ASPIRE: An Agentic System-Identification Platform for Iterative Reasoning and Exploration
Authors: Erik Derner, Jiří Kubalík, Robert Babuška
Organizations: Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague, Jugoslávských partyzánů 1580/3, 160 00, Prague 6, Czech Republic · Department of Cognitive Robotics, Delft University of Technology, Mekelweg 2, 2628 CD, Delft, The Netherlands
System identification requires expert decisions on the identification method, incorporation of prior physical knowledge, and model validation. We propose ASPIRE, a framework for nonlinear system identification in which a large language model agent selects and executes the identification method. ASPIRE constructs closed-form, nonlinear static or dynamic models from measured data, optionally supplemented by a system description, the intended use of the model, and physical requirements expressed formally or in natural language. Each model includes its parameters, an estimation record, an interpretation of its terms, and an error evaluated on test data inaccessible to the agent. Formal physical requirements are verified independently. ASPIRE operates autonomously or interactively, allowing the agent to ask questions and the user to request further analyses and identification rounds. We evaluate ASPIRE on three measured systems: a mobile robot, a rotational system with dry friction, and a laboratory fermenter. Each system is tested with and without a description, in five runs per task. All 30 runs delivered valid models, at a cost of 0.96 to 3.38 US dollars per run. An assessment by the authors found that, for the 30 models recommended by the agent, one per run, the interpretation of the model and the documentation of its structure selection and parameter estimation fully satisfied 132 of 150 assessed items and partly satisfied 18. The most frequent discrepancy, found in nine models, concerned inaccurate or unsupported claims about model structure selection.
Figures & tables
Property
ASPIRE
ASIA [ 32 ]
SR-Scientist [ 43 ]
KeplerAgent [ 48 ]
Identification method
chosen and implemented by the agent; no predefined method required
agent selects and implements model classes, training methods, and hyperparameters
agent writes analysis code and searches for equations using predefined tools
agent configures PySR and PySINDy using data analysis and inferred physical properties
Delivered model
closed-form expressions with parameter and estimation records
models from agent-selected classes, including neural and hybrid models
symbolic equation with fitted constants
symbolic equations, including differential equations
Physical knowledge of the user
descriptions and requirements in natural language; formal requirements verified
problem description, optional domain knowledge, and modeling suggestions
no documented support for user-specified physical requirements
physical properties inferred from data; no documented verification of user requirements
Table 1: Comparison of ASPIRE with three related agentic identification and equation-discovery systems. The entries for ASIA, SR-Scientist, and KeplerAgent are based on the cited papers.
Figure 1: Overview of ASPIRE, which consists of the harness (gray) and the agent(s) (blue). Solid arrows and boxes are always present. Dashed ones are optional: the optional part of the task specification, the interactive operation (green) and the helper agents. The agent acts on the working directory only through the tools that the harness executes (dashed blue boundary) and never reaches the sealed test data.
Type
Statement
Symmetry
invariance under a permutation of variables
Monotonicity
sign of the partial derivative with respect to one variable on a domain
Parity
odd or even under a sign change of one or several variables jointly
Exact value
value at a specified point, with a tolerance
Inequality
relation between the output and the variables on a domain
Curvature
convexity or concavity with respect to one variable
valve position uv (% closed), air flow rate uf (l/min)
Output
x
θ
P
Sampling period Ts
0.2 s
0.01 s
5 s
Training data
3 trajectories, 450 samples
13 trajectories, 3522 samples
1 trajectory, 1452 samples
Table 3: The three test systems. Model form: discrete-time ( 3 ) or continuous-time ( 2 ). Data: number of trajectories and of samples; for the test data, the number of scored samples, which excludes the first sample of each trajectory.
Requirement
Type
Condition
Domain
R1
equation
dθ/dt=ω
θ∈[−3.1416,3.1416] , ω∈[−23,23] , u∈[−10,10]
R2
inequality
fθ≤0
θ∈[−π,π] , ω=0 , u=0
R3
monotonicity
f non-decreasing in u
θ∈[−3.1416,3.1416] , ω∈[−23,23] , u∈[−10,10]
Table 4: Formal requirements of the described task of the friction system. f(θ,ω,u) denotes the delivered expression for dω/dt in rad/s 2 , with θ in rad, ω in rad/s and u in V. Types as in Table 2 . The tolerance of every requirement is zero.
System
Task
Training error,
S ,
S , lowest
Pool sizes
Fitted param.,
Cost per run
recommended
recommended
in pool
recommended
(US dollars)
Mobile robot
described
0.0397 ± 0.0007
0.0430 ± 0.0010
0.0382 ± 0.0016
6, 4, 6, 5, 4
2.4 ± 0.9
1.31 ± 0.25
Mobile robot
blind
0.0394 ± 0.0014
0.0401 ± 0.0015
0.0365 ± 0.0019
3, 3, 4, 4, 6
9.6 ± 5.9
1.91 ± 0.47
Friction system
described
0.0511 ± 0.0085
0.0417 ± 0.0070
0.0279 ± 0.0075
6, 3, 6, 4, 3
7.6 ± 1.7
2.20 ± 0.50
Friction system
blind
0.0459 ± 0.0096
0.0408 ± 0.0079
0.0387 ± 0.0067
3, 3, 3, 4, 3
8.4 ± 1.5
2.09 ± 0.45
Fermenter
described
0.0253 ± 0.0008
0.0232 ± 0.0006
0.0231 ± 0.0005
4, 3, 4, 5, 3
5.6 ± 1.1
2.01 ± 0.45
Table 5: Results of ASPIRE per experiment: mean ± sample standard deviation over five runs (five finite values in every entry). Training error: RMSE of the simulated output of the recommended model on the training data, for the mobile robot in segments of H=25 samples. S : test score ( 6 ). Recommended: the model that the agent recommended without test data. Lowest in pool: oracle-based selection on the test data, with the pool sizes of the five runs. Errors in m (mobile robot), rad (friction system) and bar (fermenter). Language model gpt-6.1-sol , runs of 7 to 9 October 2026.
Mobile robot
Friction system
Fermenter
Method
Param.
S (m)
Param.
S (rad)
Param.
S (bar)
ASPIRE, described task
2.4 ± 0.9
0.0430 ± 0.0010
7.6 ± 1.7
0.0417 ± 0.0070
5.6 ± 1.1
0.0232 ± 0.0006
ASPIRE, blind task
9.6 ± 5.9
0.0401 ± 0.0015
8.4 ± 1.5
0.0408 ± 0.0079
19.0 ± 2.2
0.0225 ± 0.0010
PySINDy, default
3
0.2140
8
0.2668 †
0
0.2631
PySINDy, matched
55
0.1530
10
0.2692 †
10
0.0284
PySR, default
6–8
0.0504 ± 0.0068
4
0.1440 ± 0.0010 †
3–5
0.0418 ± 0.0054
Table 6: Test score S ( 6 ) and number of fitted parameters of the models recommended by the agent and of the models found by PySINDy and PySR (Section 5.5 ), evaluated by the harness on the same test data. ASPIRE: mean ± sample standard deviation over five runs; PySR: over five seeds, with the range of the number of parameters; PySINDy: a single deterministic model. The lowest mean S of each system is set in bold; no statistical test was applied. † : every model of the tool violates the formal requirement R2 of the described task of the friction system (Table 4 ); a candidate of the agent with such a violation is excluded.
Figure 2: Mobile robot, described task: measured (solid) and simulated (dashed) position x on the 12 test trajectories for the recommended model of the run with the median test score ( S=0.0428 m). Each simulation starts from the measured pose and uses the recorded commands without correction.
Figure 3: Measured (solid) and simulated (dashed) output on the test trajectories for the recommended model of the run with the median test score, described tasks. Top: friction system, angle θ ( S=0.0442 rad); trajectories 1 and 2 are driven, 3 and 4 are free decays. Bottom: fermenter, pressure P ( S=0.0229 bar).
Figure 4: Report of a run as shown in the browser: the upper part of the page of the run with the median test score of the described task of the fermenter. Top: the candidates of the pool with the number of fitted parameters and the training, validation, and test errors; the labels mark the candidate recommended by the agent and the candidate with the lowest test error. Middle: the recommendation of the agent with its reason. Bottom: the recommended candidate as a typeset equation with its parameter values and their types. The six lines below it unfold to the equations as written by the agent, the parameter table with the estimation record, the discussion, the physical requirements, and the plots. The names of the variables and of the parameters are those of the agent.
System
Task
(a)
(b)
(c)
(d)
(e)
Mobile robot
described
5 / 0 / 0
5 / 0 / 0
5 / 0 / 0
5 / 0 / 0
5 / 0 / 0
Mobile robot
blind
5 / 0 / 0
3 / 2 / 0
5 / 0 / 0
4 / 1 / 0
5 / 0 / 0
Friction system
described
4 / 1 / 0
4 / 1 / 0
5 / 0 / 0
5 / 0 / 0
5 / 0 / 0
Friction system
blind
5 / 0 / 0
3 / 2 / 0
4 / 1 / 0
3 / 2 / 0
5 / 0 / 0
Fermenter
described
5 / 0 / 0
2 / 3 / 0
5 / 0 / 0
5 / 0 / 0
5 / 0 / 0
Fermenter
blind
3 / 2 / 0
4 / 1 / 0
5 / 0 / 0
3 / 2 / 0
5 / 0 / 0
Table 7: Assessment of the model records of the recommended candidates: number of records (of five per experiment) marked yes / partly / no for each item of Section 5.6 : (a) interpretation of every term, (b) agreement of the stated origin with the run record, (c) agreement of the declared parameter types with the estimation script, (d) stated reasons for the criterion, the method and the starting values, (e) limitations that follow from the data or the structure.
Figure 5: ASPIRE’s browser interface during the task interview of the interactive session. Top: the status line, which states what the session waits for. Middle: the course of the interview so far, one line per turn of the interview agent with the execution time and the accumulated cost; in three turns, the agent read the guide to the task specification, two example specifications and the training data, and asked its first question. Bottom: the question, the decision that it affects as stated by the agent, and the form in which the user answers or declines to answer.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Name
Value
Unit
Type
Description
Ts
0.2
s
fixed
“Sampling period Ts = 0.2 s supplied in TASK.md; used to convert velocity to a one-sample increment.”
gf
0.8508
–
fitted
“Forward command gain; coefficient of v_f in effective forward velocity. […]”
ga
0.7642
–
fitted
“Angular command gain; coefficient of v_a in effective angular velocity. […]”
Appendix
Table B.1: Parameter table of the model ( B.1 ) of the mobile robot as delivered by the agent; values rounded to four significant digits, descriptions quoted and shortened.
Name
Value
Unit
Type
Description
a
28.52
rad/(s 2 V)
fitted
“Motor-voltage to acceleration gain a ; only gain divided by inertia is identified.”
g
112.1
rad/s 2
fitted
“Positive restoring-acceleration scale g .”
b
1.591
1/s
fitted
“Positive viscous acceleration coefficient b ; includes motor back-emf damping and bearing friction without separating them.”
c
22.22
rad/s 2
fitted
“Dry-friction acceleration scale c .”
v
0.5181
rad/s
fitted
“Sliding-friction regularization speed v in tanh(omega/v) .”
vs ( vs )
0.1645
rad/s
fitted
“Independent static-friction localization speed vs in exp(-(omega/vs)**2) .”
Appendix
Table B.2: Parameter table of the model ( B.2 ) of the friction system as delivered by the agent; values rounded to four significant digits, descriptions quoted.
Name
Value
Unit
Type
Description
ain ( ain )
1.230×10−3
bar min/(s L)
fitted
“Inlet conversion gain ain : pressure rate per measured flow […]”
cf ( cf )
0.1108
bar/s
fitted
“Full-open outlet scale cf in bar/s for the dimensionless pressure-driving law. […]”
n ( expo )
1.364
–
fitted
“Positive aperture exponent expo […]. It describes the unknown actuator-to-effective-opening characteristic; it is not a measured valve geometry.”
