acceptodds
Under review as a conference paper at ICLR 2027

Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics

Abstract

A proof checker certifies that a statement is valid, not that it is worth keeping. We ask what a generator can learn about value from an incomplete record of valuable mathematics, starting from one elementary fact: a generator acts only on its information, so everything it can guarantee is decided by the worlds that its information cannot separate. First, a membership verifier localizes: a task judged in the limit is solvable with a verifier exactly when it is solvable inside each formal world, where the verifier is silent. Second, inside a formal world, a record that might be complete leaves value ambiguous. If the record lists and a larger candidate is equally consistent with it, every output in is trivia under one reading and a discovery under the other, and the two trade at an exchange rate of exactly one. If of the first points of are recorded, the best coverage of them that can be guaranteed by round is up to , for every trivia budget that grows by at most one per round. Guaranteed coverage therefore takes only two values across all trivia regimes: when trivia must stay bounded, and when it may grow without bound, however slowly, where is the relative density of in . Recovering the unrecorded region takes the inverse budget at its mass, rounds under a square-root budget. A description-length model of value realizes : bounded trivia guarantees nothing, and every unbounded budget guarantees everything.

Primary area:
learning theory
Keywords:
language generation in the limit, partial enumeration, verification, automated mathematical discovery, exploration, learning theory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YhKeOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper’s main idea is that checking correctness and recognising mathematical value are different tasks. In an abstract generation-in-the-limit model, a generator (e.g. an LLM) sees a record of valued results and can query a validity verifier such as Lean. A record cannot certify its own completeness, so the generator cannot always tell an important new result from valid but uninteresting mathematics. Guaranteeing broad discovery therefore requires continued exploration of results of uncertain value. The contributions are:

(i) a verifier identifies the valid statements in the limit but reveals nothing about value;

(ii) Theorem 4.1: if the record may be all or only part of the valuable set, guaranteed coverage of the larger set is half the recorded results plus one unrecorded result per unit of trivia allowed, tight up to lower-order terms;

(iii) hence guaranteed coverage is α/2 under bounded trivia and 1 − α/2 under unbounded trivia (α = recorded share), with the budget setting only the recovery time;

(iv) a compression-based value model in which the record has density zero.

Validation consists of proofs, a Lean formalisation of Theorem 4.1 and several other results, and simulations on integer examples. There are no experiments with LLMs or provers.

Critical strengths

Soundness. Theorem 4.1 is verified in Lean using only standard axioms, and the code compiles. I trust the Lean kernel, and the key ingredients are established in the literature, so I regard soundness as solid. Corollary 4.3 and Proposition 5.1 remain paper proofs.

Timeliness. I thank the authors for raising the problem of trivia, which has become a major concern in the community since OpenAI published openai/math.

Critical weaknesses

Contribution. Read a valid but uninteresting output as a false positive and an interesting new result as a true positive. The main result then says that guaranteeing true positives beyond the record requires accepting false positives one for one. This is a classical classifier trade-off, and the paper gives no substantially better bound:

  • the half-coverage term is from Kleinberg & Wei (2026);
  • the corollaries follow directly;
  • with every statement valid (Section 4), “trivia” is just hallucination, as in existing hallucination–coverage work (e.g. Ganju et al., 2026).

Significance and presentation. The paper explains its formal contribution more clearly than its significance; the formal results are routine derivations of the kind current LLMs produce easily. Significance is limited because the generator is given both candidate notions of value, so the hard part of discovery is assumed away. The paper is also very hard to read: coined terms and dense notation bury a simple idea.

Questions for authors

Beyond the familiar point that a system unable to tell valuable from uninteresting results must accept some uninteresting ones to find new valuable ones, what concrete lesson do your results offer people building or evaluating AI systems for mathematics? Specifically, is there a decision they would make differently because of your theorems?

4 comments
AuthorsOpenReview ↗

Response to Reviewer YhKe (1/3)

Thank you for the review and for compiling the Lean artifact. Theorem 4.1 gives an anytime frontier: one generator, fixed before the prefix length is chosen, secures the race share and the full one-for-one exchange for every prefix, from a round linear in its length on and up to a lower-order term for every sublinear budget, and no generator beats this on its own racing record. The one-for-one count it builds on is a classical two-point argument, and the bounded-trivia value α/2\alpha/2 is Kleinberg & Wei's.

