Source-linked AI summary
LLMs Don't Pay for the Jump
Paras Balani, Subhrakanta Panda
TL;DR
The paper asks what enables scientific abduction when neither induction nor deduction can supply a needed new axiom. It uses Planck’s quantization to formalize thermodynamic coupling between epistemic error and physical cost, finding that fixed transformers lack this coupling while proposing coupled systems as an untested direction.
Problem
Existing accounts have not established what physical properties a bounded predictive system needs for epistemic error to become costly enough to force abandonment of a rule.
Method
The paper analyzes Planck’s blackbody-radiation crisis, distinguishes abduction from induction and deduction, and formalizes the distinction through thermodynamic coupling.
Results
Fixed-weight transformer inference lacks thermodynamic coupling: token entropy changes little as task difficulty increases, even when accuracy changes substantially across tasks and model scales.
Takeaways & Limitations
Machine abduction may require a physical mechanism that links epistemic error to computational cost strongly enough to drive rule revision.
Takeaways & Limitations
The paper does not test whether introducing thermodynamic coupling improves abductive performance under increasing causal and model complexity.
Abstract
from arXiv · showhide
Zahavy [2026] argues that Large Language Models, despite their capabilities in induction and deduction, cannot perform the abductive "Jump" that produced Einstein's equivalence principle, and attributes this limitation to the absence of embodied simulation. Zheng-Xin [2026] and Farmer [2026] question whether embodiment is necessary for abduction, pointing to alternative routes to General Relativity and forms of abduction that require no sensorimotor grounding. Max Planck resolved the blackbody radiation problem in 1900. Planck's move to E = hν required no embodied simulation. It was motivated by a mathematical consequence of classical theory, an infinite predicted energy for a finite measured quantity, that could not be physically accepted. We show that neither induction nor deduction could have produced the postulate and argue that its adoption required a coupling between epistemic error and physical cost. We formalize this distinction through thermodynamic coupling and show that fixed-weight transformer inference lacks such coupling, regardless of model scale. This is consistent with empirical results showing that output entropy remains nearly unchanged across tasks with sharply increasing causal difficulty, even as accuracy falls from 100% to 17%. We therefore argue that the missing ingredient in machine abduction may lie deeper than embodiment: a system must have some physical mechanism through which epistemic error becomes costly enough to force revision.
1 Introduction and Related Work
The paper argues that scientific abduction requires more than induction and deduction: it can arise when a theory’s physically unacceptable error forces a new rule. Using Planck’s quantization and debates over embodiment, it shifts attention toward coupling epistemic error to physical cost.
- Limits of induction and deduction: Planck’s quantization was not supplied by induction or deduction because classical premises generated the catastrophe and contained no quantization axiom.The Rayleigh–Jeans result was a deductive consequence of classical assumptions, while E = hν required leaving their logical space.
- Limits of induction and deduction: Abduction generates a case or new rule from an otherwise inexplicable result rather than guaranteeing truth or requiring repeated cases.The paper defines abduction as a third operation distinct from deduction and induction.
- Planck as a formal case study: Planck introduced E = hν after classical theory predicted unbounded radiation energy at high frequencies despite finite measured quantities.The postulate changed the formal assumptions governing energy distribution among electromagnetic modes and required no sensory or bodily experience.
- Embodiment and alternative routes: The paper challenges embodiment as a necessary condition for abduction because Planck’s hypothesis arose from a mathematical contradiction without physical interaction.Related work instead identifies alternative routes to General Relativity and forms of identity abduction that may not require sensorimotor coupling.
- Thermodynamic motivation: The paper reframes the missing property of computational abduction as physical consequences that make maintaining an erroneous representation costly enough to force revision.This shifts the focus from representational modality toward the relation between epistemic error and physical cost.
2 Background
Planck confronted a direct conflict between classical radiation theory and observed finite blackbody emission. He introduced discrete energy exchange, using Boltzmann’s statistical framework to connect the new assumption to the observed spectrum and to a broader account of costly theoretical commitments.
- The radiation crisis: Classical theory predicted unbounded high-frequency blackbody energy, contradicting reliable observations despite agreement at low frequencies.The Rayleigh–Jeans law followed from classical electrodynamics and statistical mechanics but failed in the ultraviolet regime.
- Quantization: Planck proposed energy exchanges in discrete units, with each unit given by E = hν.This changed the assumption that energy could be exchanged continuously.
- Quantization: High-frequency exchanges became less likely at fixed temperature because each required energy increased with frequency, preventing unlimited energy accumulation.The quantization rule suppressed the high-frequency excitation responsible for the Rayleigh–Jeans divergence.
- Resolution: Planck’s formula reproduced the observed blackbody spectrum.
- Boltzmann’s role: Boltzmann’s entropy framework related macroscopic entropy to the number of microscopic configurations and supplied the counting machinery for Planck’s derivation.Using this framework required a discrete microscopic picture of matter that remained controversial.
- The cost of commitment: The paper argues that defending a rejected axiom can impose professional and personal costs, distinguishing genuine abduction from an ordinary revisable guess.
