Modern AI appears to be nothing more than a superficial statistical game. Large Language Models predict the next token from preceding words; self-supervised vision models reconstruct masked pixel patches across random photographs. Neither architecture is handed an explicit physics simulator or deductive rules of logic. On the surface, maximizing token likelihood seems completely detached from genuine understanding. Yet pushing surface likelihood to its mathematical limit forces the network to internalize the invisible structure beneath—from syntax and causal logic to 3D geometry. Crucially, this emergence relies entirely on an overwhelming abundance of observations: with trillions of tightly packed data points, competing hypotheses have virtually no room to diverge, leaving only one plausible curve standing.
In many empirical sciences, however, observational data is fundamentally sparse. In our work on Earth system science across the Qinghai-Tibetan Plateau, extreme alpine topography and harsh climates mean observational records are scarce—and long-term, high-density continuous sampling remains physically and economically unrealistic for the foreseeable future. Under such severe sparsity, data-hungry models easily mistake instrumental drift and local noise for fundamental physical law. Faced with these constraints, our approach seeks to mitigate this bottleneck through two complementary principles: leveraging mathematical isomorphisms to retrieve cross-domain evidence and refutation from data-rich proxy systems, and driving recursive self-improvement through the second axis of structural simplicity that rigorously penalizes equation complexity while expanding explanatory scope.
The Sparsity Wall: The Pitfalls of Sparse Data
In a desert of data, infinite illusions claim the crown of truth.If physical laws represent underlying reality, empirical measurements are merely low-dimensional projections cast upon our sensors. In internet text or photography, billions of dense observations tightly constrain the hypothesis space. But in data-sparse scientific disciplines, attempting to infer underlying governing equations from only a handful of scattered points is an ill-posed inverse problem: infinitely many competing, non-physical curves can pass through the exact same sparse points, trapping standard likelihood maximization in spurious regularities and hallucinations.
Earth system science on the Qinghai-Tibetan Plateau vividly illustrates this severe sparsity wall. In datasets curated by repositories like the National Tibetan Plateau Data Center (TPDC), weather and flux stations are separated by hundreds of kilometers of rugged mountain terrain (Che et al., 2019; Pan et al., 2021). Crucial nonlinear phase transitions—such as the “zero-curtain effect” where soil water and ice coexist at 0°C during spring thaw—unfold over only a couple of weeks each year. The most critical physical regimes are thus recorded by just a handful of isolated, noisy measurements.
With only a few scattered points, infinitely many mathematical curves can pass through the exact same data. Training a high-capacity neural network on ten alpine stations by maximizing likelihood alone will not discover thermodynamics. Instead, the model will effortlessly memorize sensor drift and local noise, hallucinating an arbitrary, ungeneralizable curve. In sparse regimes, surface likelihood alone is powerless to distinguish genuine physical laws from mathematical mirages.
Sampling Density and the Hypothesis Space
Dense data (Language and vision)
Sparse data (Field science)
Isomorphism: Cross-Domain Proof and Refutation
A stone from another mountain can polish jade.Faced with sparse field observations, one natural strategy is to seek structural constraints across disciplinary boundaries. The physical world exhibits profound mathematical isomorphisms: fundamentally different physical systems are frequently governed by identical partial differential equations. Regimes that are hopelessly data-starved in one domain are often abundantly instrumented and precisely measured in another.
Consider the mathematical kinship between heat diffusing through permafrost and dye dispersing through laboratory agar. In their simplest transport regimes, both follow the classical diffusion equation, ∂u/∂t = D∇2u, where u tracks temperature in frozen soil and dye concentration in a petri dish. While complete permafrost dynamics require phase-change latent heat, their transport backbone can be verified with extreme precision in a benchtop chemistry setup. A controlled lab experiment thousands of miles away can thus supply the missing empirical anchor for an inaccessible mountain range.
| Sparse field problem (Earth science) | Governing physical mechanism | Data-rich isomorphic discipline |
|---|---|---|
| Alpine airflow & snow blowing (TPCSM Drag; Pass Stations) | Boundary-layer shear and threshold particle saltation | Aerospace wind tunnels, fluidized bed reactors, desert aeolian flumes |
| Permafrost freeze-thaw cycles (Qilian / Heihe Networks) | Nonlinear diffusion with phase change (∂u/∂t = ∇·(D∇u)) | Lithium-ion battery electrodes, alloy solidification, ferromagnetic hysteresis |
| Alpine meadow patch degradation (HiWATER Ecohydrology) | Local positive feedback with critical percolation transition | Statistical physics percolation, epidemic contagion thresholds, catalytic surfaces |
Automated cross-domain retrieval transforms this mathematical isomorphism into an active discovery engine. When an AI agent formulates a hypothesis for frozen ground, it can strip away domain labels like “elevation” or “soil” to express candidate laws purely in dimensionless operators (IsoSci Benchmark, 2026; Analogical Reasoning in Science, 2026). It then queries open literature across fluid mechanics, porous electrodes, and metallurgy (CrossTrace, 2026) to find rich sister datasets. By hunting for collateral proof across disparate sciences, the agent eliminates spurious curves without waiting for years of new field expeditions.
