One Precision Map for Delay and Reward: Testing Whether Discounting Inherits Grid-Cell Geometry

People and animals devalue delayed rewards, and they are more patient when the stakes are large. One influential account says this is not taste but noise. Imagining a distant outcome is imprecise, so a sensible estimator shrinks the imagined value toward zero, and it spends more mental effort when more is at stake. This account fixes how imprecision grows with delay by assumption, and it lets precision for reward size vary on its own. This note proposes deriving both from one source. That source is a population code that preserves local distances up to a single position-dependent scale. Grid cells are thought to use exactly this kind of code for space, and the hexagonal lattice is its optimal discretization.

The central consequence is a constraint that can be falsified. If one local scale governs precision along both the delay and the magnitude directions, then the curve of delay-driven imprecision fixes the size of the magnitude effect. The two cannot be tuned separately. Existing models fit them separately. We can therefore ask a sharp behavioral question on public intertemporal choice data. Does a model with one shared precision map, estimated from one subset of trials, predict the magnitude effect on held-out trials as well as a model that gets an independent free function for each attribute?

Core Claim

The grid-cell literature defines the relevant code. On Conformal Isometry of Grid Cells trains position encoders so that distance between neural states tracks physical distance up to one uniform scale, a property called conformal isometry. Hexagonal grid maps emerge, reaching a gridness score of 1.70. The paper also proves that the hexagonal flat torus spreads distortion equally in all directions. Hexagons all the way down shows that seven cells with hexagonally arranged phases suffice for such a map. Both papers supply the method. Both also assume a flat space sampled uniformly.

Our claim is that a conformal map of the delay–magnitude attribute space, with a scale that varies across that space, implies a specific behavioral signature. The scale $\lambda(D,A)$, at delay $D$ and reward magnitude $A$, is the factor converting a small step in the attributes into a step in neural space. It is the same in every direction at a given point, which is what "conformal" means. If a finite, noisy population implements the map, the imprecision of a simulated outcome is roughly $\sigma(D,A) \propto 1/(\beta\,\lambda(D,A))$. Here $\beta$ is the effort or gain term from Rationally inattentive intertemporal choice.

Gershman and Bhui's shrinkage rule turns this imprecision into a discounted value. A single $\lambda$ must then explain three things at once: the shape of the discount curve, the magnitude effect, and how choice randomness varies across the attribute space.

We also predict where the scale map comes from, without fitting behavior. Following the ecological argument of Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice, we set $\lambda$ to grow with how often delays and magnitudes occur in a skewed, roughly power-law environment. Commonly encountered values get finer resolution.

Delay and dollars have different units, so an isotropic metric needs one exchange constant $c$ relating them. This is the model's single extra degree of freedom, and we state it openly.

Research Program

The work has three stages.

Stage 1: a behavioral model. We first build the "shared precision map" choice model. It is Gershman and Bhui's rational inattention model with its separate noise assumptions replaced by $\sigma(D,A)$. That quantity is computed from a parametric $\lambda$, which is either derived from ecological statistics or fitted with a small number of parameters. The model has one global effort parameter and the exchange constant $c$.

Stage 2: a trained network. We then make the map neural. We train the conformal-isometry network of On Conformal Isometry of Grid Cells and Conformal Isometry of Lie Group Representation in Recurrent Network of Grid Cells, using their distance-matching and transformation losses. Two changes adapt it to our setting. The distance target is the warped metric $\lambda$, and the input is the bounded quadrant of (delay, magnitude) rather than a periodic box.

We measure the network's decoding noise across the attribute space and check whether it reproduces the $\sigma(D,A)$ the behavioral model needs. The dependent quantities are local grid spacing, orientation, and decoding variance.

Stage 3: the geometric question. This stage connects to tiling a curved sphere. A conformal metric with varying scale describes a curved surface, and a perfect hexagonal tiling of a curved surface is impossible. Icosahedral sphere pixelizations, which Tegmark showed need about 20% fewer pixels than cube-based schemes for equal resolution, handle this with twelve five-sided defects. We therefore ask three things of the trained network. Does it tile the warped attribute space with locally hexagonal fields whose spacing shrinks where $\lambda$ is large? Does it introduce lattice defects where curvature accumulates? Or does it abandon hexagonal structure near the zero-delay boundary?

Grid Cells, Place Cells, and Geodesic Generalization for Spatial Reinforcement Learning is a constraint on this design. It shows that ignoring boundaries in a spatial basis lets learned value leak across barriers. That paper handled the problem with hand-built geodesic embeddings. We instead ask whether a learned conformal map handles the boundary at $D=0$ without such leakage.

Baselines

Naive reference points. These are the one-parameter hyperbolic model, the quasi-hyperbolic beta-delta model, and an exponential model, each with a softmax choice rule. Rats exhibit similar biases in foraging and intertemporal choice tasks found quasi-hyperbolic discounting best among these in rats. None of them has a magnitude effect unless one is added by hand. A model that cannot beat quasi-hyperbolic discounting on held-out likelihood fails.

Strongest mechanistic competitor. This is Gershman and Bhui's full rational inattention model. Its noise grows with delay by assumption, and effort is allocated by reward size. The original paper reports that it fits the Chávez et al. dataset of more than 1,200 participants better than hyperbolic models. However, no reported result was found on Lacuna for its numerical fit statistic on that dataset. We will therefore refit it ourselves with the authors' published form, using the same cross-validation folds, the same optimizer, and the same number of restarts as our model.

Strongest predictive competitor. This is Arithmetic-GPT. Its only number in the file is 70.8% variance explained on the risky-choice benchmark choices13k. No reported result was found on Lacuna for its performance on the intertemporal benchmarks gershman20 or agrawal23, so we will evaluate it on the same held-out trials, either by retraining from the published recipe or using released weights.

Key ablation. The ablation that isolates our ingredient is an anisotropic model. It has one free precision function for delay and an independent one for magnitude, with matched total parameter count. This is equivalent to a rectangular code instead of a conformal one.

What counts as winning. Beating the baselines means two things. First, held-out log-likelihood on gershman20, agrawal23, and Chávez must be no worse than the anisotropic model and better than rational inattention. Second, and more important, the conformal model must predict the magnitude effect on trials whose magnitudes it never saw. It does this by fitting $\lambda$ only on low-magnitude trials, and the anisotropic model has no way to do this at all.

Why not just…?

Why not just give rational inattention a flexible noise curve over delay? This is the most serious threat. A power-law or spline noise curve may fit discount shapes as well as ours. We concede in advance that discount shape alone cannot separate the models. The separation comes from the cross-attribute transfer test. A flexible delay-noise curve says nothing about magnitude precision, whereas a shared conformal scale does. If the anisotropic ablation transfers equally well, because both attributes happen to follow fitted curves of the same form, our structural claim fails and we will report that.

Why not just use Arithmetic-GPT, which already learns biases from ecological statistics? Its 70.8% on choices13k shows that ecological training data matters, and we adopt that idea. But Zhu, Yan, and Griffiths state that the mechanism linking skewed inputs to nonlinear value is unexplained. The model also makes no prediction about how choice randomness co-varies with magnitude and delay. If it matches our held-out likelihood, the geometric account adds explanation but not prediction. We will judge this by comparing the choice-randomness profile across the attribute space, not only average fit.

Why not just use a square grid, since delay and magnitude are orthogonal axes? A square code has different resolution along its axes and diagonals, so it implies the anisotropic precision of our ablation. The flat-space optimality results suggest a network penalized for distortion will reject it. We will test this directly. We train the network with a square-lattice constraint and compare its decoding noise, and the behavioral fit of that noise, against the unconstrained hexagonal solution.

Proposed Experiments

Experiment 1: the decisive test. We fit $\lambda$ and $c$ on choices at the lowest magnitude tercile of gershman20 and Chávez. We then predict discount rates and choice randomness at the upper terciles, comparing against the anisotropic model and refitted rational inattention.

Experiment 2: an effort manipulation. In the Ballard et al. data, requiring participants to justify their choices lowered discount rates for small rewards. Our model attributes this to the global effort term $\beta$. That predicts a proportional sharpening across the whole attribute space, including reduced randomness at long delays. An account in which effort acts only on magnitude predicts a pattern confined to that axis.

Experiment 3: an environment-only prediction. We derive $\lambda$ from assumed power-law delay and magnitude statistics, with no behavioral fitting except $c$. We report how much held-out likelihood survives relative to the fully fitted version. A large retained fraction would mean environment statistics alone largely determine discounting.

Experiment 4: the trained network. We train conformal-isometry networks on the warped quadrant and report three measurements. These are gridness and spacing as functions of $\lambda$, the count and location of lattice defects, and whether the measured decoding noise, inserted into the behavioral model, fits as well as the analytic $\sigma$.

Exploratory cross-species check. On the Kane et al. rat data, we ask whether one $\lambda$ fit to foraging travel delays predicts intertemporal choice. This would extend their shared-bias finding to a shared precision map.

What would make us drop the idea. Any of three outcomes would lead us to abandon the shared-map hypothesis. The first is failure of magnitude transfer in Experiment 1. The second is an effort effect confined to the magnitude axis in Experiment 2. The third is network noise that departs systematically from the required $\sigma$ in Experiment 4.