W1 (classical trade-off; no better bound).

This is a classical classifier trade-off, and the paper gives no substantially better bound

What is classical. Eq. (2) is the double count: every output in K′∖KK'\setminus K is a true positive under K′K' and a false positive under KK. The inequality "discoveries beyond the record ≤\le trivia under KK" follows from it by a two-point argument. In the submission, Section 1 ("the second is new"), Section 6 and Appendix G.2 worded the novelty of the exchange term too broadly, and we have reworded them to claim only its exactness and consequences. The submission credits the term mN/2m_N/2 and the endpoint α/2\alpha/2 to Kleinberg & Wei (2026) in Section 1 and Appendix G.2. The nested pair is also the device of their sharpness proof (Theorem 3.3 of arXiv:2511.05295v1) and, as a nested chain, of Lemma 1 of Ganju et al. (2026). Section 1 now cites both at that point, Section 6 cites Kleinberg & Wei at the nested pair and Ganju et al. in its next paragraph, and Table 4 lists both.

What Theorem 4.1 adds. In words: from a round linear in NN on, a single generator is guaranteed half of the recorded results among the first NN valuable ones, plus one unrecorded result per unit of trivia allowed, up to a lower-order term for every sublinear allowance, for every NN at once. Formally, take any budget gg that grows by at most one per round. One generator GgG_g is fixed in advance. On every complete record EE of KK, for every NN and every T≥τg(mN)T\ge\tau_g(m_N) (every T≥2N+2T\ge 2N+2 when g(T)≤Tg(T)\le\sqrt T),

∣OT(Gg,E)∩K′[1…N]∣≥mN2+min⁡{g(T),nN}−1+g(τg(mN))2.\big\vert O_T(G_g,E)\cap K'[1\ldots N]\big\vert \quad\ge\quad \frac{m_N}{2}+\min\lbrace g(T),n_N\rbrace-\frac{1+g(\tau_g(m_N))}{2}.

For every generator in the class, one record holds it to mN/2+min⁡{g(T),nN}+log⁡2mN+3m_N/2+\min\lbrace g(T),n_N\rbrace+\log_2 m_N+3 for all TT and all NN with mN≥2m_N\ge2 at once. Uniformity is the point. GgG_g does not know which prefix it will be judged on, so its sweeps must interrupt the race while prefixes are still being contested. The error term is that cost, and it is lower-order for every sublinear budget. On the converse side we claim the combination of Kleinberg & Wei's racing record with the double count on one record, not either piece.

What follows. Over complete records of KK, the bounded-trivia value α/2\alpha/2 is Kleinberg & Wei's sharpness. Corollary 4.2 adds that every unbounded trivia regime, however slow, gives 1−α/21-\alpha/2, so only boundedness matters, not rate. Under a square-root budget, Corollary 4.3 pins the recovery time to (1−α−ε)2N2(1+O(N−1/2))(1-\alpha-\varepsilon)^2N^2(1+O(N^{-1/2})).

AuthorsOpenReview ↗

Response to Reviewer YhKe (2/3)

W3 (trivia as hallucination). Once the formal world is fixed, trivia is an output outside the target under the smaller reading, which is a hallucination in the sense of Ganju et al. Theorem 3.1(i) is why dropping the verifier inside a formal world loses nothing. In the submission, Sections 1 and 6, Table 4 and Appendix G.3 contrasted the two error notions. The revision removes that contrast in all four places and states the identification; Section 6 now gives the comparison below, and Table 4 summarizes it. The results differ in generators, metric and scope, so they do not conflict:

  • Ganju et al. treat randomized generators, timely density with a deadline, and every countable collection. On a constructed nested chain, a supremum hallucination rate of o(D−1(t)/t)o(D^{-1}(t)/t) forces timely density 0 (Lemma 1). Under partial enumeration, a vanishing rate attains timely density α/2\alpha/2 given a superlinear deadline (Theorem 9).
  • Kleinberg & Wei: some eventually valid generator attains density α/2\alpha/2, and on a nested pair no eventually valid generator does better (Theorem 3.3). This is the bounded-trivia value of Corollary 4.2.
  • This paper treats deterministic generators, coverage at every round and lower density of the limit set, on one nested pair, in the worst case over complete records of KK. There, unbounded trivia at any vanishing rate raises the guaranteed density from α/2\alpha/2 to 1−α/21-\alpha/2 (Corollary 4.2), and Theorem 4.1 gives the frontier for every prefix from a round linear in its length on.

The recovery time in Corollary 4.3 has the same inverse form as the D−1(t)/tD^{-1}(t)/t scale of Ganju et al. The paper does not claim a general error–time trade-off as new (Appendix G.3); what is new there is the leading constant under a square-root budget. The effect is a property of ambiguous records: a record that leaves KK refutes the smaller reading, and from then on only the race matters (Appendix C.5).

W2 (corollaries follow directly). Both follow directly from Theorem 4.1, and the paper presents them as corollaries. Corollary 4.2 says which feature of a trivia regime matters for eventual coverage: boundedness, not rate. Its bounded value α/2\alpha/2 is Kleinberg & Wei's sharpness, and the value 1−α/21-\alpha/2 for every unbounded regime is the new half. Corollary 4.3 gives the recovery time, with an exact leading constant under a square-root budget.

W4 (significance; both candidates given; presentation).

Both candidates. The converse says that even full knowledge of both candidate accounts leaves a one-for-one price. It holds in every model that contains KK and K′K', whatever other candidates the model contains, so a generator that knows less pays at least as much. Attainment says that the price can be paid without ever learning which account is right. The positive results do use known candidates. Proposition C.4 (Appendix C.6) replaces the pair by any known countable family of candidates in a known common ambient set. Its generator outputs every unrecorded valuable point with at most T0+⌊T⌋T_0+\lfloor\sqrt T\rfloor trivial rounds, where T0T_0 is finite but instance-dependent. Its coverage is at least one minus the core's density, without the race share. The exact frontier with more than two nested readings is open (Appendix I.3). Section 4 now states these conditions (paragraph "Scope").

The hard part of discovery. Selection is a separate problem (Appendix I.1): the generator covers the unrecorded region without learning whether it is valuable. The paper prices exploration without value feedback. How much feedback selection needs is open (Appendix I.3).

Proofs. The paper does not claim proof difficulty as a contribution; Section 3 calls the proofs of Theorem 3.1 simple. The contribution is the statements: an exact anytime price, where it applies, and what it implies.

Presentation. The introduction now opens with a plain-language paragraph, and Section 2 has a table that maps the model to AI-for-mathematics practice. In the main text, "ledger", "localizes" and "clock" have given way to "double count", "decides one formal world" and "recovery time". "Fiber" is kept only as the term that Section 2 defines. "Flood" and "harvest" appear only in the title and Figure 1, and the first paragraph defines them. The general response (item 3) quotes the new opening paragraph.

AuthorsOpenReview ↗

Response to Reviewer YhKe (3/3)

Q1 (a decision made differently). Within the model, yes. These are consequences of the theorems, not tested claims about systems. They bear on whoever sets the tolerance for off-record output, for instance an evaluation that counts outputs outside a curated set of known results as failures.

  1. Only whether the tolerance keeps growing matters, not how fast. Over complete records of KK, every bounded tolerance gives the same guaranteed density, α/2\alpha/2. Every unbounded one, however slow, gives 1−α/21-\alpha/2 (Corollary 4.2). A finite count and a vanishing rate of tolerated off-record outputs therefore fall on opposite sides. A tolerance that does not grow caps what any system can be guaranteed to find beyond the record, however long it runs. At a given horizon, Theorem 4.1 says exactly what the tolerance's value there buys: one unrecorded result per unit.
  2. The growth rate sets the run length. Consider a tolerance of T\sqrt T. Coming within εN\varepsilon N of the best eventual coverage of the first NN valuable results then takes (1−α−ε)2N2(1-\alpha-\varepsilon)^2N^2 rounds, up to a factor 1+O(N−1/2)1+O(N^{-1/2}) (Corollary 4.3).
  3. Where to explore. When candidate accounts of value are available, exploration buys discoveries at rate one only where the accounts disagree. A sweep of all valid statements buys one discovery per three trivia in the running example of Section 4 at α=1/2\alpha=1/2.
  4. What would lower the price. Even a perfect validity verifier changes no guarantee inside a formal world (Theorem 3.1(i)). The price measured here is that of exploring without value feedback. In the two-reading family, a single value label settles which reading holds. How many labels are needed when the record admits many readings is open (Appendix I.3).

Appendix I.2 of the submission states these points briefly; the revised introduction now states all four (paragraph "What the theorems change").

If these answers resolve the concern about the contribution, we would appreciate it if you would reconsider your rating.

Thank you for the response.

Classical trade-off. You concede that the one-for-one count is “a classical two-point argument”; exactness and uniformity add no new idea. The proof of Theorem 4.1(a) applies the cap from Kleinberg & Wei’s racing record to the recorded part of each prefix and the double count to the rest. Attainment interleaves the race with listing the known unrecorded set in order. Since those points are never observed, the two interact only through shared rounds, at a lower-order cost (for sublinear budgets) that you identify as the price of uniformity. The frontier is thus the classical trade-off plus Kleinberg & Wei on disjoint parts, and the corollaries follow directly, as you agree.

Trivia and hallucination. The verifier is idle in the main results (Theorem 3.1(i)), whose proofs, you note, do not use Section 3. With trivia identified as hallucination, Section 4 is therefore partial-enumeration generation under a hallucination budget. The title’s necessity claim, that bounded trivia cannot guarantee more than α/2 (the bounded half of Corollary 4.2), is then, by your own account, Kleinberg & Wei’s sharpness result on a nested pair. More broadly, Ganju et al.’s Lemma 1 already shows, via nested languages, that hallucination is necessary for timely dense generation.

Both responses narrow the contribution rather than distinguish it; I will raise my confidence and I am considering lowering my rating.

Reviewer_5G5HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper is a theoretical analysis about the conditions under which an algorithm can predict new mathematical statements that are "valuable" given a prior sequence of valuable statements (in this paper being valuable implies that the statement is correct). The predictor (called a generator) receives one valuable statement every time-step in a sequence of rounds, and during each round the predictor tries to generate a new valuable statement given at most finitely many queries to a correctness oracle. The paper is theoretical and proves several theorems about how many valuable statements can be predicted in this fashion.

Critical strengths

The formalism itself of a countable universe of statements with a set of true statements versus true and "valuable" statements is clear, and seems to unearth non-trivial structure upon examination. Once the idea of a generator, record, membership oracle, and coverage are inferred from the text they are relatively natural as mathematical objects.

Critical weaknesses

There are large sections of writing (including the abstract) which are frustratingly difficult to read because each individual word is ambiguous in its meaning, so by the end of the sentence one cannot solve the system of constraints needed for understanding. I tried reading the abstract and introduction 3 times before I gave up and went to the formalism in section 2.

But then in section 2 and onwards, the connection between the motivation and the formalism is unclear at every step. No special structure is given to the set of "valuable" mathematical statements besides that it is a subset of the true mathematical statements and inherits their total ordering. So with respect to the downstream theorems, we might as well be talking about "spicy" statements or "not valuable" statements. This is not by itself fatal, but each step of the paper asks the reader to absorb more symbols without a notion of why. Another example is section 4, which discusses having a nested pair of valuable-statements sets K and K' and their relationship. But why? What are we learning from this discussion?

I have been going through these symbols for hours, and I feel that I have learned nothing about how difficult it is to produce "valuable" mathematical theorems from previous ones. What does "valuable" mean after all? That would be an interesting starting point for a paper.

Questions for authors

What problem do you feel like was solved by the existence of the main theorems of this paper? If you could have one fewer theorem or lemma, would that be enough to clearly convey the goals of the paper?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5G5H (1/3)

Thank you for the careful reading. In plain terms: the generator's record might contain everything valuable or only part of a larger valuable set, and nothing the generator observes says which. Every output in the larger set that the record never contains is then a discovery if the record is partial and trivia (valid output not worth keeping) if it is complete. The paper computes exactly how many such discoveries a generator can be guaranteed for a given allowance of trivia (Theorem 4.1).

W2 (motivation and formalism).

each step of the paper asks the reader to absorb more symbols without a notion of why

We have added a table to Section 2 ("What the objects stand for") that gives the main objects their roles in AI for mathematics:

  • the generator: a system proposing statements;
  • the formal world FF: statements with machine-checkable proofs;
  • the target HH: results worth keeping;
  • the record: the results kept so far, revealed in an order chosen by an adversary, possibly in response to the generator, so that every guarantee is worst case over such orders;
  • trivia: valid output that is not worth keeping;
  • coverage: how many of the first NN valuable results the generator has output;
  • α\alpha: the share of the valuable set that the record contains;
  • the budget g(T)g(T): the trivia tolerated by round TT.

Why value has no structure. Value enters only through what the record leaves undecided. Whenever two accounts of value both fit the record, every output on which they disagree is a discovery under one and trivia under the other. A generator that knew the valuable set would output only valuable statements and need no trivia. What the theorem prices is not knowing which account is right, so the property that defines the set does not matter. Structure matters in two places. First, it decides whether two accounts can both fit the record: Section 5 (last sentence) points to value rules under which the record identifies the valuable set and no trivia is needed. Second, when two nested accounts do fit, it sets α\alpha, and Section 5 gives a description-length rule with α=0\alpha=0.

Why the nested pair K⊂K′K\subset K'. Read KK as "the record is everything valuable" and K′K' as "the record is only part of what is valuable". Every complete record of KK is legal under both readings, and no observation or verifier query separates them (Lemma 2.1). Section 4 computes the exact price of this ambiguity; this is the case it handles exactly. Section 4 now opens with this paragraph.

AuthorsOpenReview ↗

Response to Reviewer 5G5H (2/3)

Q1 (the problem solved; one fewer result).

Problem. Fix the set of valid statements, so that a validity checker gives no further information. A generator sees a record that might be everything valuable (KK) or only part of a larger valuable set (K′K'), and nothing it observes says which. How many valuable results that the record never contains can it be guaranteed, for a given allowance of trivia?

Answer. The double count itself is elementary: an output in K′∖KK'\setminus K is a discovery under K′K' and trivia under KK (Eq. (2)). The result is the exact price. Before a recorded result is output, the record may reveal it first, and Kleinberg & Wei (2026) show that only half of the recorded results can be guaranteed. One generator, fixed in advance, secures that half and, in addition, one unrecorded result per unit of trivia allowed. It does so for every prefix of the valuable set at once, from a round linear in the prefix length on (round 2N+22N+2 when the allowance is at most T\sqrt T), losing only a lower-order term for every sublinear allowance, and no generator does better (Theorem 4.1). So bounded trivia guarantees only boundedly many discoveries beyond the record. Under any unbounded allowance, however slowly it grows, one generator eventually outputs every unrecorded valuable result (proofs of Theorem 4.1(b) and Corollary 4.2) and covers a share 1−α/21-\alpha/2 (Corollary 4.2). With more than two nested accounts, the exact frontier is open (Appendix I.3).

One fewer result. We have moved one result out of the main text. The message rests on Theorem 4.1 and Corollary 4.2, and their proofs do not use Section 3 (Figure 4). We have therefore moved Corollary 3.2 (what verification resolves, with its separations) to Appendix B, where it is now Corollary B.2. Section 3 keeps Theorem 3.1, which answers a question the motivation raises: can a proof checker stand in for a judgment of value? Inside a fixed set of valid statements, a checker answers every query the same way whichever statements are valuable, so it improves no guarantee there (Theorem 3.1(i)). Per-round trivia budgets do not carry over across sets of valid statements (Appendix B.4), so Section 4 states the frontier for one fixed formal world. Moving a result alone would not have been enough: the introduction also has to state the problem first. The new first paragraph of the introduction states the problem and answer above (W1).

W1 (abstract and introduction). We have opened the introduction with the setting and the answer in plain words. The abstract keeps its statements, and we have changed it in two places: "localizes" is now "decides only the formal world", and the main result is stated as attained by one generator fixed in advance. In the main text, "ledger", "localizes" and "clock" have given way to "double count", "decides one formal world" and "recovery time". "Fiber" is kept only as the term that Section 2 defines. "Flood" and "harvest" appear only in the title and Figure 1, and the first paragraph defines them.

The new introduction begins:

A generator proposes one statement per round, and a proof checker tells it which statements are valid. It also sees a growing record of statements that mathematicians have judged worth keeping. The record might contain everything valuable or only part of a larger valuable set, and nothing the generator observes says which. Every statement it outputs that the larger set contains but the record never does is then a discovery if the record is partial and trivia, valid output not worth keeping, if the record is complete. We show that one generator can secure, among the first NN valuable statements and from a round linear in NN on, the half of the recorded ones that an adversarial record allows together with one discovery for each unit of trivia it risks, and that no generator does better.

AuthorsOpenReview ↗

Response to Reviewer 5G5H (3/3)

W3 (what "valuable" means). The paper measures the price of exploring without knowing which account of value is right. A generator that knew the valuable set would output only valuable statements. One that cannot tell the two accounts apart must risk one trivia for each discovery. The trivia is counted in the world where the record is complete, and the discovery in the world where it is not. No strategy avoids this. So when another account fits the same record, going beyond the record is never free: no generator can be guaranteed even one valuable result that the record never contains without an output that is trivia under another account fitting the same record. A theory of what makes mathematics valuable would be a different paper. This one prices the uncertainty while that question is open, and it uses only membership in a latent target set (Appendix H, Table 5). The first paragraph of the introduction now states this scope.

If these changes make the goal of the paper clear, we would appreciate it if you would reconsider your rating.

Thank you for your detailed response.

I think largely the prose style is unchanged in a way that leaves it difficult to read. The issue is a kind of non-locality. At the sentence level, this manifests as a very tangled parse tree, which is difficult to decipher while you're reading in forward order. It's difficult to say this precisely, but the following sentence is an example. “Every statement it outputs that the larger set contains but the record never does is then a discovery if the record is partial and trivia, valid output not worth keeping, if the record is complete.”

In terms of non-locality in the structure of the paper itself, there is language in the introduction which won't make sense until later in the paper. Example in the first paragraph: “We show that one generator can secure, among the first N valuable statements and from a round linear in N on, the half of the recorded ones that an adversarial record allows together with one discovery for each unit of trivia it risks, and that no generator does better.”

And the introduction, which is two pages long, asks the reader to swallow (inline) formulas for which the corresponding mathematical objects have not yet been defined.

My understanding of the main content is that you could have some set of hypotheses of what is the true collection of valuable formulas. When you see examples of valuable formulas one at a time, you can learn if a particular hypothesis is incorrect if the formula is outside that set, but in this story you do not learn anything if the formula is in a hypothesis. The goal is to guess as many valuable sentences as possible before the ground truth valuable formula generating process produces that formula. So if you have any unbounded budget for incorrect guesses of valuable formulas, then you can slowly but surely keep picking examples outside of the conservative set that's in the intersection of all the hypotheses. Since we're modeling this whole space as being countable, then you can eventually get everything that's outside of the intersection, but is within the union of hypotheses. The eventual thing we care about is the density of predictions that you make (before the ground truth process) with respect to the true valuable set.

The reason I'm saying all this is because I want to highlight that for me it still feels like there's a large disconnect between the motivation and the actual mathematical formalism that's being presented. Before I was talking about there being nothing in here that's unique to being valuable, and that was true. But if I reflect on what I have actually learned from this exercise about the difficulty of guessing valuable formulas, I still find myself at a loss.

For instance, K and K' are two hypothesis sets of possible valuable formulas, and one is a subset of the other. Suppose we are somehow in a situation where you have these sets of hypotheses and that there is such a thing as a true valuable set of valid formulas. Telling me about the limiting density of which valuable formulas you can pick in the two different cases of whether it's the bigger set or the smaller set is far removed for me from the question of how difficult in principle is it to produce valuable mathematics?

As a result of this, I will keep my score at a 2.

Reviewer_7Dk4OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies an important setting: generating from an incomplete record of valuable objects while the verifier can decide validity but provides no feedback on value. Its main result characterizes a tradeoff between coverage under a larger world and tiriva under an smaller but equally consistent world. It also derives more results about asymptotic coverage and recovery-time results. Main arguments are validated with Lean formalization.

Critical strengths
  1. The paper is well-motivated and focuses on an important problem: what can an incomplete record of valuable mathematics tell a generator about statements that have never appeared in it.
Critical weaknesses

None.

Questions for authors
  1. I'm not an expert in TCS. Could the authors provide a more intuitive explanation of what the quantitative findings add beyond the observation that exploration risks producing trivia?
2 comments

Thank you for the review. The quantitative results add three things that the observation that "exploration risks producing trivia" does not: an exact exchange rate that applies in one specific region, a threshold that depends on whether exploration stops rather than on how much of it there is, and a recovery time.

Q1 (what the quantitative findings add).

what the quantitative findings add beyond the observation that exploration risks producing trivia?

1. The risk is an identity, at rate one, in one region. Suppose the record of valuable results might be complete. Then every discovery beyond it is trivia under another account of value that agrees with everything the generator has seen. This double count is Eq. (2). Theorem 4.1 shows that no generator does better than one discovery per unit of trivia, and that a simple generator, given the two candidate accounts, achieves this rate. Only outputs in the region where the two accounts disagree count. A generator that sweeps all valid statements pays more: in the running example of Section 4 at α=1/2\alpha=1/2, two of every three trivial outputs of such a sweep are odd numbers, so it buys one discovery per three trivia.

2. For eventual coverage, only whether exploration stops matters. Let α\alpha be the recorded share of the valuable set. If trivia must stay bounded, the guaranteed share of the valuable set that the generator covers is α/2\alpha/2. If trivia may grow without bound, however slowly, it is 1−α/21-\alpha/2 (Corollary 4.2). Nothing lies in between. For this guarantee, a cap of a thousand trivia is no better than a cap of one.

3. The rate sets the recovery time. By round TT, at most g(T)g(T) unrecorded results can be guaranteed, where g(T)g(T) is the trivia allowed by then. With α=1/2\alpha=1/2 and g(T)=⌊T⌋g(T)=\lfloor\sqrt T\rfloor, the first 1000 valuable statements contain 500 recorded and 500 unrecorded ones. After 10410^4 rounds the best guarantee on them is about 250+100250+100, and it approaches 750 only after about 2.5×1052.5\times10^5 rounds (Section 4; Corollary 4.3 gives the general formula).

Finally, Section 5 shows that the extreme case arises naturally. Under the description-length rule of Section 5 (Proposition 5.1), the recorded share is α=0\alpha=0, so bounded trivia guarantees nothing.

The revised introduction now states these points in a new paragraph, "What the theorems change" (Section 1, in blue).

We are glad to answer further questions.

Thank you to the authors for your detailed response and prompt update to the submission. TBH, I had a similar experience to Reviewer 5G5H: I failed to get an intuitive overview of this paper even after repeatedly reading the introduction. However, I believe this submission is meaningful to potential readers, so I maintain my positive score.

Comments on the paper

2 comments

General Response (1/2)

We thank the three reviewers. Below are the paper's question and answer in plain words, what the theorems change within the model, and the revision. Changed text in the revised PDF is in blue.

Strengths stated in the reviews. The problem is important (Reviewer 7Dk4: "well-motivated and focuses on an important problem"; Reviewer YhKe on timeliness). Reviewer 5G5H finds the formalism itself clear and says it "seems to unearth non-trivial structure upon examination." Reviewer YhKe reports that "Theorem 4.1 is verified in Lean using only standard axioms, and the code compiles."

1. The question and the answer (Reviewer 5G5H Q1, Reviewer 5G5H W1, Reviewer 7Dk4 Q1). Fix the set of valid statements, so that a validity checker gives no further information (Theorem 3.1). A generator sees a record of valuable results. The record might be everything valuable (KK) or only part of a larger valuable set (K′K'), and nothing the generator observes says which. Every output in the larger set that the record never contains is then a discovery if the record is partial and trivia, valid output not worth keeping, if it is complete. That double count is elementary. The result is the exact price. The record can reveal a recorded result before the generator outputs it, so only half of the recorded results can be guaranteed (Kleinberg & Wei, 2026). One generator, fixed in advance, secures that half and, in addition, one unrecorded result per unit of trivia allowed. It does so for every prefix of the valuable set at once, from a round linear in the prefix length on (round 2N+22N+2 for the first NN results when the allowance is at most T\sqrt T), up to a lower-order term for every sublinear allowance, and no generator does better (Theorem 4.1). So bounded trivia guarantees only boundedly many discoveries beyond the record. Under any unbounded allowance, however slowly it grows, one generator eventually outputs every unrecorded valuable result (proofs of Theorem 4.1(b) and Corollary 4.2) and covers a share 1−α/21-\alpha/2 (Corollary 4.2). With more than two nested accounts, the exact frontier is open (Appendix I.3). A generator that knew the valuable set would need no trivia: the paper prices the uncertainty about value, not value itself.

