acceptodds
Under review as a conference paper at ICLR 2027

Reliable Behavior Discovery via False Discovery Rate Control

Abstract

Reliable methods for discovering behavior patterns in large generative models are increasingly essential as models become both more complex and capable. We formalize behavior discovery as a problem class of interest and highlight the task of controlling false discovery rate (FDR) as crucial for guaranteeing reliability in behavior discovery. Building upon the multiple hypothesis testing literature, we propose DISCO (Discovery through Iterative Search with COarsened e-processes), a method which provably controls FDR in behavior discovery under mild conditions. DISCO improves power over previous related eBH-based approaches by allowing for adaptivity to a data-dependent number of hypotheses to be tested. We demonstrate empirically that DISCO controls FDR with higher power than comparable approaches on both synthetic data and a novel empirical evaluation of FDR-controlling methods in settings where large generative models act as both discovery engines and targets.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
behavior discovery, red teaming, auditing, alignment, false discovery rate, e-BH, sequential hypothesis testing, e-values, failure discovery, behavior verification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MRUFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper frames behavior discovery in large generative models as budgeted multiple hypothesis testing with false discovery rate control. It shows that wrapping heuristic LLM auditors with eBH lowers the false discovery rate at the cost of fewer verified findings, and proposes DISCO, a bandit-style search procedure in which the hypothesis selector only observes coarse e-process bins so that the final eBH step can use the realized number of tested hypotheses KobsK_{obs} instead of an upper bound Kˉ\bar{K}. Pre-hoc and post-hoc bounds are derived, and experiments on a synthetic bandit and five text and image tasks report more verified discoveries than eBH baselines.

Critical strengths
  1. Framing model auditing as hypothesis testing with false discovery rate control is a useful perspective, since Figure 3 shows that existing heuristic auditors produce high false discovery proportions.

  2. Appendix L documents the prompts, task seeds, and spatial-tolerance scoring rules for visual symmetry in detail.

Critical weaknesses
  1. The paper compares itself with online eBH in Section 4 and Appendix F, but the problem it actually studies is an offline audit with a fixed query budget BB, where all rejections are made once after the budget is spent. Within this offline setting the paper still requires the agent to spend the budget in rounds of bb queries and to run a formal test on every round, and this protocol is the only source of the random hypothesis count KobsK_{obs} that DISCO is designed to handle. A practical auditor with the same budget could explore freely, fix KK candidate hypotheses, and test them on fresh queries with BH, so KK would be known before testing, while the main benefit of interleaving, reading earlier outputs YY to propose better hypotheses, is ruled out by Assumption 2. Since the paper neither describes a deployment that needs this round-by-round protocol nor compares against such an explore-then-test split with the same agent, the practical relevance of the setting is unclear.
  2. Figure 2 claims that all compared methods guarantee FDR≤α=0.1\mathrm{FDR} \le \alpha = 0.1 and that DISCO achieves the highest power. However, Section 6.1 and Appendix L.3 state that the DISCO guarantees hold with probability at least 96.9%, which corresponds to the conditional certificate of Theorem 2 and only bounds unconditional FDR by α+E[δpost]\alpha + \mathbb{E}[\delta_{post}] with δpost\delta_{post} up to 0.031 in Table 8, whereas BUW-eBH and the wrapped baselines control unconditional FDR at α\alpha. In addition, Section 4.2 notes that DISCO uses boosted eBH while Section 3 uses base eBH, and Section 6 averages runs with ∣P∣=1|\mathcal{P}| = 1 and ∣P∣=∞|\mathcal{P}| = \infty even though Remark 2 admits that ∣P∣=∞|\mathcal{P}| = \infty does not use coarsening at all. Across the seven task settings in Tables 1 and 2, DISCO has the most verified discoveries among FDR-controlling methods in only three settings.
  3. Theorem 7 extends the post-hoc certificate to multiple bins, which is the setting of Section 6 because hypotheses are abandoned once Eh<emin=0.99E^h < e_{min} = 0.99. Its proof reuses the equality in Equation 38, which charges each null the exit probability PHP_H at its realized count nhn_h, but this equality requires nhn_h to be independent of the null walk, whereas the emine_{min} rule, which Assumption 2 permits, stops weak walks after one round and keeps only walks already above emine_{min} for later rounds, so E[δpost]\mathbb{E}[\delta_{post}] need not bound the exit probability and the proof of Theorem 7 does not hold. As a result, the post-hoc certificate behind the 96.9% guarantee in Section 6.1 is not established for the procedure actually run.
  4. Assumption 3 requires the p-values of each null hypothesis to follow a known distribution P0P_0, and Algorithm 2 multiplies the per-round e-values of a hypothesis into one e-process. In Appendix L.9, however, the agent writes a new sub-hypothesis and pre-registers a new test type and null baseline in every round, and the system prompt asks for methodological diversity across sub-hypotheses. Section 6 itself states that such sub-hypotheses are only conceptually valid in falsification-based experiments, which behavior discovery is not. The e-values multiplied into one e-process can therefore come from different nulls, so the product is only valid for the intersection of these nulls, and a rejection shows that at least one operationalization departs from its own baseline rather than that the stated behavior holds.
Questions for authors

None

4 comments
AuthorsOpenReview ↗

Rebuttal response

We thank you for your feedback and are glad that you found our perspective useful; we agree that it is valuable to understand how heuristic methods can produce high FDRs currently.

We respond here to your feedback, and have uploaded a new draft with relevant edits:

  1. We agree that DISCO is not an online algorithm; indeed we state as such in Sec. 4. We only discuss the online e-BH literature as a relevant point of comparison. Additionally, we agree that an auditor which, given a budget, explores freely and then runs e-BH is a compelling structure: this is our baseline method shown in Alg. 1 and Figs. 4 and 5 (also as “Baselines + eBH” in Fig 2). In particular, the “Investigator Agent+eBH” and “ART+eBH” baselines are highly flexible agentic baselines. We find that this effectively controls FDR but does so with fewer verified discoveries than DISCO - we hypothesize that DISCO’s advantage is in combining the flexibility of an iterative structure (even in offline settings) with greater adaptivity to hypothesis count.
  2. There are 4 sub-questions in this point; we respond one by one.
  • (on the conditional certificate) We thank you for pointing out this slight empirical discrepancy. To remedy this, we add Table 3 and 6 in Appendix A.2, which demonstrate that even at smaller α=0.025\alpha = 0.025, we observe average verified discoveries at 6.56 on text tasks and 6.70 on image tasks, which is still more than all baselines (the best baselines+eBH are 3.03 and 3.68 respectively, while BUW-eBH yields 4.39 and 4.24).
  • (on boosting) We apologize for the confusion. We clarify in the new draft that in fact all “baselines + eBH” experiments were run with boosted e-BH, so the comparison is fair.
  • (on ∣P∣=∞|\mathcal{P}| = \infty) As we discuss in Remark 1, the ∣P∣=∞|\mathcal{P}| = \infty case is the infinite limit of the multi-bin setting of DISCO; in this case, the theory is simpler and can be justified directly from the pathwise bound of Eq. 19. Therefore, it is valid to abandon the coarsening step in this case, as we mention there. However, to demonstrate that both settings work empirically, we have added Tables 4, 5, 7 and 8 to the new draft, which break out results by ∣P∣=3|\mathcal{P}| = 3 and ∣P∣=∞|\mathcal{P}| = \infty. We can see in these tables that both settings have strong performance.
  • (on performance comparisons) We note that a) DISCO had by far the most verified discoveries on average (53% more than the best baselines), and b) in the four settings where DISCO was not best, a different baseline was best in each, meaning no other method was generally stronger. Additionally, in our new draft, we have rerun experiments after fixing a bug in error handling and found even better results; with DISCO “winning” 6/7 times (Tables 1 and 2).
  1. We thank the reviewer for the correctness note: this is a sharp example that points out a small issue with our formulation of δpost\delta_{\mathrm{post}} in the post-hoc theorems.
  • We note first that this does not change any experimental result (which use α⋆=αϵpost\alpha^{\star} = \frac{\alpha}{\epsilon_{\mathrm{post}}} (see Eq. 8)); all conclusions regarding FDR and verified discoveries (i.e. power) remain identical (Sec 5 and 6).
  • However, we have updated the definition of δpost\delta_{\mathrm{post}} to rely on $P_H (n_{\mathrm{max}}) $ rather than PH(nh)P_H (n_{h}), and clarify the wording of the conditional guarantee (Thm 2/6/7) to be regarding E[δpost]E[\delta_{\mathrm{post}}] rather than δpost\delta_{\mathrm{post}} - these revisions restore validity under adaptive stopping and clarify the role of δpost\delta_{\mathrm{post}} in the post-hoc bound.
  • The only numbers that this changes throughout the entire paper are in Tables 14/15 of Appendix L.3, which we recompute: this increases the average value of δpost\delta_{\mathrm{post}} from roughly 0.02 to 0.03.
  • Again, we appreciate the reviewer for catching this as it strengthens the work greatly: we have added an acknowledgement section recognizing the contribution of anonymous reviewers for this.
  1. We thank the reviewer for catching this error: the prompts in that portion of the appendix represent an old setting that were not actually used in experiments. We confirm that we did not use sub-hypothesis generation in experiments (as stated in Sec. 6) and have fixed the prompt appendix in the new draft.
AuthorsOpenReview ↗