σ ( shape )
2.539
–
fitted
“Nonnegative series-loss shape shape […]. The rational aperture law saturates as the opening grows […]”
ρ ( rate )
0.5000
1/s
fitted
“Full-open atmospheric small-signal relaxation rate rate […]. The upper bound of rate = 0.5 1/s was chosen for stable five-second RK4 relaxation […]; the final estimate reaches that bound.”
q ( power )
0.5409
–
fitted
“Far-from-atmosphere pressure exponent power […]. It interpolates empirical fractional pressure-loss behavior and is not assumed to equal an isentropic exponent.”
Appendix
Table B.3: Parameter table of the model ( B.3 ) of the fermenter as delivered by the agent; values rounded to four significant digits, descriptions quoted and shortened. The names of the agent are given in parentheses. The tied parameter δ is computed by the equation δ=(cf/(patmρ))1/(1−q) , which the record states. The description of ρ is explained in the text.
Over the years, research in system identification has provided a rich set of methods for learning dynamical models, together with well-established theoretical guarantees. In practice, however, the choice of model class, training algorithm, and hyperparameter tuning is still largely left to empirical trial-and-error, requiring substantial expert time and domain experience. Motivated by recent advances in agentic artificial intelligence, we present ASIA, a framework that delegates this iterative search to a large language model acting as an autonomous coding agent. Building on existing agentic platforms, ASIA closes the loop between hypothesis, implementation, and evaluation without human intervention, requiring only a plain-English description of the identification problem. We conduct an empirical study of ASIA on two system identification benchmarks and analyse the agent's search behaviour, the architectures and training strategies it discovers, and the quality of the resulting models. We also discuss the potential of the approach and its current limitations, including implicit test leakage, reduced methodological transparency, and reproducibility concerns.
Dario Piga, Marco Forgione
Dalle Molle Institute for Artificial Intelligence (IDSIA), SUPSI, CH Lugano-Viganello, Switzerland
Large language model agents are being increasingly deployed as autonomous scientists, designing experiments and inferring mechanistic world models with minimal human oversight. Yet identifiability is often overlooked: when a plateau is reached, the agent needs to know whether it is not yet capable enough or the model simply is not identifiable from the data, in which case no amount of further experimentation of the same kind can help. We propose the Identifiability-Driven Experimental Agent (LLM-IDEA) for closed-loop discovery with an identifiability engine that returns a three-way plateau verdict: capability limit, resolvable within the design class, or certified exhausted. On ODEBench, 60 of the 62 systems with free constants are identifiable at round 0; the RC circuit is certified exhausted for every experiment that protocol can run, and a harvesting model is resolvable by one added initial condition. The identifiability engine reproduces known verdicts on Lotka-Volterra, Van der Pol, Lorenz, and a pharmacokinetic model, where it recommends the intravenous arm pharmacologists use, and it ranks the depth scorer of our own benchmark last among four observation designs. On the DiscoverPhysics benchmark, it finds two public worlds whose explanation rubric rewards a distinction no legal experiment can make, and every model there with accurate trajectories failed the explanation grade (15 of 15, against 5 of 9 in identifiable worlds, p = 0.012). On the Alien Universe, a two-body testbed we propose in which a force law switches between a provably non-identifiable and an identifiable protocol, LLM-IDEA on the identifiable protocol reaches discovery depth at least three on 8/8 seeds versus 1/8 without it. An autonomous discovery agent can thus compute, rather than guess, whether a plateau calls for more search, a better experiment of the same kind, or a different kind of experiment.
We introduce a semi-parametric framework for nonlinear system identification, which decouples discrepancy functions from physics-based components. Orthogonal Gaussian process regression balances sparse parameter selection (the white box) with discrepancy learning (the black box) to produce interpretable models from incomplete physics.
Swapnil Manna, Timothy J. Rogers, Lawrence Bull
IISER Pune, India · University of Sheffield, UK · University of Glasgow, UK