- Boltzmann’s role: Planck adopted Boltzmann’s statistical counting after classical approaches failed, despite earlier reservations about its microscopic assumptions.
3 The Limits of Inductive Inference
The paper argues that induction and compression cannot explain Planck’s move because the decisive failure lay in an unobserved extrapolation rather than an existing dataset. Future confirmations of quantization were unavailable in 1900, so the new principle required treating the predicted contradiction as intolerable before direct confirmation.
- Compression and induction: Compression objectives reward rules that fit existing observations, not rules that prevent contradictions in unobserved regimes.
- Compression and induction: The Rayleigh–Jeans failure was a mathematically certain extrapolation to infinite energy, not a pattern recoverable from accumulated low- and mid-frequency data.
- Future confirmation: Experiments establishing quantization, including the photoelectric effect, Bohr’s atomic model, and Compton scattering, were unavailable in 1900.Induction can infer only from cases and results already available.
4 The Limits of Deduction
Deduction could derive the Rayleigh–Jeans catastrophe from classical premises but could not generate the new axiom E = hν. Identifying an anomaly likewise does not determine whether theory revision is warranted without a criterion for the cost of retaining the existing framework.
- Deduction after quantization: After E = hν was fixed, deriving the blackbody spectrum and recovering known radiation laws became deductive tasks.
- Deduction after quantization: Modern formal systems can derive complex results from explicitly stated premises, including through dependent type theory and reinforcement-learned systems such as AlphaProof.
- The boundary of deduction: Deduction cannot produce an axiom absent from its premises, so classical reasoning could derive the divergence but not replace the continuous-energy assumptions.
- Anomaly and revision: Detecting the Rayleigh–Jeans divergence as an anomaly does not determine whether it requires changing theory.
- Anomaly and revision: A system needs a criterion for when an error becomes costly enough to justify abandoning an existing rule.Planck’s decade of unsuccessful attempts to preserve the classical account illustrates this distinction.
5 Thermodynamic (De)coupling: Formal Framework
The paper distinguishes uncertainty that merely describes predictions from thermodynamic coupling, where epistemic error changes physical cost and can force rule revision. It argues that fixed-weight transformer inference lacks this coupling, while cost-coupled abduction requires local error-dependent pressure to generate new axioms.
- 5.1 Descriptive versus Regulatory Uncertainty: Thermodynamic coupling makes physical operation cost increase with epistemic error, creating pressure to revise an incorrect prediction or its generating rule.The paper defines epistemic error as ε = |ŷ − y*| and treats error-dependent cost as regulatory uncertainty.
- 5.1 Descriptive versus Regulatory Uncertainty: In thermodynamically decoupled systems, uncertainty describes predictive distributions without changing computational cost or subsequent dynamics based on correctness.Correct and incorrect predictions incur the same energetic cost conditional on the physical substrate.
- 5.1 Descriptive versus Regulatory Uncertainty: Landauer’s bound applies to every irreversible bit erasure regardless of correctness, so the framework distinguishes ordinary computation cost from cost dependent on epistemic quality.Each irreversible erasure dissipates at least kBT ln 2 joules.
- 5.1 Descriptive versus Regulatory Uncertainty: Planck’s infinite predicted energy could not be absorbed as a minor residual, and preserving the classical framework eventually became costlier than abandoning its assumptions.The paper presents this history as consistent with regulatory uncertainty and cost-coupled revision.
- 5.2 The Softmax–Gibbs Analogy: The softmax distribution has the same mathematical form as the Gibbs–Boltzmann distribution, with inference temperature τ formally analogous to physical temperature.This analogy motivates examining whether token entropy carries a physical cost signal.
- 5.2 The Softmax–Gibbs Analogy: For fixed weights θ and inference temperature τ, token entropy is determined by θ, τ, and context, independent of whether the context requires causal extrapolation.Theorem 1 characterizes this as softmax decoupling.
- 5.2 The Softmax–Gibbs Analogy: Transformer token entropy omits hardware temperature, dissipated power, and claim correctness, so equal entropy can accompany different predictive accuracy.The computation contains no variable representing these physical or epistemic outcomes.
- 5.3 Cost-Coupled Abduction: A Proposed Criterion: Cost-coupled abduction requires an internal cost causally linked to epistemic error on a specific anomaly until maintaining the erroneous rule becomes increasingly costly.The criterion specifies local coupling rather than a merely global uncertainty signal.
6 Empirical Signatures of Decoupling
Across models and task types, token entropy remains nearly flat as causal difficulty increases, whereas accuracy changes substantially. This pattern is consistent with the paper’s proposed entropy–correctness decoupling.
- Empirical entropy and accuracy: Across three Llama models from 3B to 70B parameters, entropy remained statistically flat across retrieval, causal reasoning, and out-of-distribution extrapolation.Within-model entropy ranges were 0.011 to 0.028 nats, with all p ≥ 0.568.
- Empirical entropy and accuracy: Accuracy ranged from 0% to 100% across the task categories despite the nearly unchanged entropy signal.The supplied passage reports this contrast across the same model comparisons.