2. What the theorems change, within the model (Reviewer YhKe Q1, Reviewer 7Dk4 Q1). These are consequences of the theorems, not tested claims about systems. They bear on whoever sets the tolerance for off-record output, for instance an evaluation that counts outputs outside a curated set of known results as failures.

  1. Only whether the tolerance keeps growing matters, not how fast. Over complete records of KK, every bounded tolerance gives the guaranteed density α/2\alpha/2, which is Kleinberg & Wei's sharpness, and every unbounded one, however slow, gives 1−α/21-\alpha/2 (Corollary 4.2). A tolerance that does not grow caps what any system can be guaranteed to find beyond the record, however long it runs.
  2. The growth rate sets the run length. Consider a tolerance of T\sqrt T. Coming within εN\varepsilon N of the best eventual coverage of the first NN valuable results then takes (1−α−ε)2N2(1-\alpha-\varepsilon)^2N^2 rounds, up to a factor 1+O(N−1/2)1+O(N^{-1/2}) (Corollary 4.3). This has the inverse form of the D−1(t)/tD^{-1}(t)/t scale of Ganju et al. (2026); the leading constant is what is new.
  3. Where to explore. When candidate accounts of value are available, exploration buys discoveries at rate one only where the accounts disagree. A sweep of all valid statements buys one discovery per three trivia in the running example of Section 4 at α=1/2\alpha=1/2.
  4. What would lower the price. Even a perfect validity verifier changes no guarantee inside a formal world (Theorem 3.1(i)). The price measured here is that of exploring without value feedback.

3. Readability and attribution (Reviewer 5G5H W1, Reviewer YhKe W4, Reviewer YhKe W1). We have opened the introduction with the question in plain words. The new introduction begins:

A generator proposes one statement per round, and a proof checker tells it which statements are valid. It also sees a growing record of statements that mathematicians have judged worth keeping. The record might contain everything valuable or only part of a larger valuable set, and nothing the generator observes says which. Every statement it outputs that the larger set contains but the record never does is then a discovery if the record is partial and trivia, valid output not worth keeping, if the record is complete. We show that one generator can secure, among the first NN valuable statements and from a round linear in NN on, the half of the recorded ones that an adversarial record allows together with one discovery for each unit of trivia it risks, and that no generator does better.

General Response (2/2)

We have also:

  • added a table to Section 2 ("What the objects stand for") that maps the main objects to their roles in AI for mathematics;
  • opened Section 4 with a paragraph on why the analysis uses a nested pair K⊂K′K\subset K';
  • replaced "ledger", "localizes" and "clock" in the main text with plain words, kept "flood" and "harvest" only in the title and Figure 1, and defined them in the first paragraph of the introduction;
  • credited Kleinberg & Wei (2026) and Ganju et al. (2026) for the nested-pair device in Section 1, Section 6 and Table 4;
  • removed the contrast between trivia and hallucination in Sections 1 and 6, Table 4 and Appendix G.3: inside one formal world, trivia is a hallucination relative to the smaller reading, and Section 6 now compares the results with Ganju et al. by generators, metric and scope;
  • reworded the novelty claims of Section 1, Section 6 and Appendix G.2 to claim the exactness of the frontier and its consequences;
  • stated in Section 4 that the positive results use known candidate accounts of value, and that with more than two nested accounts the exact frontier is open;
  • added a paragraph "What the theorems change" to the introduction, with the four points of item 2;
  • changed two phrases of the abstract: "localizes" is now "decides only the formal world", and the main result is stated as attained by one generator fixed in advance;
  • moved Corollary 3.2 to Appendix B, where it is now Corollary B.2, since the main results do not use it (Figure 4).

The theorem statements are unchanged, and the main text is within the 10-page limit. One further correction, not raised in the reviews: Section 7 and Appendix I.1 said that the generator covers the unrecorded region without learning where it is; the generator is told where the region is, and what it never learns is whether the region is valuable. Both places now say so.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.