch_sim/integrators/md.py
@@ -4,6 +4,7 @@
 import warnings
 from collections.abc import Callable
 from dataclasses import dataclass
+from typing import Annotated
 
 import torch
 
@@ -17,6 +18,13 @@
 logger = logging.getLogger(__name__)
 
 
+# The metadata is the factor that converts the kwarg to internal units
+TimeArg = Annotated[float | torch.Tensor | None, MetalUnits.time]
+InverseTimeArg = Annotated[float | torch.Tensor | None, 1 / MetalUnits.time]
+PressureArg = Annotated[float | torch.Tensor, MetalUnits.pressure]
+InversePressureArg = Annotated[float | torch.Tensor | None, 1 / MetalUnits.pressure]
+
+
 @dataclass(kw_only=True)
 class MDState(SimState):
     """State information for molecular dynamics simulations.
diff --git a/torch_sim/integrators/npt.py b/torch_sim/integrators/npt.py
index 35cacd32f..f12fc39a5 100644
--- a/torch_sim/integrators/npt.py
+++ b/torch_sim/integrators/npt.py
@@ -11,9 +11,13 @@
 import torch_sim as ts
 from torch_sim._duecredit import dcite
 from torch_sim.integrators.md import (
+    InversePressureArg,
+    InverseTimeArg,
     MDState,
     NoseHooverChain,
     NoseHooverChainFns,
+    PressureArg,
+    TimeArg,
     construct_nose_hoover_chain,
     initialize_momenta,
     momentum_step,
@@ -407,9 +411,9 @@ def npt_langevin_anisotropic_init(
     *,
     kT: float | torch.Tensor,
     dt: float | torch.Tensor,
-    alpha: float | torch.Tensor | None = None,
-    cell_alpha: float | torch.Tensor | None = None,
-    b_tau: float | torch.Tensor | None = None,
+    alpha: InverseTimeArg = None,
+    cell_alpha: InverseTimeArg = None,
+    b_tau: TimeArg = None,
     **_kwargs: Any,
 ) -> NPTLangevinAnisotropicState:
     """Initialize NPT Langevin state with independent per-dimension cell lengths.
@@ -511,7 +515,7 @@ def npt_langevin_anisotropic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
+    external_pressure: PressureArg,
 ) -> NPTLangevinAnisotropicState:
     r"""Perform one NPT Langevin step with independent per-dimension cell lengths.
 
@@ -886,9 +890,9 @@ def npt_langevin_isotropic_init(
     *,
     kT: float | torch.Tensor,
     dt: float | torch.Tensor,
-    alpha: float | torch.Tensor | None = None,
-    cell_alpha: float | torch.Tensor | None = None,
-    b_tau: float | torch.Tensor | None = None,
+    alpha: InverseTimeArg = None,
+    cell_alpha: InverseTimeArg = None,
+    b_tau: TimeArg = None,
     **_kwargs: Any,
 ) -> NPTLangevinIsotropicState:
     """Initialize an NPT Langevin state using logarithmic strain coordinate.
@@ -986,7 +990,7 @@ def npt_langevin_isotropic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
+    external_pressure: PressureArg,
 ) -> NPTLangevinIsotropicState:
     r"""Perform one NPT Langevin step using logarithmic strain coordinate.
 
@@ -1643,8 +1647,8 @@ def npt_nose_hoover_isotropic_init(
     chain_length: int = 3,
     chain_steps: int = 2,
     sy_steps: int = 3,
-    t_tau: float | torch.Tensor | None = None,
-    b_tau: float | torch.Tensor | None = None,
+    t_tau: TimeArg = None,
+    b_tau: TimeArg = None,
     **kwargs: Any,
 ) -> NPTNoseHooverIsotropicState:
     """Initialize the NPT Nose-Hoover state.
@@ -1810,7 +1814,7 @@ def npt_nose_hoover_isotropic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
+    external_pressure: PressureArg,
 ) -> NPTNoseHooverIsotropicState:
     r"""Perform a complete NPT integration step with Nose-Hoover chain thermostats.
 
@@ -2515,8 +2519,8 @@ def npt_crescale_triclinic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    external_pressure: PressureArg,
+    tau: TimeArg = None,
 ) -> NPTCRescaleState:
     r"""Perform one NPT integration step with anisotropic stochastic cell rescaling.
 
@@ -2643,8 +2647,8 @@ def npt_crescale_anisotropic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    external_pressure: PressureArg,
+    tau: TimeArg = None,
 ) -> NPTCRescaleState:
     """Perform one NPT integration step with cell rescaling barostat.
 
@@ -2719,8 +2723,8 @@ def npt_crescale_triclinic_average_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    external_pressure: PressureArg,
+    tau: TimeArg = None,
 ) -> NPTCRescaleState:
     """Perform one NPT integration step with cell rescaling barostat.
 
@@ -2795,8 +2799,8 @@ def npt_crescale_isotropic_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    external_pressure: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    external_pressure: PressureArg,
+    tau: TimeArg = None,
 ) -> NPTCRescaleState:
     r"""Perform one NPT integration step with isotropic stochastic cell rescaling.
 
@@ -2902,8 +2906,8 @@ def npt_crescale_init(
     *,
     kT: float | torch.Tensor,
     dt: float | torch.Tensor,
-    tau_p: float | torch.Tensor | None = None,
-    isothermal_compressibility: float | torch.Tensor | None = None,
+    tau_p: TimeArg = None,
+    isothermal_compressibility: InversePressureArg = None,
 ) -> NPTCRescaleState:
     """Initialize the NPT cell rescaling state.
 
diff --git a/torch_sim/integrators/nvt.py b/torch_sim/integrators/nvt.py
index 644b10897..9007e63dc 100644
--- a/torch_sim/integrators/nvt.py
+++ b/torch_sim/integrators/nvt.py
@@ -8,9 +8,11 @@
 import torch_sim as ts
 from torch_sim._duecredit import dcite
 from torch_sim.integrators.md import (
+    InverseTimeArg,
     MDState,
     NoseHooverChain,
     NoseHooverChainFns,
+    TimeArg,
     construct_nose_hoover_chain,
     initialize_momenta,
     momentum_step,
@@ -143,7 +145,7 @@ def nvt_langevin_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    gamma: float | torch.Tensor | None = None,
+    gamma: InverseTimeArg = None,
 ) -> MDState:
     r"""Perform one complete Langevin dynamics integration step using the BAOAB scheme.
 
@@ -291,7 +293,7 @@ def nvt_nose_hoover_init(
     *,
     kT: float | torch.Tensor,
     dt: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    tau: TimeArg = None,
     chain_length: int = 3,
     chain_steps: int = 3,
     sy_steps: int = 3,
@@ -711,7 +713,7 @@ def nvt_vrescale_step(
     *,
     dt: float | torch.Tensor,
     kT: float | torch.Tensor,
-    tau: float | torch.Tensor | None = None,
+    tau: TimeArg = None,
 ) -> NVTVRescaleState:
     r"""Perform one complete V-Rescale (CSVR) dynamics integration step.
 
diff --git a/torch_sim/runners.py b/torch_sim/runners.py
index cd46cc7ce..2f7bb7ccd 100644
--- a/torch_sim/runners.py
+++ b/torch_sim/runners.py
@@ -18,7 +18,7 @@
 
 import torch_sim as ts
 from torch_sim.autobatching import BinningAutoBatcher, InFlightAutoBatcher
-from torch_sim.integrators import INTEGRATOR_REGISTRY, Integrator
+from torch_sim.integrators import INTEGRATOR_REGISTRY, INTEGRATOR_UNIT_KWARGS, Integrator
 from torch_sim.integrators.md import MDState
 from torch_sim.models.interface import ModelInterface
 from torch_sim.optimizers import OPTIM_REGISTRY, FireState, Optimizer, OptimState
@@ -249,7 +249,7 @@ def _write_initial_state(
             trajectory_reporter.report(state, 0, model=model)
 
 
-def integrate[T: SimState](  # noqa: C901
+def integrate[T: SimState](  # noqa: C901, PLR0915
     system: StateLike,
     model: ModelInterface,
     *,
@@ -287,7 +287,7 @@ def integrate[T: SimState](  # noqa: C901
             it's passed to `tqdm` as kwargs.
         init_kwargs (dict[str, Any], optional): Additional keyword arguments for
             integrator init function.
-        **integrator_kwargs: Additional keyword arguments for integrator init function
+        **integrator_kwargs: Additional keyword arguments for integrator step function
 
     Returns:
         T: Final state after integration
@@ -320,6 +320,16 @@ def integrate[T: SimState](  # noqa: C901
             f"integrator must be key from Integrator or a tuple of "
             f"(init_func, step_func), got {type(integrator)}"
         )
+
+    # Like `timestep` above, unit-carrying kwargs (e.g. `tau`, `gamma`,
+    # `external_pressure`) must be converted to internal units per
+    # INTEGRATOR_UNIT_KWARGS
+    init_kwargs = dict(init_kwargs or {})
+    channels = {"init": init_kwargs, "step": integrator_kwargs}
+    for key, meta in INTEGRATOR_UNIT_KWARGS.get(integrator, {}).items():
+        kwargs = channels[meta.channel]
+        if kwargs.get(key) is not None:
+            kwargs[key] = kwargs[key] * meta.factor
     # batch_iterator will be a list if autobatcher is False
     batch_iterator = _configure_batches_iterator(
         initial_state, model, autobatcher=autobatcher
@@ -351,9 +361,7 @@ def integrate[T: SimState](  # noqa: C901
         batch_kT = (
             kTs[:, system_indices] if (system_indices and len(kTs.shape) == 2) else kTs
         )
-        state = init_func(
-            state=state, model=model, kT=batch_kT[0], dt=dt, **init_kwargs or {}
-        )
+        state = init_func(state=state, model=model, kT=batch_kT[0], dt=dt, **init_kwargs)
 
         # set up trajectory reporters
         if autobatcher and trajectory_reporter is not None and og_filenames is not None:
diff --git a/torch_sim/units.py b/torch_sim/units.py
index 19cd6c4fc..146adab22 100644
--- a/torch_sim/units.py
+++ b/torch_sim/units.py
@@ -15,7 +15,7 @@ class BaseConstant(float, Enum):
 
     References:
         http://arxiv.org/pdf/1507.07956.pdf
-        https://wiki.fysik.dtu.dk/ase/_modules/ase/units.html#create_units
+        https://docs.ase-lib.org/_modules/ase/units.html#create_units
     """
 
     def __new__(cls, value: float) -> Self: