The Sleeping Beauty Problem: Sleeping Kelly is a Thirder

arXiv:2510.15911 · q-fin.GN, cs.AI · Submitted 2026-08-19 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder".

Jane: The paper was written by Ben Abramowitz from.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Authors: Tom: Welcome back to the show, everyone. Today we're digging into a paper that's been making the rounds on arXiv, and it's called "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder." Jane, I have to say, the title alone got me excited — it's like a puzzle wrapped in a paradox wrapped in a betting strategy.

Jane: It really is, Tom. And for anyone just tuning in, the Sleeping Beauty problem is this classic thought experiment from philosophy. Sleeping Beauty gets put to sleep on Sunday, a coin is flipped, and if it lands heads she's woken up once on Monday, but if it lands tails she's woken up twice — Monday and Tuesday. Her memory is erased each time, so when she wakes up she doesn't know which day it is. The big question is: what probability should she assign to the coin having landed heads?

Tom: Right, and philosophers have been arguing about that for decades. Halfers say it's fifty-fifty, thirders say it's one-third heads. But this paper takes a totally different angle — it doesn't ask what she should believe, it asks what she should bet. And the author, Ben Abramowitz, shows that if Sleeping Beauty wants to grow her wealth over many repetitions of the experiment, she ends up acting like a thirder.

Jane: That's the beautiful part. Instead of just arguing about abstract beliefs, the paper sets up a concrete betting game. Sleeping Beauty can wager a fraction of her wealth before she goes to sleep, and again each time she's woken up. The question becomes: what fraction should she bet to maximize her long-term wealth?

Lu: And this is where it gets really interesting. The paper borrows from something called the Kelly criterion, which is a well-known strategy for maximizing the growth rate of your wealth when you're making repeated bets. It's not about maximizing the expected dollar amount — it's about maximizing the multiplier on your wealth over time.

Tom: Exactly, Lu. And when you apply that logic to the Sleeping Beauty problem, something remarkable happens. The optimal strategy is to bet nothing before going to sleep, but to bet one-third of your wealth on tails each time you're woken up. And that betting pattern implies she's assigning a two-thirds probability to tails and one-third to heads — which is exactly the thirder position.

Meng: So the math just falls out naturally? She's not being told to believe anything — she's just trying to be smart with her money, and the optimal strategy makes her a thirder?

Jane: Precisely, Meng. That's the core insight. The paper is saying that if you take wealth growth seriously as the goal, the thirder position isn't just a philosophical preference — it's the mathematically optimal way to behave. And that's a pretty powerful argument.

Tom: And we haven't even gotten to the Dutch book stuff yet, which is where things get really spicy. But before we go there, I want to make sure everyone understands why the Kelly criterion matters here. It's not just some obscure finance trick — it's about how you survive when you're betting repeatedly.

Jane: Right, and that's the hook for our next segment. Because the paper doesn't just stop at showing that thirders grow wealth faster — it also shows that halfers can be exploited. Stay with us.

Summary and Key Findings: Tom: So we're back with "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder," and Jane, I think we need to talk about why the expected value approach fails here, because that's where the paper really makes its case.

Jane: Absolutely. The traditional way to think about betting is to maximize expected value — you calculate the average payoff and take the bet if it's positive. But this paper shows that approach leads you straight to ruin. If Sleeping Beauty goes all-in on every bet, her expected value looks amazing — she could quadruple her money on average. But if she repeats that strategy across multiple experiments, she goes bankrupt with probability one.

Lu: That's the classic distinction between ensemble averages and time averages. The expected value calculation assumes you can play many bets in parallel and average the results. But in real life, you play bets in sequence — your wealth after one bet becomes your stake for the next. And in sequence, going all-in is a disaster because one loss wipes you out.

Meng: So the paper is saying that the standard decision theory framework — maximize expected value — is just wrong for this kind of sequential problem?

Jane: Not wrong exactly, but incomplete. The paper shows that if you instead maximize the wealth multiplier — the factor by which your money grows over many repetitions — you get a completely different answer. And that answer is the Kelly criterion strategy: bet one-third of your wealth on tails each time you're woken, and bet nothing before going to sleep.

Tom: And that strategy corresponds to assigning a two-thirds probability to tails. Which is the thirder position, just flipped around. Instead of starting with beliefs and deriving bets, you start with optimal betting and derive the beliefs.