Inside a recursive discovery loop, cross-domain evidence provides an essential filter for hypothesis validation:
- De-semanticize the Hypothesis. Stripping domain-specific terminology translates candidate mechanisms into universal dimensionless conservation laws.
- Cross-Domain Proof and Refutation. If an empirical formula fitted to sparse alpine stations fails in a fluid-mechanics lab experiment where the same dimensionless law must hold, it is instantly discarded. Conversely, agreement in a sister domain provides independent corroboration under controlled conditions.
Toward Simplicity: The Other Axis of Scientific RSI
Truth is ever to be found in simplicity, never in multiplicity.Cross-domain evidence expands the empirical frontier, but an autonomous discovery system still requires a clear optimization compass. As I emphasized at the outset, modern foundation models and conventional Recursive Self-Improvement (RSI) evaluate progress almost exclusively through maximum likelihood and its variants—relentlessly optimizing the statistical fit of observed data. In data-sparse sciences, however, chasing maximum likelihood alone is perilous: without structural constraint, an agent will simply bloat its equations with free parameters to force-fit sparse numbers.
In my view, self-improvement over scientific hypotheses demands a fundamentally dual-objective framing. When iteratively refining AI-generated hypotheses, one goal is undeniably aligned with traditional RSI—achieving greater accuracy in explaining empirical reality (maximizing likelihood and predictive fit across broader scenarios). But the second, equally vital goal is compressing that hypothesis to become progressively simpler, leaner, and more compact. Science does not merely ask how accurately an equation fits data; it asks how purely that law is expressed.
Figure 3 below illustrates this two-axis dynamic through a representative example. Horizontally, progress follows the traditional RSI criterion of empirical fit—not by driving residuals on a single alpine station from 0.05 down to 0.001 (which merely memorizes sensor noise), but by extending bounded predictions across an expanding suite of physical regimes. The objective is not zero residual on one dataset, but stable predictive validity across many.
Vertically, the second axis drives relentless symbolic pruning and structural distillation. When assimilating an unfamiliar physical regime, the agent temporarily tolerates over-parameterized transitional expressions (moving rightward). It then refactors its equation tree—drawing inspiration from symbolic distillation in LASR and recursive physical decomposition in AI Feynman—compressing tangled terms into minimal analytical expressions (moving straight up). By continually trading off empirical scope against structural simplicity, the agent climbs toward true physical invariants.
The Discovery Loop: Iterative Self-Improvement in Science
To uncover an invariant law, one must constantly dare to falsify.In our research observations, scientific discovery has never been a passive, one-shot regression problem; it is a prolonged, deliberate cycle of proposing hypotheses, confronting refutations, and revising structure. When field data is fundamentally sparse, we increasingly recognize that naive likelihood maximization quickly hits a hard ceiling. If AI is to become a genuine collaborator in vertical scientific inquiry, we believe the critical frontier lies in orchestrating an autonomous loop of recursive self-improvement (RSI) over the hypotheses themselves.
In our conceptual framework, this recursive iteration advances along the two structural axes explored above. Whenever candidate formulas are proposed, we envision guiding the system to leverage cross-domain isomorphisms—drawing benchmark observations from well-instrumented sister disciplines to promptly discard hypotheses that fail under controlled physical analogues (IsoSci; Analogical Reasoning; Verification in AI-Driven Discovery). Surviving models then advance along the simplicity axis, where symbolic refactoring relentlessly prunes equation trees down to their minimal invariant core (LASR; AI Feynman).
We further observe that the true power of an iterative discovery loop is its ability to guide active experimentation. When competing parsimonious hypotheses fit the sparse data equally well, the model can pinpoint exactly where their predictions diverge most sharply, directing scarce alpine expedition resources toward targeted, highly informative measurements. Ultimately, these ideas reflect our subjective reflections and working hypotheses from the front lines of Tibetan Plateau Earth system research. We do not expect AI to leap to universal truths overnight, but by unifying cross-domain falsification with a relentless drive toward simplicity, a self-improving scientific loop may well help us see past empirical mirages and edge closer to the underlying mechanics of nature.