Every one of the NeurIPS 2026 main-track papers was labelled core, adjacent or out from its title and abstract. The question asked was what the paper contributes, not which words it uses.

1. Core ( papers): the main contribution controls, identifies, estimates or certifies a dynamical system. That means a physical, engineered, biological or abstract system whose state evolves in time under inputs. Six subcategories:
   • Optimal control & MPC: optimal control, model predictive control, LQR/LQG, HJB and Pontryagin methods, trajectory optimization with dynamics constraints, differentiable MPC, stochastic and robust optimal control, learning-based MPC.
   • System identification & learned dynamics for control: identifying linear, nonlinear, hybrid or switching models from data, Koopman and operator methods (EDMD, Koopman autoencoders, dictionary discovery), equation discovery (SINDy-style), dynamic mode decomposition, data-driven (Willems-type) control, and learned dynamics or world models whose purpose is control, planning or stability analysis.
   • Stability, safety & certificates: Lyapunov, barrier and contraction certificates, control barrier functions, reachability, safety filters, formal verification of controllers and closed loops, robust stability, dissipativity, input-to-state stability, neural certificate synthesis.
   • Estimation & filtering: state estimation, Kalman and nonlinear filters, observers, smoothing, data assimilation, sensor fusion for state, learned filters.
   • Learning-based & adaptive control theory: adaptive control, online and non-stochastic control with regret bounds, control of unknown systems, sample complexity of control, control-theoretic analysis of RL, system-level synthesis, networked, distributed and multi-agent control, control of PDEs and delay systems.
   • Robotics & embodied control: locomotion, manipulation and flight where the contribution is a controller, a dynamics-aware planner, whole-body control, contact dynamics for control, or sim-to-real of controllers.
2. Adjacent ( papers): a control researcher would read it, but control is not the contribution. Five subcategories:
   • Model-based RL & world models: learned world models or dynamics for RL and planning without control-theoretic framing, Dyna-style methods, model-based offline RL.
   • Neural ODE/PDE surrogates & physics-informed dynamics: neural ODEs, SDEs and CDEs, PDE solvers and neural operators, physics-informed networks, simulation surrogates, Hamiltonian and Lagrangian networks, used for modelling rather than control.
   • Dynamical-systems view of learning: gradient-flow and ODE analyses of training, stability of optimizers, feedback or control-theoretic analyses of learning algorithms, dynamical-systems theory of networks.
   • Time series & dynamical-system reconstruction: forecasting of physical or dynamical systems with black-box models, chaotic attractor reconstruction, causal discovery in dynamical systems, tools that compare or embed dynamics.
   • Sequential decision making close to control: safe, constrained, robust or risk-sensitive RL, POMDPs and planning under uncertainty in physical domains, linear-quadratic and stochastic games, bandits with dynamics, inverse optimal control, stochastic optimal control used for sampling (Schrödinger bridges, diffusion as control), and end-to-end robot policy learning.
3. Out: everything else. These look-alikes are out unless the abstract makes a genuine control or dynamics contribution:
   • “state space models” meaning Mamba or S4 sequence models for language, vision or audio;
   • “controllable” or steerable generation, ControlNet, control tokens, LLM steering;
   • “training dynamics” or “training stability” described empirically, numerical stability, Stable Diffusion;
   • RLHF, DPO, GRPO and other RL for language-model reasoning, and LLM agents;
   • vision-language-action robot policies, and robotics papers on perception, datasets or language grounding without a controller;
   • “trajectories” of a diffusion sampler or an LLM agent, and video “world models” with no planning use;
   • adversarial robustness of classifiers, and “Kalman” or “filter” used for image or audio denoising;
   • human feedback, control groups, access, congestion, version and quality control, FDR control;
   • generic time-series benchmarks (retail, traffic counts, finance) and generic deep-RL algorithms on standard benchmarks.
4. Borderline rules:
   • A learned model is core when it is used or analysed for control, planning or stability. It is adjacent when it only predicts or simulates. Koopman, system identification and equation discovery are core either way.
   • Robot papers are core when they reason about dynamics, constraints, stability or MPC-like structure. They are adjacent when a generic RL algorithm learns a policy on robot tasks.
   • A paper is core when it controls a system. It is adjacent when the “system” is the learning algorithm itself.
   • Papers on power grids, traffic, HVAC, plasma, fluids and chemical processes are core when they control the system, adjacent when they only simulate it.
   • Mean-field control is core. Mean-field games are adjacent.
5. Subcategory: each paper gets exactly one, the best fit. Borders are fuzzy, so treat an institution's mix as indicative.
6. Rankings and counting: rankings count core and adjacent papers. A paper counts once for every institution with at least one author on it. The fractional count splits each paper equally across its authors, then across each author's institutions. The institution table can also rank by core papers only, and topic search can be limited to core papers.
7. Author name search: authors are grouped by their neurips.cc profile ID, with names and affiliations as published in the program. Search matches full or partial names, first or last names, and names typed without accents (“Anden” finds “Joakim Andén”). Each author page lists their DSC papers, core papers first, with poster session, co-authors and institutions. The author table ranks people by their number of DSC papers.
8. Institution names: affiliations come from authors' OpenReview profiles and are merged with a hand-checked table. Google includes DeepMind, Meta includes FAIR, and all Chinese Academy of Sciences institutes are merged. Unmatched affiliations are kept as written.
9. How the labels were made: a keyword and embedding filter picked candidates, and an LLM read every candidate's abstract against the written rubric. Non-candidates were checked by random sample and by a title screen, with flagged papers read in full. A second, independent LLM pass agreed on 91.5% of a 284-paper sample, and the disagreements were settled by hand. Authors' publication history was used only as a tie-breaker, never to exclude. More detail is under “How this was made”.
10. Scope: main track only. The Evaluations & Datasets, Position Paper and Journal tracks are not included.