Lu: What's really elegant is that the paper formalizes this with a theorem. It says Sleeping Beauty maximizes her wealth growth rate as a thirder who sizes bets according to the Kelly criterion. It's a clean mathematical statement, not just a philosophical argument.

Meng: But hold on — does this depend on the specific setup? Like, what if the odds aren't one-to-one, or what if she can bet on heads instead of tails?

Jane: Great question. The paper uses one-to-one odds for simplicity, but the logic generalizes. The key insight is that the optimal bet size encodes the probability — with even odds, the Kelly criterion says bet the difference between your probability and its complement. So a one-third bet on tails means she thinks tails has a two-thirds chance.

Tom: And that's the part that's going to make halfers uncomfortable. Because the paper doesn't just say thirders do better — it says halfers can be actively exploited. There's a specific Dutch book construction in the paper that shows a halfer Sleeping Beauty will accept a series of bets that guarantee she loses money no matter how the coin lands.

Meng: Wait, so it's not just that halfers grow wealth slower — they're actually leaving money on the table that someone else can take?

Jane: Exactly. The paper constructs explicit bets that a halfer would accept individually because each one looks positive, but together they guarantee a loss. And the thirder version of Sleeping Beauty is immune to that kind of exploitation.

Lu: That's a really strong result. It moves the debate from "which belief is more intuitive" to "which belief is actually defensible when you're putting your money where your mouth is."

Tom: And that's exactly where we're heading next — the Dutch book details and why thirders are invulnerable while halfers are sitting ducks. Stick around.

Improvements and Methodology: Tom: Welcome back. We're still on "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder," and now we need to get into the Dutch book argument, because that's where the paper really separates the thirders from the halfers.

Jane: Right. So a Dutch book is a set of bets that someone would accept individually — because each one looks like a good deal — but together they guarantee a loss no matter what happens. The paper shows that a thirder Sleeping Beauty can't be Dutch-booked, but a halfer can.

Meng: How does that work exactly? I mean, if she's only accepting bets that have a positive wealth multiplier, how can you force a loss?

Lu: That's the clever part. The paper defines what it means for a bet to be acceptable: the wealth multiplier has to be greater than one. For the thirder, that means before sleeping she requires the product of the heads and tails multipliers to be greater than one, and when woken she requires the heads multiplier times the tails multiplier squared to be greater than one — because she thinks tails is twice as likely as heads.

Tom: And the Dutch book conditions are the opposite — you want the wealth to shrink whether heads or tails comes up. When heads, the product of the before-sleep and waking multipliers has to be less than one. When tails, the before-sleep multiplier times the waking multiplier squared has to be less than one.

Jane: And here's the kicker — if you multiply those two Dutch book conditions together, you get that the total wealth multiplier is less than one. But the thirder's acceptance conditions, when multiplied together, guarantee it's greater than one. So the two sets of conditions are mathematically incompatible. You can't satisfy both at once.

Meng: So the thirder's beliefs are internally consistent in a way that prevents exploitation?

Lu: Exactly. The thirder's probability assignments — one-third heads, two-thirds tails — create a set of acceptance conditions that are exactly tight enough to block any Dutch book. It's like her beliefs are calibrated to the actual structure of the problem.

Tom: Now contrast that with the halfer. The halfer thinks heads and tails are equally likely when she's woken, so her acceptance condition is just the product of the multipliers being greater than one — no squared term. And that's where the vulnerability comes in.

Jane: The paper gives a concrete example. You offer the halfer a bet when she's woken that wins half her wealth if heads and loses a third if tails. Then before she sleeps, you offer a bet that doubles her wealth if tails and loses a third plus a tiny bit if heads. Each bet individually looks good to the halfer — the wealth multipliers are positive. But together, she loses money whether the coin lands heads or tails.

Meng: That's brutal. So the halfer's mistake isn't just philosophical — it's financially exploitable.

Lu: And that's the improvement this paper makes over previous work. Earlier Dutch book arguments against halfers assumed Sleeping Beauty maximizes expected dollar value, which is itself a flawed assumption. This paper shows the halfer is vulnerable even under the more defensible Kelly criterion framework.

Tom: Right — the paper is saying, look, even if you grant the halfer the most reasonable betting behavior — only accept bets that grow your wealth — she still gets exploited. The thirder doesn't. That's a huge step forward for the thirder position.

Jane: And it's not just about this one thought experiment. The methodology — using wealth multipliers and time-average growth rates instead of expected values — has implications for how we think about decision-making under uncertainty more broadly.

Meng: So what does this mean for real-world applications? Is this just philosophy, or does it actually matter outside the Sleeping Beauty problem?

Tom: That's the perfect question for our final segment. Because I think this paper has implications that go way beyond a sleeping woman and a coin flip.

Conclusion: Tom: And we're back for the final segment on "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder." Jane, I think we should wrap up by talking about why this paper matters beyond the philosophy department.

Jane: Absolutely. The paper's core result is that if you take wealth growth seriously — not expected dollar value, but actual growth over time — then the thirder position falls out naturally. Sleeping Beauty bets one-third of her wealth on tails when woken, which implies she assigns two-thirds probability to tails. And that position is immune to Dutch books.

Lu: And the halfer position isn't. That's the part that should really shake things up. The paper constructs explicit bets that a halfer would accept and that guarantee a loss. It's not just that thirders do better — halfers are actively exploitable.

Meng: So the practical takeaway is that if you're making decisions under imperfect recall — where you forget what you've already decided — you need to be really careful about how you update your beliefs. The wrong probability assignment leaves you vulnerable.

Jane: And that's where this paper connects to the real world. There are lots of situations where we make decisions without full memory of our past choices — financial decisions, medical decisions, even everyday choices. The paper suggests that the way we handle uncertainty in those situations should account for the sequential nature of wealth and decision-making.

Tom: The broader implication, I think, is that the Kelly criterion framework — maximizing growth rate rather than expected value — is more than just a betting strategy. It's a way of thinking about rational decision-making that respects the fact that our choices happen in sequence, not in parallel.

Lu: And that's a genuinely important contribution. The paper takes a decades-old philosophical puzzle and shows that the answer depends on what you're optimizing. If you optimize expected dollar value, you get nonsense — go all-in and go bankrupt. If you optimize growth rate, you get a clear, defensible answer.

Meng: And the Dutch book result gives us a way to test which position is actually rational. Thirders are invulnerable, halfers aren't. That's a concrete, testable distinction.

Tom: So to sum up — "The Sleeping Beauty Problem: Sleeping Kelly is a Thirder" argues that the thirder position isn't just a philosophical preference. It's the position that maximizes wealth growth and survives Dutch book scrutiny. The halfer position fails both tests.

Jane: And the paper does it with clean mathematics — wealth multipliers, Kelly criterion, explicit Dutch book constructions. It's a satisfying blend of philosophy, probability, and finance.

Tom: Well, I think we've given this paper a proper send-off. It's been a great discussion — from the Sleeping Beauty paradox to Kelly betting to Dutch books. Thanks to Lu and Meng for joining us, and to all our listeners for tuning in.

Jane: Next up on the show, we've got a paper on — actually, let's keep that a surprise. But trust me, it's going to be just as thought-provoking. See you all next time.

Ben Abramowitz

q-fin.GN, cs.AI

Submitted: 2026-08-19

Updated: 2026-08-20

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 49/100

The gist: “Some researchers are going to put you to sleep.

Key concepts

Sleeping Beauty Problem: A classic thought experiment where a person is put to sleep with a coin flip determining if they wake up once or twice. Her memory is erased between awakenings, creating a problem of assigning probabilities to her situation.
Kelly Criterion: A strategy used for maximizing wealth growth when making repeated bets. It focuses on maximizing the multiplier of wealth over time rather than just the average expected dollar amount.
Dutch Book: A set of bets where an individual would accept each bet individually because it looks positive, but collectively they guarantee a loss regardless of the outcome. The paper shows that thirder beliefs are immune to such exploitation.
Expected Value vs. Growth Rate: The episode contrasts maximizing expected dollar value, which can lead to ruin in sequential betting, with maximizing the wealth growth rate. The latter leads to the Kelly criterion strategy and a mathematically defensible result.

Terminology

Core — 3× in this paper · explained in 2 episodes
Adam — 1× in this paper · explained in 2 episodes

Summary

Summary

The paper addresses the Sleeping Beauty problem, a problem of imperfect recall first presented by Elga [2000]. The problem is stated as follows: “Some researchers are going to put you to sleep. During the two days that your sleep will last, they will briefly wake you up either once or twice, depending on the toss of a fair coin (Heads: once; Tails: twice). After each waking, they will put you to back to sleep with a drug that makes you forget that waking. When you are first awakened, to what degree ought you believe that the outcome of the coin toss is Heads?”

The two most common answers are the “halfer” position, assigning a probability of 1/2 to Heads, and the “thirder” position, assigning a probability of 1/3. Elga argued for the thirder position, while Lewis prominently argued for the halfer position, but debate continues. The halfer argument is Bayesian: “Initially you were certain that the coin was fair, and so initially your credence in the coin’s landing Heads was 1/2. Upon being awakened, you receive no new information (you knew all along that you would be awakened). So your credence in the coin’s landing Heads ought to remain 1/2.” The thirder argument is frequentist: “Imagine the experiment repeated many times. Then in the long run, about 1/3 of the wakings would be Heads-wakings – wakings that happen on trials in which the coin lands Heads. So on any particular waking, you should have credence 1/3 that that waking is a Heads-waking, and hence have credence 1/3 in the coin’s landing Heads on that trial.”

The paper argues that existing decision-theoretic approaches to the problem are flawed because they assume Sleeping Beauty should maximize the expected value of her bets. The core problem is that these assumptions “confuse bet payoffs with von Neumann-Morgenstern utilities.” This is illustrated by the PGM coin flip of Peters and Gell-Mann [2016]: a bet where you gain 50% on heads and lose 40% on tails has a positive expected value of 1.05x per bet, but if played sequentially with wealth reinvested, an equal number of heads and tails leads to a wealth decrease by a factor of 0.9 per pair, making it a bad bet in the long run. The expected value calculation corresponds to parallel plays, while sequential plays require a different criterion.

The paper sets up a sequential betting version of the Sleeping Beauty problem. Sleeping Beauty begins with a certain amount of money. Before being put to sleep on Sunday, she can bet any fraction of her wealth in [0,1] on the coin flip at 1:1 odds. The coin is flipped after she is put to sleep. If heads, she is woken on Monday; if tails, she is woken on Monday and Tuesday. Each time she is woken, she can wager any fraction of her money on the coin flip at 1:1 odds. She is not informed of her wealth, so wagers are fractions. Her memory is erased each time, so she has no recollection of previous wakings or wagers. The goal is to determine the optimal fraction to wager.

Let a be the fraction wagered on tails before sleep, and b be the fraction wagered on tails each time she is woken. If maximizing expected value, the paper shows the optimal strategy is a = b = 1, going all-in on every wager, yielding an expected value of 4 times original wealth. However, this strategy guarantees eventual bankruptcy if the experiment is repeated sequentially, because eventually a bet will be lost.

Instead, the paper considers maximizing the wealth multiplier for sequential repetitions of the experiment. If the experiment is run twice with one heads and one tails, the wealth multiplier is (1−a)(1−b)(1+a)(1+b)(1+b). Maximizing this gives a = 0 and b = 1/3. By the law of large numbers, as the number of experiments tends to infinity, half the flips are heads and half are tails, so the wealth multiplier per experiment is sqrt((1−a)(1−b)(1+a)(1+b)(1+b)), and the optimal strategy is (a = 0, b = 1/3). This means Sleeping Beauty should not bet before sleep, but each time she is woken she should wager 1/3 of her wealth that the coin landed tails.

The paper then infers the probability implied by this betting strategy using the Kelly criterion. With 1:1 odds, the Kelly criterion says to wager a wealth fraction b = p − (1 − p), where p is the probability that the coin landed tails. Setting b = 1/3 gives p = 2/3. Therefore, each time she is woken, Sleeping Beauty implicitly assigns probability 2/3 to tails and 1/3 to heads. Before sleep, a = 0 implies p = 1/2. The paper states: “Sleeping Kelly is a thirder.” This is formalized as Theorem 1: “Sleeping Beauty maximizes the growth rate of her wealth as a thirder who sizes bets according to the Kelly criterion.”

The paper then examines vulnerability to Dutch books. A diachronic Dutch book is “a set of bets that the agent in question would all accept individually, but that together ensure that the agent incurs a strict loss overall,” with bets offered at different times. The paper considers a setting where Sleeping Beauty is offered a bet before sleep and each time she is woken, with the same bet offered each time, and she can only accept or reject. The bet cannot depend on information unavailable to her.

The paper defines bet acceptance conditions for a thirder who only accepts bets with a wealth multiplier greater than 1. Let αOH and αOT be the fractions of wealth won from the pre-sleep bet if heads or tails occurs, and αWH and αWT be the fractions won from the waking bet if heads or tails occurs. As a thirder, before sleeping she assigns probability 1/2 to tails, so she accepts the pre-sleep bet if (1+αOH)(1/2)(1+αOT)(1/2) > 1, which simplifies to (1+αOH)(1+αOT) > 1. Upon waking, she assigns probability 2/3 to tails, so she accepts the waking bet if (1+αWH)(1/3)(1+αWT)(2/3) > 1, which simplifies to (1+αWH)(1+αWT) squared > 1.

A Dutch book requires that she loses regardless of the coin outcome: losing when heads means (1+αOH)(1+αWH) < 1, and losing when tails means (1+αOT)(1+αWT) squared < 1. Multiplying these gives (1+αOH)(1+αWH)(1+αOT)(1+αWT) squared < 1. However, by reordering terms, the acceptance conditions give (1+αOH)(1+αOT)(1+αWH)(1+αWT) squared > 1, which is a contradiction. Therefore, Theorem 2 states: “Sleeping Kelly cannot be subjected to a Dutch book if she is a thirder and only accepts bets with a wealth multiplier greater than 1.”

In contrast, the paper shows that a halfer is vulnerable. If Sleeping Beauty assigns probability 1/2 to heads and tails each time she is woken, her acceptance conditions become (1+αOH)(1+αOT) > 1 before sleep and (1+αWH)(1+αWT) > 1 upon waking. The Dutch book conditions remain the same. The paper provides an explicit example: for any suitably small ε, offer αWH = 1/2, αWT = −1/3 + ε, αOT = 1, αOH = −1/3 − ε. Before sleeping, the wealth multiplier is 4/3 − ε, and upon waking it is 1 + ε. However, if the coin lands heads, wealth shrinks by a factor of 1 − ε, and if tails, wealth shrinks by a factor of 8/9 − ε. Thus, Theorem 3 states: “Sleeping Beauty is vulnerable to Dutch books if she assigns a probability of 1/2 to the coin landing heads or tails each time she is woken, even if she only accepts bets with a wealth multiplier greater than 1.”

The paper concludes: “If Sleeping Beauty is allowed to choose bets on the outcome of the coin flip before being put to sleep and each time she is awakened, then she maximizes the growth rate of her wealth as a thirder sizing bets using the Kelly criterion under multiplicative wealth dynamics. This stands in contrast to the assumption that a bettor rationally maximizes the expected payoff of their bets. Furthermore, thirders cannot be subjected to diachronic Dutch books, while halfers remain vulnerable.”

Improvements for AI systems

Based on the paper, I can improve AI systems in the following specific ways:

1. Implement Kelly Criterion-based decision-making for AI agents in sequential decision problems with imperfect recall.

Improvement: Replace expected-value maximization with wealth-multiplier (geometric growth rate) maximization when AI agents make repeated decisions under uncertainty, especially in environments where the agent cannot remember past states or actions.
What the improved AI can do: An AI agent operating in a partially observable Markov decision process (POMDP) with memory constraints will now size its bets/actions as fractions of its current wealth (e.g., resources, energy, or capital) using the Kelly criterion, rather than risking everything on a single high-expected-value action. This prevents catastrophic ruin in long-horizon tasks.

2. Correct probability estimation in self-locating belief problems.

3. Make AI agents invulnerable to Dutch-book-style exploitation.

4. Design AI systems that maximize long-term wealth growth in repeated experiments.

5. Correctly handle asymmetric information in sequential betting with memory erasure.

6. Implement a Dutch-book-proof acceptance rule for AI agents.

7. Use the paper's proof to validate AI decision policies.

Abstract

The Sleeping Beauty problem is a problem of imperfect recall that has received considerable attention. One approach to solving the Sleeping Beauty problem is to allow Sleeping Beauty to make decisions based on her beliefs, and then characterize what it takes for her decisions to be "rational". In particular, she can be allowed to make monetary bets based on her beliefs, with the assumption that she wants to gain wealth rather than lose it. However, this approach is often coupled with the assumption that Sleeping Beauty should maximize the expected value of her bets. Here, show that Sleeping Beauty maximizes the expected growth rate of her wealth as a "thirder" sizing bets using the Kelly Criterion under multiplicative dynamics. Furthermore, this position is shown to be impervious to Dutch books. By contrast, the "halfer" position is shown to be vulnerable to Dutch books under similar circumstances.