Following Up

This note is just to check in before the end of the discussion period - let us know if you have any questions or concerns regarding our response and updated draft. We appreciate your thoughtful feedback!

Thank you for the response. Three points remain.

  1. The response still does not explain what practical scenario requires testing in rounds rather than exploring freely first. Without a clear use case for round-by-round testing, the motivation for handling a random hypothesis count remains weak.

  2. Why controlling FDR over the reported set is the right goal also remains unclear. Unlike standard multiple testing where the hypothesis family is fixed and homogeneous and the selected set is used as a whole, here the agent generates hypotheses of varying scope and potential duplicates on the fly, and the reported findings read as separate claims. Both FDR and the count of verified discoveries therefore depend on how the agent splits behaviors into hypotheses, so Figure 2 does not show that DISCO finds more distinct behaviors.

  3. On Algorithm 1, my review suggested testing fixed candidates on fresh queries with BH rather than e-BH. Because the candidates are fixed before testing and use independent queries, BH applies directly and is less conservative than converting p-values into e-values, so Figure 2 does not rule out this simpler alternative.

AuthorsOpenReview ↗

Thank you for the followup

We greatly appreciate the reviewer’s thoughtful feedback and clarifying distinctions. We have uploaded a new draft with new Appendices M and N. Appendix M shows that “baselines+BH” does improve over “baselines+eBH” but DISCO still yields far more verified discoveries on average; we will incorporate this more fully into the final draft. Appendix N uses an LLM to cluster hypotheses, collapsing near-duplicates and parent-child relationships; we find all methods (DISCO and baselines) lose 5-12% of verified hypotheses and FDR results are extremely similar; all general trends remain the same.

Details below, including expanded motivation for iterative hypothesis testing and FDR as a metric:

  1. We argue that iterative hypothesis testing (as opposed to explore-then-test approaches) is an important class of methods for the following reasons, which we will clarify in the next draft. 1) Explore-then-test approaches have a fundamental challenge of query allocation. If we allocate too little to exploration, we risk outputting few discoveries if hypothesis search is hard; allocate too much, and we risk underpowering the testing stage if discovered hypotheses are borderline. Iterative methods like DISCO navigate this tradeoff smoothly by combining the exploration and testing stages, and naturally yield a random, data-dependent hypothesis count. 2) Iterative testing can use previous statistically-calibrated results to inform future hypotheses. We empirically show this is useful - see Tables 4 and 7 in the new draft, where DISCO yields more verified discoveries than baselines even in the ∣P∣=3|\mathcal{P}| = 3 case, where model outputs and exact e-values are hidden; 3) iterative approaches are useful for audits run on restricted timelines; they may be interrupted partway through and still yield useful results by running eBH on the hypotheses explored so far.
  2. We defend FDR as a useful metric on two grounds. 1) We agree with the reviewer’s criticism around accounting for hypothesis duplication, and have added Appendix N (Tables 20 and 21) in the paper, where we use an LLM-based clustering which groups near-duplicate/parent-child hypotheses. We find that average verified discoveries over all tasks drop by 5-12% across all methods, FDR changes are very small, and general trends stay the same: DISCO/BUW-eBH/best baseline+BH have 6.02/3.92/3.31 and 6.81/4.05/2.80 verified discoveries on average on text and image tasks. 2) Conceptually, we argue FDR is an intuitive metric for user trust as it gives a sense of reliability across the output set, even if the claims may overlap or vary in scope. FDR can diverge wildly from alternatives such as Type-1 error; if 30 hypotheses are tested at type-1 level α=0.1\alpha=0.1, with 1 true one, then the output set will have FDR ≈0.68\approx 0.68.
  3. Thank you for clarifying the comment about BH; we apologize for the confusion. In the new draft, we add Appendix M (Tables 16-19) where we adapt all “baselines+eBH” methods to “baselines+BH” methods. We find that this does increase verified discoveries, but that on average across all tasks, DISCO still has more than 2x as many verified discoveries as any of the baselines+BH methods. As BH is a strong and valid approach, we will incorporate these results more fully into the next draft, and we thank the reviewer for the suggestion.
Reviewer_Bxz9OpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies adaptive discovery of reproducible behavior patterns in generative models under a query budget. DISCO accumulates evidence through calibrated (p)-values, adaptively allocates queries, freezes hypotheses, and applies e-BH using the realized hypothesis count. Coarsened feedback and a known null distribution underpin its proposed FDR certificates. Synthetic experiments and five generative-model tasks evaluate discovery power and replication-based FDR estimates.

Critical strengths

The paper addresses an important problem: discovering reproducible patterns while controlling false discoveries. Experiments across text and image tasks provide more informative evidence.

Critical weaknesses

Assumption 3’s common, known null distribution is not established for the heterogeneous tests. See the questions below as well.

Questions for authors
  1. For each experimental test, what null distribution is simulated? Please justify its validity throughout the composite null. 2.Does the knapsack bound remain valid under multi-bin adaptive allocation? Appendix H optimizes over fixed allocations, but Theorems 7 and Theorem 8 also use this bound under multi-bin feedback.
  2. How are Monte Carlo errors in crossing probabilities, conditional expectations, incorporated into the guarantee?
  3. If the composite-null applications rely on superuniformity rather than an exact null law, what establishes that uniform calibration conservatively bounds EA(m)E_A(m) and the boosted survivor expectations? Stochastic dominance can control unconditional crossing probabilities without preserving the ordering of survival-conditioned expectations. For instance, a superuniform variable p=max⁡(U,u)p=\max(U,u) can yield a (much) larger EA(1)E_A(1) than uniform calibration.
3 comments
AuthorsOpenReview ↗

Rebuttal response

We thank you for your feedback: we are glad that you found the paper interesting and the experiments informative.

We respond to your comments here. We have uploaded a new draft, and will upload another one in the near future during the discussion period with edits corresponding to your feedback.

  1. One by one:
  • (on the composite null) We use either one-sided binomial proportion tests (e.g. for bandit experiments we simulate Binom(30,0.25)\mathrm{Binom}(30, 0.25)), Student’s t-test, two-sample Fisher’s or McNemar’s test throughout (in Sec. 6, applying Tocher (PIT) randomization to all discrete tests to simulate a common U(0,1)U(0, 1)). These are all from exponential families whose test statistics satisfy the monotone likelihood ratio property, yielding likelihood-ratio dominance. Therefore, noting that the p-to-e calibrator is decreasing in p, the boundary parameter is the worst case in H_0, and we are justified in using the composite null.
  • (on the knapsack bound) Thank you for catching the subtle point. We note that the knapsack bound is a minor contribution and not used empirically or in our central theorems, and we have demoted it to the appendix in our new draft. In an upcoming draft during the discussion period, we will clarify the definition in pre-hoc multi-bin settings to handle J-ary tree stopping policies (where J is the number of active bins).
  1. (on Monte Carlo errors) We simulate 10^5 trajectories, meaning that standard errors are on the order of 10^{-3}-10^{-4}, much smaller than our statistics of interest. Formally, we can apply Clopper-Pearson or Empirical Bernstein/Hoeffding confidence bounds (for P_H and E_A respectively). We will note this in an upcoming draft (uploaded during the discussion period).
  2. (on super-uniformity) We thank the reviewer for this subtle point. We note that experimentally, as stated previously, we use exponential families (e.g. proportion tests and Student’s t-tests) which are well-behaved and as above yield likelihood-ratio dominance - which, unlike super-uniformity, is preserved under survival/interval conditioning and thus bounds E_A/boosted survivor expectations at the boundary. However, considering a generic distribution where only super-uniformity holds, we can additionally derive distribution-free upper bounds on the value of E_A in both the binary and multi-bin settings: these will be presented in our next uploaded draft.

Noting that we have uploaded an additional revision which adds the details we mentioned we "to come" above: in particular, discussing further the handling of the knapsack bound, how Monte Carlo errors can be incorporated and a distribution-free upper bound on E_A under generic super-uniformity of p-values. Thank you again for your thoughtful feedback!

Thank you for the revisions.

The corrected exit-probability and the distinction between the integrated guarantee and the realized certificate are indeed helpful. I also appreciate the added formulation and Monte Carlo error discussion.

I remain my score and confidence level.

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

This paper studies reliable behavior discovery in generative models through false discovery rate control. The authors propose DISCO, which combines adaptive hypothesis generation, sequential e-processes, and e-BH. Its main contribution is enabling FDR control with a data-dependent number of tested hypotheses through e-value coarsening and freezing. Experiments on synthetic data and LLM-driven behavioral auditing tasks demonstrate improved discovery power over several baselines.

Critical strengths
  1. The paper addresses an important problem in automated model auditing: distinguishing statistically reliable behavioral patterns from potentially spurious findings.

  2. Adapting e-BH to a data-dependent number of hypotheses is a meaningful technical direction. The proposed coarsening mechanism and associated FDR analysis are interesting.

  3. The experiments cover both synthetic settings and real generative models, including text and image behavior discovery tasks.

Critical weaknesses
  1. Potential correctness issue in the post-hoc FDR guarantee.

The proof of Theorem 2 in Appendix H.5 appears to contain an invalid step. Specifically, Eq. (38) claims P(B)\leq \mathbb E[\delta_{\mathrm{post}}], where (\delta_{\mathrm{post}}=\sum_h P_H(n_h)) uses the realized, adaptively determined sample counts.

However, this inequality does not generally hold under adaptive stopping. For example, consider two testing rounds with (n_{\max}=2), (e_{\max}=10), and null p-values independently distributed as Uniform(0,1), using the paper's p-to-e calibrator with (\kappa=0.5). Test a new hypothesis in the second round if the first hypothesis crosses the threshold; otherwise, continue testing the first hypothesis. This policy satisfies the binary-coarsening restriction, yet yields (P(B)\approx0.006244665) and (\mathbb E[\delta_{\mathrm{post}}]\approx0.006241554), contradicting Eq. (38).

Moreover, even an unconditional bound involving (\mathbb E[\delta_{\mathrm{post}}]) would not directly justify a per-run high-probability guarantee with the realized random (\delta_{\mathrm{post}}).

This raises a fundamental concern about the claimed post-hoc certificate, which is central to the proposed method and its experimental evaluation. A corrected proof or a revised theorem statement is needed.

  1. The conditional-independence argument relies on restrictions on how the hypothesis selector accesses model outputs (Assumption 2), together with a known, query-invariant null p-value distribution (Assumption 3).

However, the implementation appears to use richer information, including evidence summaries and previously observed behavioral patterns, to generate new hypotheses. It is unclear whether these information flows satisfy the assumptions required by the FDR guarantees.

Questions for authors

see the above correctness question.

2 comments
AuthorsOpenReview ↗

Rebuttal response

Thank you for the feedback - we are glad you found our motivation compelling and our key technical contribution (extending e-BH to a data-dependent hypothesis count) to be meaningful.

We have uploaded a new draft with edits in response to your feedback:

  1. We thank the reviewer for the correctness note: this is a sharp example that points out a small issue with our formulation of δpost\delta_{\mathrm{post}} in the post-hoc theorems.
  • We note first that this does not change any experimental result (which use α⋆=αϵpost\alpha^{\star} = \frac{\alpha}{\epsilon_{\mathrm{post}}} (see Eq. 8)); all conclusions regarding FDR and verified discoveries (i.e. power) remain identical (Sec 5 and 6).
  • However, we have updated the definition of δpost\delta_{\mathrm{post}} to rely on $P_H (n_{\mathrm{max}}) $ rather than PH(nh)P_H (n_{h}), and clarify the wording of the conditional guarantee (Thm 2/6/7) to be regarding E[δpost]E[\delta_{\mathrm{post}}] rather than δpost\delta_{\mathrm{post}} - these revisions restore validity under adaptive stopping and clarify the role of δpost\delta_{\mathrm{post}} in the post-hoc bound.
  • The only numbers that this changes throughout the entire paper are in Tables 14/15 of Appendix L.3, which we recompute: this increases the average value of δpost\delta_{\mathrm{post}} from roughly 0.02 to 0.03.
  • Again, we appreciate the reviewer for catching this as it strengthens the work greatly: we have added an acknowledgement section recognizing the contribution of anonymous reviewers for this.
  1. Thank you for pointing this out. Passing this richer information is in fact permitted in several settings: 1) as part of an isolated subagent within a single hypothesis (not passing information between hypotheses), 2) in the “infinite bins” DISCO setting, where we do not require restrictive assumptions (as discussed in Remark 1 in Appendix I.2), and 3) in unrestricted (non-DISCO) baselines such as BUW-eBH. We clarify in the new draft that these settings are the only cases where these prompts are used.
AuthorsOpenReview ↗

Following Up

This note is just to check in before the end of the discussion period - let us know if you have any questions or concerns regarding our response and updated draft. We appreciate your thoughtful feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Main Response

We thank all the reviewers for their time and feedback. We feel as though the questions and notes have helped to strengthen the submission significantly. We have uploaded a new draft with a number of edits, including:

  1. An improvement to the formulation of δpost\delta_{\mathrm{post}} in the post-hoc theorems (e.g. Theorem 2) in response to reviewer concerns about validity under adaptive stopping; we have also added an acknowledgement to anonymous reviewers for pointing out this issue. We note that this issue does not change any experimental results as δpost\delta_{\mathrm{post}} is an accounting measure: all empirical measurements of FDR and verified discoveries remain the same.
  2. Tables 3-8, which break out results in two ways: 1) different values of α\alpha, demonstrating that even accounting for the extra slack provided by δpost\delta_{\mathrm{post}}, DISCO still results in the highest verified discoveries (e.g. DISCO at level α−δpost\alpha - \delta_{\mathrm{post}} has more discoveries than baselines at level α\alpha); and 2) different values of ∣P∣|\mathcal{P}| - demonstrating that DISCO performs well in both the default ∣P∣=3|\mathcal{P}| = 3 as well as the unusual ∣P∣=∞|\mathcal{P}| = \infty setting (infinite limit of bins). Future work is needed to understand the regimes in which each of these settings work best; for the moment, we keep our default results as the average of the two in order to avoid unprincipled cherrypicking.
  3. We updated some incorrect prompts in Appendix L.9, and clarified that DISCO a) only passes richer, output-dependent information in cases where this is permitted, and b) sub-hypothesis generation was not used.
  4. We additionally have updated Tables 1 and 2 with new results that are directionally after a) running more replications to improve robustness, and b) fixing some bugs in error handling.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.