Machine Learning · arXiv:2610.01751 v1 · submitted 1 Oct 2026 · 19 references
Evidence-Gated Research: Statistically Controlled Model Adoption in Adaptive Search
The paper says
“Adaptive model search is path dependent: once a challenger is adopted, it becomes the reference from which later candidates are generated. A statistically unsupported replacement can therefore alter hypotheses that have not yet been proposed.”
more of the abstract
We introduce Evidence-Gated Research (EGR), a statistical adoption layer for moving-incumbent search. EGR freezes each challenger before decision evidence is revealed, builds anytime-valid evidence across a predeclared set of environments, routes evidence predictably toward unresolved components, composes a persistent candidate e-value, and passes that e-value to an online controller. Under explicit conditional-validity and predictability conditions, the resulting procedure controls false discovery rate for the declared all-environment adoption target even though earlier adoptions change later challengers. In a 5,000-trajectory closed-loop benchmark, development-only e-LOND attains persistent FDR 0.621, whereas no persistent false-adoption path is observed for the audited EGR variants in that finite run. In matched replay over 600 challenger–incumbent pairs, Stagewise EGR preserves fixed-anytime alternative crossing decisions while using 56.1% less decision evidence at the representative threshold. A three-environment public-data study and a 40,000-sample controlled neural benchmark reproduce the evidence-efficiency pattern. These results identify model replacement as a distinct statistical control point in adaptive model development.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 2 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A genuinely clean anytime-valid adoption layer for moving-incumbent search, with sound conditional validity proofs; the honest gap is that every experiment constructs an evidence unit that satisfies Assumption 1 by design. This line is wrong
2 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 19 references say.
The paper says, at Sec. 4.1 (Assumption 1, Eq. 5) and Sec. 5.3 (b76)
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.80):
“Larger prospectively sampled image, language-model, and fully agentic search loops remain important tests of external scope.”
Pith says: Objection 1 of 2 · major
The sequentially valid evidence unit is the single load-bearing condition of the framework, and the paper's own diagnostic sets the stakes: treating correlated raw rows as fresh sequential units crosses the persistent null with probability .728 at rho=.95, versus .0127 for a block-aggregated unit.
more of objection 1
Yet in every experiment the unit is constructed so that Assumption 1 holds by design — synthetic paired draws are independent, and public-data decision evidence is resampled with replacement from fixed empirical pools (b213), making draws i.i.d. conditional on the pool. No experiment exercises the motivating setting of real, prospectively sampled evaluation data with unknown correlation, where the block/cluster unit must be identified rather than stipulated, and where finding it by outcome-dependent peeking would violate the freeze/restart contract in Sec. 4.3. Because the authors state this scope honestly (b111, b113–115), this is an unresolved hinge between theorem and practice rather than a concealed flaw; the revision should either validate a valid-unit construction on data with unknown correlation, provide a non-outcome-dependent procedure for predeclaring the unit, or rescope the headline claim to settings where a predeclared valid unit is established.
The paper says, at Table 4 and Sec. C.2
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The power-advantage claims for EGR-ALLOC over fixed IUT are not supported by the reported Monte Carlo intervals. For the three-environment digits study, fixed IUT power is .3137 [.2981, .3293] while EGR-ALLOC is .3349 [.3185, .3514], and these intervals overlap; the same holds for Stagewise EGR (.3221 [.3064, .3379]).
more of objection 2
In the controlled 40k benchmark the intervals (.2149 [.1886, .2413] versus .2666 [.2414, .2919]) touch at the boundary. The text asserts that EGR-ALLOC 'improves power to .3349 and .2666' without flagging this overlap, and Table 4 displays only the point estimates. The evidence-efficiency claims are unaffected — they are tight by construction in the replay design — but the power claims should either show the intervals at the claim site or be softened to reflect an unconfirmed power gain within Monte Carlo error.
4 minor objections
The paper says, at Sec. 4.3 / A.5, Theorem 2
Pith says: minor objection 1 of 4
The phrase 'under the dependence conditions of Xu and Ramdas (2024)' is never unpacked; state that the inherited property is conditional e-value superuniformity, E[E_pers_t | F_{t-1}] <= 1 under the persistent null, which is exactly what Section A.5 supplies, so it is transparent to the reader which property is handed to e-LOND.
