- Rating
- 2
- Confidence
- 1
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 3
Read review
Paper summary
This paper asks when an AI system can keep particular answers hidden from an evaluator while still doing ordinary work to a required standard. It defines a hiding class (response policies whose observation distribution is the same across a declared group of tasks) and a hiding value (the best score such a policy can guarantee), so hiding and adequate performance coexist exactly when the required score is at most that value. Worked examples show that hiding can be possible under an expected requirement but not a hard one, and Proposition 1 (checked in Lean) bounds how many interval answers endpoint subtraction still recovers when some inventory entries are wrong. The toolkit finds the cheapest audit menu with a certificate, turns it into fixed-horizon repeated tests under explicit sampling assumptions, and includes software for saved logs, ordinary-request runs and a prospective audit.
Beyond authored examples and control programs, the model studies are mostly negative or unresolved on the hiding question: a registered inventory prediction failed, a prompted concealment pilot (Result 1) recovered answers mainly through a route the instruction barely changed, the prospective program study (TRACE-1) has a broader bound that depends on collection facts no record establishes, and two trained-concealer studies (TC-2 and TC-3) ended not established, with one of eight runs meeting every requirement. A replay of public sandbagging-game logs found that the upstream scorer credited 336 responses containing a wrong value. The empirical part is a set of smaller studies. Authored control programs show how copied prompt wording or a private coin can make an audit look successful or unsuccessful. In an inventory study, one hosted model answered all 32 direct questions wrong, report extraction recovered 24, and a frozen off-by-one correction recovered all 32, so the registered prediction that extraction would beat the direct readers failed. In a prompted concealment pilot (Result 1), the mean recovery score stayed high (23/24 vs. 1 without the instruction) mainly through a configuration that had failed qualification and whose answers the instruction barely changed. A prospective program study recorded 127/128 correct answers, but its broader error bound depends on collection facts no record establishes. Two registered studies that trained concealers on a 27B open-weight model (TC-2 and TC-3) both ended not established, with one of eight runs meeting every requirement. Finally, a replay of 69 public logs (14,969 records) from a sandbagging auditing game found that the upstream scorer credited 336 responses containing a wrong value.
Critical strengths
-
I think the core framing is useful for sandbagging and hidden-capability evaluations. The key point, which the paper makes well, is that ordinary work doesn't carry more information than a direct question by default; what matters is whether the system is required to keep doing that work accurately while it gives misleading answers to the test. Turning this into a hiding class, a hiding value and a gap makes it precise, and the examples give evaluators concrete design lessons: an expected requirement can allow hiding where a hard one doesn't, the request distribution matters (if every request ends on the corrupted close, one bad balance can make every requested difference wrong), and a recognisable evaluation request with no performance requirement places no constraint on hiding at all. These are the parts I'd expect others to reuse, and they're the main reason I see value in the paper.
-
Key predictions are registered and failed ones are reported as failed (the inventory extraction prediction, both trained-concealer verdicts), withdrawn claims are kept next to their corrections, Proposition 1 is checked in Lean, the menu certificates are re-verified with a standard-library verifier (10 of 10), and Table 2 cleanly separates observations, rejections of a stated null, further interpretations and claims about mechanism. The replay also produces a concrete finding others can use: the upstream scorer used in the public sandbagging-game logs credited 336 wrong-value responses, and recorded BBH accuracy exceeds strict accuracy by 7.2 to 16.5 points in every log.
Critical weaknesses
- I don't think the paper shows that the framework or toolkit works on a model that is actually hiding something, which is the question in the title. None of the model studies shows the approach catching a model that is trying to hide across the whole menu. The inventory study's registered prediction failed. The authors themselves describe Result 1 as a failed concealment induction rather than evidence that the audit can defeat a working concealer. The prospective program study's bound and capability reading are unresolved because no record establishes the required collection facts (one account both held the sealed answer material and ran the collection), and the account holder answered "unknown" to the conduct questions afterwards. TC-2 and TC-3 both ended NOT ESTABLISHED, with one of eight runs meeting every requirement, and the paper's own claim map lists a model that conceals against the full permitted reader as not observed. The closest case is a single-seed development result in Appendix D: an adapter trained to replace arithmetic answers on one narrow interface, whose targets the full menu reader still recovered (32/32). The repeated-test round counts (31 or 61 rounds for dispatch, 812 for ordering with the KL test) come from authored examples.
The model studies also never realize the regime where the theory has teeth. In TC-2 the concealers were trained to give substitutes on evaluation wording and the correct report on ordinary requests, so recovering totals from ordinary reports mostly checks that training didn't break those reports (H3 held on all three runs; it was hiding, H1, that held on only one). A model that tries to make its ordinary work uninformative while staying within the usefulness requirement is exactly the case where the pivot-report argument says recovery is forced. The prompted pilot asked for roughly this, but on the route that supplied most of the recovery, the instruction barely changed the answers. So right now I think the empirical contribution is a set of useful cautionary lessons (copied wording, consistent offsets, scorer bugs, unqualified configurations), not evidence that this approach catches a concealer that hides across the menu.
- The theoretical contribution is modest, and the presentation makes the paper very hard to evaluate. By the authors' own account, the core hiding and recovery statements are standard minimax and testing mathematics and not claimed new (Table 3), the menu design is a procedure rather than a new foundation (Table 1), and the concentration bounds and dynamic programming are standard, with the contribution being the reduction and the collection protocol (Appendix H). Proposition 1 is an elementary counting bound. That would be fine for a paper whose value is a clear framing plus a convincing demonstration, but here the framing is buried. The main text and appendix read like a claims ledger: a long list of study codenames (e.g., TC-2, TC-2X, TC-3, TRACE-1, RW-1, UWA-1, HPE-1, S1, P-IMP), three packages with version histories, test pass counts by platform in the reproducibility statement, withdrawn claims and internal corrections interleaved with results, and process details such as a random tape whose hash a custodian holds, studies added after the evidence freeze at the authors' request, and a trained-concealer study (S1) that arrived but wasn't integrated. Table 1 alone packs 19 claims, each with evidence and limits. Working out what is actually established requires going deep into the appendix, and I think most readers won't. I'd want the paper restructured around the framework, the design lessons and one or two studies, with the provenance material moved to the supplement.
Questions for authors
-
Apart from the single-seed adapter in Appendix D, which was trained to replace arithmetic answers on one narrow interface, is there any model result where a model trying to hide across the menu is caught through its required ordinary work? For example, what did the S1 trained-concealer study find, or what happens when a concealer is trained to minimize recovery from its ordinary reports subject to the usefulness requirement? One end-to-end positive demonstration would change my assessment a lot.
-
The framework's leverage depends on evaluation requests either not being recognisable or carrying a binding requirement: a recognisable request without one places no constraint on hiding, and one work-framed request defeated the coin control. Given that frontier models can distinguish evaluation from deployment transcripts above chance (Needham et al., 2025, which you cite), how would an evaluator get indistinguishable evaluation and ordinary requests in practice, and can the framework handle partial recognisability?
-
Appendix H notes that scoring ordinary traffic often requires knowing the target answer. Outside constructed task families, who enforces the performance requirement, and on what kinds of tasks would you expect the framework to apply?
-
The paper places its contribution in specifying the setting, the menu-design procedure, and the reduction from a usefulness certificate to a repeated test with its collection protocol, while describing the underlying minimax, concentration and dynamic-programming tools as standard. Which of these do you see as the main technical novelty, and how does it change evaluation practice?
-
Could TRACE-1 be rerun with the sealed material held by a different account from the collector, and with logged access, so that the certified 1/20 bound applies to at least one model study?