- Empirical entropy and accuracy: The results are consistent with Theorem 1: predictive entropy changes little from familiar retrieval to difficult causal extrapolation even when accuracy changes substantially.Larger models become more confident overall without a corresponding increase in confidence–correctness alignment.
- Related empirical evidence: Other studies report that hypothesis quality degrades as world-model complexity increases, under both in-context learning and reinforcement learning from verifiable rewards.A synthesis of more than sixty studies across eleven model families identifies limited mechanistic understanding of abduction as an open problem.
7 Toward Thermodynamically Coupled Architectures
The paper compares biological prediction-error regulation with transformer inference and proposes hardware that makes inference-time updates increasingly energy-costly as errors grow. This architecture is intended to create a physical analogue of revision pressure.
- Biological comparison: Biological nervous systems provide an example in which prediction error has consequences for physical state through changes in synaptic activity, attention, and related processes.The paper connects this regulatory structure to Friston’s free-energy principle.
- Biological comparison: Thermodynamic coupling would make larger or more persistent errors alter physiological or computational resource allocation, unlike current transformer token-level uncertainty.The comparison identifies error-dependent physical consequence as the missing property without proving that biology abducts while transformers cannot.
- Proposed architecture: A proposed engineering direction uses neuromorphic hardware where prediction errors drive inference-time weight updates whose energy cost increases with error magnitude.The design corresponds to the “Intrinsic Cost” mechanism and aims to couple computational cost to epistemic error.
- Proposed architecture: The proposed update costs would create a physical analogue of the pressures associated with Boltzmann’s defense of atomism and Planck’s attempts to preserve classical theory.The comparison concerns costs of maintaining commitments rather than direct evidence that the hardware would produce abduction.
8 Limitations
The paper presents two falsifiable tests for its criterion but does not perform them. The proposed coupling could improve calibration without improving abductive hypothesis quality.
- Falsifiability: A coupled architecture should produce an internal cost or uncertainty signal that tracks accuracy as causal demand increases, unlike current transformer entropy.A flat signal despite declining accuracy would count as evidence against the criterion itself.
- Falsifiability: The paper does not test whether thermodynamic coupling reduces abductive degradation as world-model complexity increases.It warns that improved uncertainty calibration could occur without improved hypothesis quality, weakening the stronger criterion.
9 Conclusion
The paper argues that Planck’s quantization arose from an untenable physical consequence of classical theory, not embodied simulation, and formalizes this as cost-coupled abduction. It concludes that fixed-weight transformers lack this coupling, while whether coupling improves abduction remains untested.
- Planck’s route to E = hν began with a finite measured quantity that classical theory predicted to be infinite, requiring a new axiom without embodied simulation.The paper presents this as a case where neither accumulated data nor deduction supplied the postulate.
- Neither induction nor deduction could supply Planck’s postulate: available data fit without quantization, while classical premises generated the catastrophe.
- Thermodynamic coupling means that physical computational cost depends on epistemic error; fixed transformers lack this dependence in token entropy.Theorem 1 establishes decoupling for fixed transformers, and later sections examine possible coupled systems.
- Whether introducing cost coupling improves abductive performance under increasing causal and model complexity remains an open empirical question.
- The argument is limited to physical sciences, where theoretical error can have measurable physical consequences and cost has a concrete referent.The paper does not claim that the criterion applies unchanged to mathematics or computer science.
A Author’s Note
The author connects mathematical and computational training with questions about assumptions in physical theories and the generation of new axioms. Work on modified gravity and entropy-flatness results informs the paper’s synthesis.
- The author’s mathematics and computer science background made assumptions in physical theories concrete through work on modified gravity.Changing parameters of an imposed energy condition can produce different field equations and physical theories.
- Reading Zahavy’s work alongside research on thermodynamic decoupling led to the synthesis developed in this paper.
- The author’s computational background shaped the interpretation of entropy-flatness results, while modified-gravity work connected axiom generation to theoretical physics.
B Supplementary Formal Details
The supplementary details connect classical equipartition, Boltzmann’s entropy, Planck’s discrete energy postulate, and correspondence requirements to the paper’s formal decoupling condition.
- Classical Equipartition: Classical equipartition assigns each electromagnetic mode energy k_BT, while the growing high-frequency mode density yields the ultraviolet catastrophe.The mode density is 8πν^2/c^3, so unboundedly many high-frequency modes receive the same energy.
- Boltzmann’s Statistical Entropy: Boltzmann’s entropy relates macroscopic states to compatible microscopic configurations through S = k_B ln W.The framework treats matter as composed of discrete, countable microstates.
- Planck’s Quantization: Planck proposed discrete energy exchange because continuous exchange was untenable alongside the observed finite blackbody spectrum.Combined with Boltzmann’s counting, the postulate supplied the framework for quantization.
- The Correspondence Requirement: Any replacement for classical equipartition must recover the classical law where it already matched observation, namely at low frequency or high temperature.This correspondence requirement excludes replacements that resolve the divergence but fail the classical limit.
- Thermodynamic Decoupling: Thermodynamic decoupling is defined for fixed weights, fixed inference temperature, and inference-time output entropy in a bounded computational system.The formal statement uses H_t as defined in Theorem 1.