The paper says, at Sec. 3, Eq. (2)
Pith says: minor objection 2 of 4
The paired utility gain requires a pairing of Z for challenger and incumbent; make explicit the assumed joint law (e.g., common uniform seeds or a fixed reference assignment) that defines the pairing and is implemented in the experiments, since Delta^e_t is not well defined without it.
show 2 more
The paper says, at Sec. 4.1, Eq. (6)
Pith says: minor objection 3 of 4
Clarify the relationship between the bounding scale B^e (which makes Y in [-1,1]) and the per-environment caps used in the experiments (B=60/120/400/800); the paper says 'Choose B^e so that Y in [-1,1]' but does not state whether the experiments set B^e equal to the cap or to a separately computed bound.
The paper says, at Table 2, cross-environment independence row
Pith says: minor objection 4 of 4
The cell 'Minimum composition is pointwise' is terse to the point of obscuring the argument; expand the 'Role' cell to say that the minimum is pointwise bounded by the first true component null via the Lemma 1 partition on F_{t-1}-measurable events.
What would settle it
Reproduce the clustered-dependence diagnostic (b76) on real correlated validation data: use a prospectively-sampled stream with genuine temporal or panel correlation (video validation frames, or a dataset with known subject/session grouping), plant challengers exactly on the audit boundary so every candidate is a true persistent null, and run the EGR evidence machinery with (i) raw rows and (ii) a pre-declared block unit whose cluster boundaries are fixed from metadata only — session ID, subject ID, timestamp window — before any outcome is opened. Compare persistent-null crossing rates at q=20. If the metadata-only blocking holds the crossing rate near nominal (analogous to .0127) while raw rows inflate (analogous to .728), Assumption 1 transfers and Theorem 2 holds in that real setting.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 2 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Read it. The core idea is real and the theory is in better shape than the abstract suggests. The new thing is the framing: an unsupported adoption changes the reference model and can bias every proposal that comes after, so model replacement becomes a statistical control point. Nobody in the e-value literature treats that directly. The persistent candidate e-value as a pointwise min over component e-processes (Lemma 1) is clean and genuinely independence-free — the partition-by-first-true-null argument needs no cross-environment independence at all. The Appendix A proofs hold up; I checked the predictable-routing step and the measurability of the frozen challenger. I don't share the reader's secondary worry about Theorem 2: the composition delivers exactly the conditional e-value condition e-LOND consumes, and nu_t is Ft−1-measurable because the challenger is frozen before decision evidence. The moving-incumbent dependence is absorbed, not asserted away.
The real hinge, and the only one I'd spend a referee's attention on, is Assumption 1. Validity lives or dies on what counts as the sequential evidence unit. The clustered-observation diagnostic (.728 vs .0127 null-crossing at rho=.95) is the sharpest object in the paper, and it makes the point that a marginal mean condition on correlated rows is not enough. But in every experiment the unit is constructed so Assumption 1 holds by design: synthetic paired draws are independent, and public-data decision evidence is resampled with replacement from fixed empirical pools (b213). No experiment uses real, prospectively-sampled, correlated evaluation data where the block or cluster unit has to be found rather than assumed. The authors state this scope honestly (b89, b113, b114), so it is an unresolved hinge between theorem and practice, not a concealed flaw. As written, the guarantee holds relative to what the user can predeclare as a valid unit.
The empirical discipline is decent: finite-run zeros are flagged as descriptive (b63), the 0/5000 headline is one finite run with a ~6e-4 one-sided upper bound, and the leak+reuse versus leak+restart stress test isolates candidate-shaping evidence as the cause of inflation rather than adaptivity — that is a well-designed experiment. Mechanical gripe: the reproducibility statement promises runnable code and SHA-256 checksums but gives no public link or commit hash in the text; a referee should ask for it.
Bottom line: sound within its stated assumptions, honest about scope, one genuinely unresolved hinge between the theory and real deployment data. It deserves a serious review. Send it out and ask the author to either demonstrate Assumption 1 on real correlated data or state plainly that the guarantee applies to evidence units the user can predeclare.
The strongest objection
The paper's clustered-observation diagnostic (Sec. 5.3, b76) is its sharpest object: treating correlated raw rows as fresh sequential evidence units crosses the persistent null at .728, while an aggregated block unit drops that to .0127. This gap is a direct statement about Assumption 1 (Eq. 5) — validity depends entirely on what the 'actual unit consumed by the sequential update' is. Yet in every experiment the unit is constructed so Assumption 1 holds by design: synthetic paired Bernoulli draws are independent, and public-data decision evidence is resampled with replacement from fixed empirical pools (b213), so draws are i.i.d. given the pool. No experiment uses real, prospectively-sampled evaluation data where the correlation structure is unknown and the block/cluster unit must be found, not assumed; the stress test knows its cluster boundaries and rho. If a valid predeclared evidence unit does not exist for such streams, or can only be discovered by outcome-dependent peeking (which the freeze/restart contract forbids), then by the paper's own .728-versus-.0127 logic the conditional persistent-FDR guarantee of Theorem 2 fails to transfer to the setting the paper motivates. The authors state the scope honestly (b89, b111, b113, b114, b213), so this is the unresolved hinge between theorem and practice, not a concealed flaw. I do not share the reader's secondary worry literally: the across-candidate composition is cleaner than 'asserted, not verified' — Theorem 2 / Appendix A.5 delivers E[nu_t E^pers_t | F_{t-1}] <= nu_t, exactly the conditional-e-value condition e-LOND consumes, and nu_t is F_{t-1}-measurable because the challenger is frozen before evidence, absorbing the moving-incumbent dependence. The theory is in good shape; Assumption 1's real-data instantiation is the weak link.
The reader's weakest_assumption is the same as mine: Assumption 1 (Eq. 5) holding for the actual evidence unit on real correlated data. I agree. My added value is narrowing why it is load-bearing and pruning one of the reader's mechanical reasons. On the reader's reason that Theorem 2's across-candidate control is 'asserted but not verified' under e-LOND's dependence conditions: I inspected the composition and it is sound. Lemma 1 (Appendix A.2) proves E[nu_t E^pers_t | F_{t-1}] <= nu_t using disjoint F_{t-1}-measurable events selecting the first true component null; this is exactly the conditional-e-value property e-LOND consumes, and nu_t being F_{t-1}-measurable (challenger frozen before evidence, Section 4.3) absorbs the moving-incumbent history dependence. So the across-candidate composition is not the weak point; the premise is. Assumption 1 is instantiated in all experiments only by independent draws or by resampling with replacement from fixed pools (b213), which makes unit validity trivial. The clustered diagnostic (b76) shows how catastrophic unit misspecification is (.728 vs .0127) on a KNOWN correlation structure. On real deployment streams the structure is unknown; a valid predeclared block may not exist, and deriving blocks from outcomes would violate the paper's own freeze/restart contract (Section 4.3). The authors state this scope honestly (b89, b111, b113, b114, b213): I am not flagging a concealed flaw, I am identifying the unresolved hinge between the theorem and its motivating application. I also verified the headline 0/5000 figure is internally consistent (one-sided 95% binomial upper bound approximately 6e-4) and that the matched-replay evidence-efficiency claim (56.1%) is well-controlled by construction — identical streams and identical crossing events, differing only in evidence placement. The reproducibility statement names bundle contents and SHA-256 checksums but gives no public link or commit hash in the reviewed text, which, together with the unresolved real-correlation question, supports keeping the reader's CONDITIONAL verdict. My stress test confirms the reader's weakest-assumption identification and does not move the verdict.
The referee's report
If Theorem 2 holds, EGR identifies model replacement as a distinct statistical control point in adaptive search, separating flexible proposal generation from adoption validity — a framing that goes beyond fixed-family comparisons and beyond candidate-level anytime validity (b68). The conditional-validity machinery is the paper's finest asset: Lemma 1's partition-by-first-true-null argument is pointwise and requires no cross-environment independence, and the Appendix A proofs are complete and checkable. The paper is exemplary on reproducibility (runnable code, run manifests, SHA-256 checksums) and on honest accounting of what the finite-run numbers do not establish (b63). The evidence-efficiency claims are cleanly isolated by construction in the matched replay (identical crossing events across acquisition rules by design), and the b76 diagnostic honestly quantifies the cost of violating Assumption 1.
Pith's reading of the paper (a machine's summary, not the authors')
Adaptive model search is path-dependent: every adopted challenger becomes the reference from which the next candidate is generated, so a statistically unsupported replacement silently biases hypotheses that have not yet been proposed. This paper claims that this state-changing adoption can be made a statistically controlled decision point. The Evidence-Gated Research (EGR) layer freezes each challenger before decision evidence is opened, builds anytime-valid evidence across a predeclared set of environments, combines it into a persistent candidate e-value, and passes that e-value to an online controller. Under explicit conditional-validity and predictability conditions, the persistent-target false discovery rate stays controlled even though earlier adoptions change later challengers. The empirical pattern across closed-loop, replayed, multi-environment, and neural benchmarks is that target mismatch, candidate-time leakage, or missing stream correction drives unsupported adoptions, while predictable evidence placement cuts validation cost substantially.
What carries the argument. The carrying object is the persistent candidate e-value, $E^{\mathrm{pers}}_t = \min_{e\in\mathcal{E}} E^e_t$, a minimum over environment-specific betting e-processes. Each component $E^e_t$ is a fixed mixture of nonnegative supermartingale products built on precommitted betting fractions, and its validity rests on Assumption 1: whenever a component null is true, the acquisition mechanism yields bounded selected updates with non-positive conditional mean, for the actual evidence unit consumed by the sequential update rather than for marginally valid raw rows. Because the persistent null is a union over environments, the minimum is pointwise dominated by any true component, so it is a valid e-value for the union null without cross-environment independence (Lemma 1). Theorem 1 shows predictable allocation and stopping preserve each component's supermartingale property when environments are interleaved, and Theorem 2 feeds the stopped minima to e-LOND to convert candidate-level conditional validity into stream-level persistent FDR control under the controller's stated dependence conditions.
Core claim. The central claim is that model replacement deserves its own statistical control point in adaptive development, and that such control is achievable. A challenger is called persistently beneficial only if it beats a predeclared practical margin in every declared environment, giving an intersection-union target. The paper constructs a persistent candidate e-value, $E^{\mathrm{pers}}_t = \min_{e\in\mathcal{E}} E^e_t$, the pointwise minimum of component betting e-processes, and proves (Lemma 1) that this minimum is conditionally valid under the union null without any cross-environment independence, priced only in possible conservatism when one environment accumulates evidence slowly. Theorem 2 then shows that feeding these e-values to e-LOND yields online persistent-target FDR control at level $\alpha$ under the controller's stated dependence conditions, even though an adopted challenger becomes the incumbent from which future candidates are generated. Empirically, development-only e-LOND reaches persistent FDR 0.621 in a 5,000-trajectory closed loop, while the audited EGR variants show no persistent false-adoption path (one-sided 95% binomial upper bound near $6\times 10^{-4}$), and matched replay preserves alternative crossing decisions while using 56.1% less decision evidence.
The premise it rests on. The entire guarantee rests on Assumption 1: whenever a component null is true, the acquisition mechanism must yield bounded selected updates whose conditional mean is non-positive, and this must hold for the actual evidence unit consumed by the sequential update, not merely for marginally valid correlated raw rows.
If the paper is right (Pith's inferences, not the authors')
A wrong replacement no longer needs to silently bias every subsequent challenger: earlier adoptions may change future proposals without invalidating the current candidate's evidence, because validity is conditioned on the realized adaptive-search history.
Development-only validation is a distinct failure mode that online multiplicity correction alone cannot repair: in the closed loop it stays valid for the development hypothesis yet attains persistent FDR 0.621.
Candidate-level anytime validity does not control a repeated adoption stream: reusing an anytime-valid two-environment test without stream correction yields all-null FDR and $P(V_T>0)$ of 0.999 by $T=500$.
Once the target, evidence unit, and candidate identity are fixed, evidence placement becomes an efficiency problem: in matched replay over 600 pairs, Stagewise EGR uses 56.1% less decision evidence at identical power, and weakest-component balancing stays within a few percent of an oracle's evidence use.
Reusing decision evidence to shape a candidate inflates persistent-null crossing roughly 73-fold (from 0.052% to 3.802%), while discarding the leaked observations and restarting on fresh evidence restores the frozen rate.
If the persistent-target guarantee transfers to real deployment data, automated research agents gain a principled stop rule for model replacement: adoption is permitted only when the persistent e-value clears the controller threshold, making 'when is there enough evidence to replace the incumbent' a computable quantity rather than a judgement call.
The controller sweep suggests that pairing the EGR candidate interface with a stronger online e-value controller (e-LORD or e-SAFFRON) would recover the long-horizon power loss of e-LOND at fixed FDR; the paper measures this sensitivity on identical e-value streams but does not couple controller choice with acquisition routing, a natural next step.
The first-order balancing bound (Proposition 2) implies evidence cost scales with the inverse of the weakest component's evidence rate; on heterogeneous, time-varying environments this scaling is untested and could be refined or falsified by experiments where evidence rates are unknown and drifting.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Adaptive model search is path dependent: once a challenger is adopted, it becomes the reference from which later candidates are generated. A statistically unsupported replacement can therefore alter hypotheses that have not yet been proposed. We introduce Evidence-Gated Research (EGR), a statistical adoption layer for moving-incumbent search. EGR freezes each challenger before decision evidence is revealed, builds anytime-valid evidence across a predeclared set of environments, routes evidence predictably toward unresolved components, composes a persistent candidate e-value, and passes that e-value to an online controller. Under explicit conditional-validity and predictability conditions, the resulting procedure controls false discovery rate for the declared all-environment adoption target even though earlier adoptions change later challengers. In a 5,000-trajectory closed-loop benchmark, development-only e-LOND attains persistent FDR 0.621, whereas no persistent false-adoption path is observed for the audited EGR variants in that finite run. In matched replay over 600 challenger–incumbent pairs, Stagewise EGR preserves fixed-anytime alternative crossing decisions while using 56.1% less decision evidence at the representative threshold. A three-environment public-data study and a 40,000-sample controlled neural benchmark reproduce the evidence-efficiency pattern. These results identify model replacement as a distinct statistical control point in adaptive model development.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-02
The authors' main claim, as the classifier picked it out 0.86
“These results identify model replacement as a distinct statistical control point in adaptive model development.”
How the main claim is stated: asserted 0.99
What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.48 · human participants 0.04 · controlled comparison 0.84 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.95 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Adaptive model search is path dependent: once a challenger is adopted, it becomes the reference from which later candidates are generated. | background 1.00 | 0.10 | 0.04 | asserted 1.00 |
| A statistically unsupported replacement can therefore alter hypotheses that have not yet been proposed. | background 0.97 | 0.06 | 0.05 | asserted 0.53 |
| We introduce Evidence-Gated Research (EGR), a statistical adoption layer for moving-incumbent search. | contribution 0.99 | 0.84 | 0.08 | asserted 0.99 |
| EGR freezes each challenger before decision evidence is revealed, builds anytime-valid evidence across a predeclared set of environments, routes evidence predictably toward unresolved components, composes a persistent candidate e-value, and passes that e-value to an online controller. | method 0.98 | 0.82 | 0.08 | asserted 0.99 |
| Under explicit conditional-validity and predictability conditions, the resulting procedure controls false discovery rate for the declared all-environment adoption target even though earlier adoptions change later challengers. | contribution 0.41 | 0.78 | 0.51 | asserted 1.00 |
| In a 5,000-trajectory closed-loop benchmark, development-only e-LOND attains persistent FDR 0.621, whereas no persistent false-adoption path is observed for the audited EGR variants in that finite run. | result 1.00 | 0.95 | 0.15 | asserted 1.00 |
| In matched replay over 600 challenger--incumbent pairs, Stagewise EGR preserves fixed-anytime alternative crossing decisions while using 56.1% less decision evidence at the representative threshold. | result 1.00 | 0.95 | 0.08 | asserted 1.00 |
| A three-environment public-data study and a 40,000-sample controlled neural benchmark reproduce the evidence-efficiency pattern. | result 0.91 | 0.91 | 0.07 | asserted 1.00 |
| These results identify model replacement as a distinct statistical control point in adaptive model development. | interpretation 0.99 | 0.80 | 0.77 | asserted 0.99 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from simulation or computation 0.78
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-02: 1 request, 6,999 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (19)
- [1] A. N. Angelopoulos, S. Bates, E. J. Candès, M. I. Jordan, and L. Lei. Learn then test: Calibrating predictive algorithms to achieve risk control. The Annals of Applied Statistics, 19(2):1641--1662, 2025 on Pith
- [2] E. Aygün, A. Belyaeva, G. Comanici, et al. An AI system to help scientists write expert-level empirical software. Nature, 654:909--916, 2026 on Pith
- [3] M. Baratchi, C. Wang, S. Limmer, J. N. van Rijn, H. H. Hoos, T. Bäck, and M. Olhofer. Automated machine learning: Past, present and future. Artificial Intelligence Review, 57:122, 2024 on Pith
- [4] R. L. Berger and J. C. Hsu. Bioequivalence trials, intersection--union tests and equivalence confidence sets. Statistical Science, 11(4):283--302, 1996 on Pith
- [5] J. S. Chan, N. Chowdhury, O. Jaffe, et al. MLE-Bench: Evaluating machine learning agents on machine learning engineering. In International Conference on Learning Representations, 2025 on Pith
- [6] C. Dwork, V. Feldman, M. Hardt, T. Pitassi, O. Reingold, and A. Roth. The reusable holdout: Preserving validity in adaptive data analysis. Science, 349(6248):636--638, 2015 on Pith
- [7] A. E. Ghareeb, B. Chang, L. Mitchener, et al. A multi-agent system for automating scientific discovery. Nature, 655:497--505, 2026 on Pith
- [8] E. Han, C. Huang, and K. Wang. Model assessment and selection under temporal distribution shift. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235:17374--17392, 2024 on Pith
- [9] R. Johari, P. Koomen, L. Pekelis, and D. Walsh. Always valid inference: Continuous monitoring of A/B tests. Operations Research, 70(3):1806--1821, 2022 on Pith
- [10] C. Lu, C. Lu, R. T. Lange, Y. Yamada, S. Hu, J. Foerster, D. Ha, and J. Clune. Towards end-to-end automation of AI research. Nature, 651:914--919, 2026 on Pith
- [11] T. Pandeva, P. Forré, A. Ramdas, and S. Shekhar. Deep anytime-valid hypothesis testing. In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics, PMLR 238:622--630, 2024 on Pith
- [12] A. Ramdas and R. Wang. Hypothesis testing with e-values. Foundations and Trends in Statistics, 1(1--2):1--390, 2025 on Pith
- [13] A. Ramdas, P. Grünwald, V. Vovk, and G. Shafer. Game-theoretic statistics and safe anytime-valid inference. Statistical Science, 38(4):576--601, 2023 on Pith
- [14] E. Toledo, K. Hambardzumyan, M. Josifoski, et al. AI research agents for machine learning: Search, exploration, and generalization in MLE-bench. In Advances in Neural Information Processing Systems, volume 38, 2025 on Pith
- [15] I. Waudby-Smith and A. Ramdas. Estimating means of bounded random variables by betting. Journal of the Royal Statistical Society Series B: Statistical Methodology, 86(1):1--27, 2024 on Pith
- [16] Z. Xu and A. Ramdas. Online multiple testing with e-values. In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics, PMLR 238:3997--4005, 2024 on Pith
- [17] Z. Xu, L. Fischer, and A. Ramdas. Improving online FDR procedures via online analogs of e-closure and compound e-values. In Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337:7552--7561, 2026 on Pith
- [18] M. Zecchin, S. Park, and O. Simeone. Adaptive learn-then-test: Statistically valid and efficient hyperparameter selection. In Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:74018--74036, 2025 on Pith
- [19] Y. Zhang, Z. Wei, H. Ren, and C. Zou. e-GAI: e-value-based generalized -investing for online false discovery rate control. In Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:76308--76331, 2025 on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment