Cosmic-ray muon scattering tomography (MST) recovers the inverse radiation length λ ( r ) = 1 / X 0 ( r ) of dense or shielded objects, but classical algorithms (PoCA, MLEM) discretise the volume into a fixed voxel grid and become noise-limited at fine resolution. We introduce MINT (Muon Implicit Neural Tomography), a differentiable neural-field framework that represents λ by a coordinate-based network with multi-resolution hash encoding and optimises the per-track multiple Coulomb scattering log-likelihood end-to-end by stochastic gradient descent. Validated on two Geant4-based benchmarks spanning nuclear security and geophysical surveying, and on real detector data, MINT consistently outperforms MLEM in structural fidelity, dense-object segmentation, and background smoothness, while matching or exceeding its intensity accuracy in both the logarithmic and the linear domain. In a small-object detection study inside a full-scale ISO-like 20-foot container (twelve high- Z cubes of decreasing size, 10–1 cm, exposed in air and concealed inside 30-cm-side steel cubes), at a fixed 0.5% background-voxel exceedance fraction MINT detects 10 of 12 objects — down to a 1 cm tungsten cube hidden in steel — versus 4 of 12 for both PoCA and MLEM; and when the grid is refined at fixed track budget PoCA and MLEM lose the targets while MINT still localises them. In a mine-tunnel ore-body survey ( 1000 × 800 × 400 cm, 12.5 cm voxels, 1.6 × 1 0 6 tracks), MINT reduces the log-RMSE to 0.413 against 1.070 for MLEM and 1.895 for PoCA and the linear RMSE to 0.164 against 0.440 cm −1 ; under grid refinement at a fixed track budget its structural-fidelity advantage over MLEM widens from × 1.3 to × 10 as PoCA and MLEM degrade to noise. Finally, on real data from the INFN Padova large-volume prototype ( 1.26 × 1 0 6 cosmic-ray tracks, six material samples from Al to W), MINT with a Gaussian single- Φ likelihood locates all six samples over a markedly cleaner background than MLEM and with slightly better reconstructed scattering densities, confirming the simulation results under real detector conditions. • MINT: differentiable neural-field reconstruction of the inverse radiation length λ ( r ) from cosmic-ray muon scattering data using a multi-resolution hash-grid backbone. • Closed-form Gaussian NLL for multiple Coulomb scattering (Highland formula) cast as a differentiable line integral, optimised end-to-end by stochastic gradient descent. • Small-object detection limit in an ISO-like 20-foot container (high- Z cubes 10–1 cm, exposed and concealed inside 30-cm-side steel cubes): at a fixed 0.5% background-voxel exceedance fraction MINT detects 10 / 12 objects — down to a 1 cm tungsten cube concealed in steel — against 4 / 12 for both PoCA and MLEM; refining the grid at fixed track budget degrades PoCA/MLEM to noise while MINT still localises the targets. • Geant4 mine-tunnel ore-body slice (12.5 cm voxels): MINT gives the lowest log-RMSE (0.413 vs. 1.070 for MLEM) and, under grid refinement at fixed track budget, retains structural fidelity (Spearman 0.43 at 3.125 cm) while PoCA and MLEM degrade to noise. • Experimental validation on real data from the INFN Padova large-volume prototype (two CMS drift chambers, six samples from Al to W, 1.26 × 1 0 6 tracks): MINT with a Gaussian single- Φ likelihood locates all six samples over a markedly cleaner background than MLEM and with slightly better reconstructed scattering densities.