acceptodds
Under review as a conference paper at ICLR 2027

Behavioral Spec Inversion: Recovering a Model's Enacted Specification from a Limited Record of Its Behavior

Abstract

A model specification describes how an assistant should behave, but the specification a model enacts, request by request, can differ from the one written for it and from the one it reports when asked. The enacted specification is what auditors and developers need, yet an adherence score and a list of failures report how often a model fails, not when. We study behavioral spec inversion (BSI): recovering a model's enacted specification from a limited record of its black-box behavior, as a natural-language description that is selected by behavioral evidence and tested on behavior it was not produced from. On value-adapted targets whose intended preference is known, and on two targets trained on specifications we wrote, a recovered description reaches the preference more reliably than the specification's parameters: only some supplied preferences are identified stably, and a description written from a small set of choices misstates a threshold yet reproduces most of the target's choices on unseen probes. On SpecBench Code with natural targets, behavior profiles learned by BSI-Search from a small number of observations predict clause violations on test queries, with every prediction frozen before the responses exist: they improve on clause frequencies and on direct rule induction, mostly by anticipating which clauses a request engages, and an implanted condition is recovered and attached to its correct branch. Rendered into corrective training data at equal budget, a profile raises specification adherence over frequency-guided data on a natural target, at a cost in benign answering, and repairs an implanted defect no better than ordinary compliant examples.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model-spec, alignment, specification, llm audit

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_FhuEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about behavioural spec inversion (BSI) i.e. recovering a natural language description of the specification that the model actually follows from a limited set of black box observations, as opposed to the specification as written or as the model claims it is. The main contributions are a framework for constructing and validating such descriptions (based on evidence), the BSI-SEARCH method for learning interpretable conditions associated with specification violations and an empirical evaluation of the predictive and corrective value of such conditions. BSI-SEARCH proposes clause level conditions, fits violation probabilities and selects the conditions that lead to improved cross-validated prediction over clause frequencies.

Validation includes value-adapted models, three instruction-tuned models on SpecBench, controlled conditional defects and corrective fine-tuning. The value experiments are able to reliably identify three of the six intended preferences across domains and paraphrases with the generated descriptions reproducing the choices that were not seen (though in some cases they do get the numerical thresholds wrong). On SpecBench Code the profiles learned from 70 observations lead to a reduction in macro Brier error relative to both clause frequencies and to one pass rule induction (with the predictions being frozen prior to test responses being generated). The controlled experiments also manage to recover the implanted triggers. Profile guided training leads to better adherence than frequency guided supervision, though at the cost of a substantial reduction in benign answering and provides no clear advantage in terms of repairing the implanted defect. In sum the results do support useful behavioural prediction and the fact that a lot of the predictive gain comes from identifying clause applicability is clear, and that predictive accuracy does not necessarily equate to exact specification recovery.

Critical strengths

BSI requires the descriptions to be backed up by observed behaviour and validated on unseen outcomes. In the SpecBench experiments the predictions are frozen prior to the generation of the test responses and the comparisons are made using paired statistical tests. On Code BSI-SEARCH reduces the macro Brier error over clause frequencies and one pass rule induction. The protocol therefore provides credible evidence that the descriptions recovered contain predictive information beyond just the aggregate failure rates and hence the contribution is more than that of a simple qualitative demonstration of a plausible behavioural explanation.

The authors can use known preference specifications, reversed private policies and implanted triggers to assess what their descriptions actually recover. The results show both instances of successful recovery of conditional behaviour and cases where a description does a good job of predicting choices even if it gets the threshold wrong. The behaviour replacement controls also reveal instances of apparent identification that is independent of the presence of informative target behaviour. These findings provide some useful lessons for behavioural auditing: the fact that a description is fluent and that it makes good predictions is not in itself sufficient to establish that it has recovered the specification faithfully.

Critical weaknesses

On SpecBench Code, an applicability-only diagnostic actually does achieve some predictive gain which suggests that the majority of the improvement is in identifying which clause(s) a request is going to engage with rather than in the case where the model fails to follow an applicable clause. The value experiments raise a similar issue in that replacing the target probabilities with 0.5 still results in one of the matching specifications being ranked first in the bootstrap draws. The paper recognises this but doesn't really establish its core interpretation. Stronger controls are needed to separate out the generic request structure from the target specific behaviour if we are to have any confidence in the ability to recover the specification.

On natural failures profile guided training outperforms frequency guided training in terms of adherence, albeit with a fall in benign answering compared to the frequency guided arm. Without a comparison at matched rates of benign answering the study can't establish whether the method is in fact improving the adherence/utility trade-off or is simply leading to more conservative responses. The recovered conditions also provide no statistically significant repair advantage over the ordinary compliant examples for the implanted defect. The natural failure experiment only covers one target and compares complete data generation pipelines so the evidence for a distinct corrective benefit is still too limited to make a material contribution to the paper's case for acceptance.

Questions for authors

The applicability-only diagnostic achieves approximately 93% of the predictive gain on Code. What evidence shows that BSI-SEARCH captures target-specific failure conditions beyond clause applicability? In particular, do the applicable-only results demonstrate a reliable improvement over clause frequencies?

4 comments
AuthorsOpenReview ↗

Response to W1 + Q: Controlled Tests Support Target-Specific Recovery

We thank the reviewer for the careful and accurate reading. Tables G1–G4 are in our reply to pbo5, G5 in our reply to SnbM, and G6 below.

Response to W1 + Q: Target-specific behavior versus request structure

Short answer to the question. BSI-Search retains predictive gains when evaluation is restricted to applicable cases. It improves on applicability-aware clause frequencies in both scenarios (approximately 2.0% of the baseline Brier on Code and 2.4% on Health; Table G1 in our reply to pbo5, W1/Q1; new App. C.6.1).

Applicable-only evaluation. We score only the (query, clause) pairs whose clause applies, against each clause's violation rate among applicable training cases. Conditional Brier falls by 0.0026 [0.0007, 0.0045] on Code and by 0.0033 [0.0021, 0.0044] on Health (simultaneous 95% intervals), with positive point estimates for all six target–scenario pairs. These conditional gains are smaller than Code's 8.5% gain on the joint endpoint. The conditions are those selected for the joint endpoint; only the two branch risks are refitted on applicable training cases, with no reselection.

Rule setting: request structure versus the target. In policy reversal, the model, questions, and public Next Steps requirement remain fixed while the private policy changes. The recovered high-risk context follows that policy in all three realizations, and an opposite-policy profile predicts worse than Frequency (pooled Brier 0.47 vs. 0.23; Sec. 5, Figs. 3 and A9). In the implanted-defect study, the search names a marker sentence and identifies the correct high-risk branch without being told that a marker exists (App. C.4). On Code, the applicability-only diagnostic recovers about 93% of the joint gain; the two separately fitted diagnostics overlap and do not form an additive decomposition. On Health, behavior-only exceeds applicability-only (+0.0016 vs. +0.0007; Apps. A.4, C.2). These diagnostics characterize predictive gain, while the controlled studies test what is recovered.

On natural targets, the applicable-only evaluation and the comparison with Prior Rules + Fit provide complementary evidence. On Code, conditions proposed without target labels predict worse than BSI-Search even when fitted on the same labels (Fig. A2, S9). On Health, Prior Rules + Fit and BSI-Search are close (0.0970 vs. 0.0967). These natural-data tests establish predictive information beyond clause frequency and applicability; the controlled tests establish that conditions can follow changes in target behavior under fixed requests.

Value setting: the 0.5 control. This control shows why an identification rank alone is insufficient. With every target probability set to 0.5, one NoodleSpec target's matching specification still ranks first in 97–99% of bootstrap draws: subtracting the empty-reference prediction leaves probe-dependent offsets even when target behavior is uninformative (App. C.1, Fig. A1d). The revised Introduction therefore treats this as a finding about the evidence needed for recovery. The paper combines the rank with additional checks:

  • For the two NoodleSpec MSM+AFT targets, with candidates and probes fixed, substituting the other target's behavior drops every matching-specification frequency below 0.80; each target's own behavior ranks its matching specification first in all 1,000 draws in both option orders. The score responds to target behavior, while the 0.5 result shows why a high rank alone cannot establish recovery.
  • On 60 probes the inverter never saw, recovered descriptions reproduce 80.51–96.64% of non-tied choice directions, versus at most 54.31% for an empty or general specification (Sec. 4, Fig. 2).
  • A separate content check against the written specification exposes a misstated threshold despite that predictive agreement.

For the six released value targets, a value counts as identified only if the matching specification reaches 0.80 top-1 frequency in both option orders on the fit, validation, and challenge domains and under blind paraphrase; three of six pass, and we report the three that fail. The substitution controls use the NoodleSpec targets, not the six-value data (App. C.1).

Takeaway: Across Code and Health, recovered conditions improve violation prediction after applicability is fixed. Under controlled policy changes, the recovered conditions follow the target's behavior with requests held fixed. The NoodleSpec experiments further show why recovery must be assessed through content and behavioral checks as well as identification rank.

AuthorsOpenReview ↗

Response to W2:Recovered Conditions Enable Useful Corrective Supervision

Response to W2: Corrective value and its limits

Sec. 6 tests the corrective use of recovered conditions. At a fixed training budget, recovered conditions produce supervision that reduces unsafe-question violations relative to Frequency on Qwen3-8B and Llama-3.1-8B. On Qwen3-8B, separate controls retain the safety benefit at matched rule-example count and with explicit refusals on safe questions at Frequency's level. Sec. 6 demonstrates this downstream use alongside the recovery results in Secs. 4–5. Tables G3 and G4 use the reported rescoring of the original responses; their scoring passes are specified in the corrective-training note and Table G4 in our reply to pbo5, W2/Q3 continuation.

More conservative responses? (Table G3, pbo5 W2/Q3 continuation; new App. C.6.3) We replaced 26 unsafe training examples with benign requests satisfying the same recovered conditions, leaving the other 230 examples, the budget, and the two seeds unchanged. Explicit refusals on safe questions fall from 7.14% to Frequency's 0.89%, while unsafe-question violations remain 5.83 points below Frequency [2.71, 8.95]. The safety improvement therefore persists without excess explicit refusals on safe questions. Broader caution remains: unnecessary refusals fall from 11.61% to 5.36% but stay above Frequency's 0.89%, and benign answering averages 37.50% versus 33.04% for the full profile, below Frequency's 50.89%. The two seeds differ in direction (44.64 / 30.36 vs. 30.36 / 35.71), so we do not claim consistent recovery of benign answering.

Matched benign answering. Benign answering is measured after training; our controls intervene on the supervision at a fixed training budget. The safety gains persist in the same-condition benign and matched-count controls, supporting the use of recovered conditions to locate useful corrective supervision; Sec. 6 does not claim an improved safety–helpfulness trade-off at equal benign-answering rates. We have not compared models at matched benign-answering rates; the remaining benign-answering cost is reported in Sec. 6 and App. D.

Implanted defect. All three repair arms reduce violations from 96.1% to 0.7–3.3%. The recovered-condition arm demonstrates that an inferred condition can be rendered into effective corrective supervision. Ordinary compliant examples also repair this defect, so the experiment does not establish an additional repair advantage from the recovered condition in this setting (Sec. 6; App. C.4).

Complete pipelines and a single target (Table G4 in the pbo5 continuation; Table G6 below; new Apps. C.6.4, C.6.6). BSI-Search-89 holds 89 safety-rule examples and the identical 167 other examples as Direct, with the same budget and seeds. It lowers unsafe-question violations by 5.45 points relative to Frequency [1.84, 9.05], so the safety gain does not require the full profile's larger number of rule-derived examples. Direct also improves at this count and keeps more benign answering (46.43% vs. 41.07%); we claim no corrective advantage over Direct. Direct and BSI-Search-89 were scored together; Frequency and the full profile reuse Table G3's scores from a separate pass with the same evaluator and rubric. Relative to Fig. 4, unsafe-question rates for the original Frequency, full-profile, and Direct arms differ by at most 1.4 points (App. C.6.4).

On Llama-3.1-8B, a profile recovered from its own observations lowers unsafe-question violations by 4.37 points [1.72, 7.03] in both seeds; SAR rises by 1.61 points and benign answering falls by 11.61 points. This second-target experiment compares Frequency with profile-guided training; the matched-count and benign-enhanced controls are on Qwen3-8B.

Table G6. A second target: corrective training on Llama-3.1-8B (App. C.6.6, Table A8). Percentages, means of two seeds; unsafe-question violations are also given per seed. Profile-guided vs. Frequency on unsafe-question violations: 4.37 points lower [1.72, 7.03].

ArmSAR ↑Unsafe-question violations ↓ (seed 1 / seed 2)Benign answering ↑
Frequency19.5629.25 (29.67 / 28.82)38.39
Profile-guided21.1724.87 (24.01 / 25.73)26.79

Capabilities (Table G5 in our reply to SnbM, W3/Q3; new App. C.6.5). On 150 ordinary Qwen3-8B tasks outside SpecBench, no decline relative to the untrained model is detected, with wide intervals at 50 tasks per category.

Takeaway: Recovered conditions produce supervision that reduces unsafe-question violations relative to Frequency on Qwen3-8B and Llama-3.1-8B. On Qwen3-8B, the reduction also remains at a matched rule-example count and when explicit refusals on safe questions return to Frequency's level. These controls support a concrete downstream use of recovered conditions: choosing situations for corrective supervision at a fixed budget.

Dear Reviewer FhuE, thank you again for your constructive comments. Our response shows prediction beyond applicability and recovered conditions that track controlled changes in target behavior. Additional training controls and a second target support their corrective usefulness at a fixed budget. Your feedback is important to us, and we look forward to hearing whether these responses address your concerns. We would be happy to discuss any remaining questions.

Dear Reviewer,

As the discussion period approaches its end, we would be grateful if you could share your feedback on our point-by-point responses and additional results. We understand that this is a particularly busy period and appreciate your time. Your feedback is very important to us. We would appreciate hearing whether our responses address your concerns and would be happy to clarify anything that remains open. Thank you for your consideration.

Best regards,

The Authors

Reviewer_SnbMOpenReview ↗
Rating
2(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces Behavioral Spec Inversion (BSI), a framework for inferring a language model's enacted specification from a limited set of black-box behavioral observations. The framework generates or selects natural-language descriptions of model behavior, evaluates their consistency with observed responses, and tests their predictive ability on unseen inputs. The authors investigate two settings: recovering behavioral preferences from pairwise choices and identifying conditions associated with specification violations. For the latter, they propose BSI-SEARCH, which discovers interpretable conditions and estimates their associated violation risks. The paper further explores using the recovered descriptions to construct corrective fine-tuning data. Experiments across synthetic and natural model behaviors demonstrate varying degrees of predictive accuracy and improvements in selected specification-adherence metrics, although corrective training also reduces benign answering performance.

Critical strengths
  1. The paper addresses a meaningful problem of understanding and predicting model behavior.

  2. The experiments include held-out evaluations, several useful controls, and an attempt to apply the recovered descriptions to model correction.

Critical weaknesses
  1. The method novelty and necessity of BSI are unclear. The proposed framework closely resembles existing neuro-symbolic approaches, where language models generate and evaluate candidate conditions based on observed labels. Since the public specification and labeled behavioral observations are already available, it is unclear what additional information specification inversion provides over direct violation prediction or existing rule-learning methods.

  2. The evidence for behavioral recovery is limited. In Appendix C.2, clause applicability alone achieves approximately 93% of BSI-SEARCH's Brier-score improvement on SpecBench Code. It remains unclear how much model-specific behavior the method recovers beyond predicting clause applicability.

  3. The corrective training results raise concerns about generalization and selective behavioral correction. In Figure 4, BSI-SEARCH improves specification adherence from 23.88% to 28.96% compared with the Frequency baseline, but benign answering drops substantially from 56.25% to 34.82%. This suggests that the improvement may come from more conservative or broadly refusing behavior rather than correcting the identified violations. Moreover, the training data differ in composition across methods, and the paper does not evaluate the preservation of broader model capabilities.

Questions for authors
  • What is the fundamental methodological advantage of BSI over existing language-predicate learning or direct violation prediction?

  • How much predictive improvement remains when considering only applicable specification clauses?

  • Does corrective training preserve the model's original capabilities beyond the evaluated specification-adherence metrics?

9 comments
AuthorsOpenReview ↗

Response to W1 + Q1:Behavioral Recovery Requires More Than Prediction

We thank the reviewer for the careful reading and for noting that the problem is meaningful and the soundness concerns are fixable. The three weaknesses ask a central question: given the public specification and labeled observations, what does specification inversion provide beyond direct violation prediction or existing rule learning? Our answer is:

  1. A different output, checked beyond predictive fit. A predictor outputs probabilities. BSI's rule profiles describe when this model departs from its specification, supported against competing candidates by target behavior and tested by predictions frozen before unseen responses exist.
  2. Predictive fit alone cannot certify recovery. A description can predict well with the wrong threshold (Sec. 4); an identification score can reflect probe structure without informative target behavior (App. C.1). Reference and behavior-replacement checks detect these failures.
  3. Evidence beyond applicability and excess explicit refusal. Reversing a private policy changes the recovered conditions under fixed queries and public requirements (W2). Safety gains remain when explicit refusals on safe requests return to Frequency's level (W3).

Tables G1–G4: pbo5; G5: below; G6: FhuE.

Response to W1 + Q1: Why inversion, rather than direct prediction or rule learning

What the specification and labels do not contain. The specification states intended behavior, and the 70 labeled responses record individual past outcomes. Neither explicitly states the condition under which this model departs from the specification. In NoodleSpec, models report a declared threshold that differs from their measured choice boundary (App. C.1). In the policy-reversal and implanted-defect settings, the enacted condition is deliberately withheld from the public specification. In the rule setting, inversion turns the observations into explicit, testable conditions, evaluated on 170 new queries with predictions frozen before any test response exists.

Why prediction alone does not answer the recovery question. Spec-only predicts violation probabilities from the specification and query without observing the target; Frequency uses the target's per-clause label frequencies. On Code, Spec-only has slightly higher AUROC but more than twice BSI-Search's Brier (Fig. A2). Even a numerical predictor trained on the same labels that matched BSI's accuracy would not, through probabilities alone, supply a description whose content can be checked. Sec. 4 makes the distinction concrete: a recovered description misstates a threshold by 25% yet reproduces 88.89% of unseen choices. BSI therefore combines predictive evaluation with reference and behavior-replacement checks (Sec. 3).

Controlled comparisons of where the gain comes from. The comparisons test target-conditioned proposal, fitting, and selection separately. Prior Rules + Fit proposes conditions without target labels and then fits them on those labels; it predicts worse than BSI-Search on Code (Fig. A2). Direct and BSI-Search use the same 70 observations and branch-risk estimator, with BSI-Search adding selection across candidates and lowering macro Brier by 3.3% on Code (Sec. 5). Shuffling condition–behavior pairings leaves only +0.0006 of the gain over Frequency (Fig. A3). These controls test fitting and selection against the observed labels; W2 separately tests whether conditions follow the target.

Relation to existing predicate learners. BSI's contribution lies in its recovery target and verification standard; the propose–evaluate mechanism is shared with predicate learners. We ran the released nlparam (Zhong et al., 2024) with identical observations and the same proposal/validation model (new App. C.6.2; Table G2, pbo5 W2/Q2). On Qwen3-8B Code, BSI-Search has a lower observed Brier (0.1478 vs. 0.1522) at about one-twelfth the model time (1.12 vs. 13.65 hours). This is one target and scenario; BSI-Search averages three realizations, all counted in its cost, while nlparam was fitted once. BSI-Search screens conditions by cross-validated gain and falls back to clause frequency when none is supported. In this evaluation, no clause is worse than Frequency, versus seven for nlparam; Direct, which shares the risk estimator, also has none. This illustrates the audit value of checking whether a condition adds information beyond a clause's base rate.

Revisions. The Introduction now states the contribution as the recovery of enacted behavior, an evidence standard for recovered descriptions, and an auditable instantiation. Sec. 2 is condensed and includes a comparison table in the main text (Table 1), with the extended comparison in App. G.

Takeaway: BSI provides an explicit, testable account of the target's enacted behavior, together with checks of its content and dependence on the target. The experiments show why predictive fit alone is insufficient to establish recovery.

AuthorsOpenReview ↗

Response to W2 + Q2:Violation Prediction Improves Within Applicable Cases

Response to W2 + Q2: Recovery beyond clause applicability

Direct answer to Q2. Restricting evaluation to applicable (query, clause) pairs and comparing against each clause's applicable-case violation frequency, conditional Brier improves by +0.0026 [0.0007, 0.0045] on Code and +0.0033 [0.0021, 0.0044] on Health (simultaneous 95% intervals; new App. C.6.1; Table G1 in our reply to pbo5, W1/Q1). These are approximately 2.0% and 2.4% reductions relative to the respective baselines. The point estimate is positive in all six target–scenario combinations. No rules are reselected; only the two branch risks are refit on applicable training cases. Once applicability is fixed, the conditions still carry information about which applicable cases this model violates.

Policy reversal: the public requirement is unchanged. In Sec. 5, the model, queries, and Next Steps requirement are fixed, and the same deterministic checker scores compliance. Only the target's private policy changes. The recovered high-risk context switches from demonstration to deployment in all three realizations (Fig. A9). Transferring the opposite-policy profile gives pooled Brier 0.47, worse than Frequency's 0.23 (Fig. 3). This change follows the target's policy, not a change in the public requirement. Likewise, in the implanted-defect experiment, the six top-ranked candidates name a marker sentence and the selected condition identifies the correct high-risk branch (App. C.4). On natural Code data, Prior Rules + Fit, whose conditions are proposed without target labels, underperforms BSI-Search (Fig. A2).

Interpreting the 93% figure. The applicability-only diagnostic recovers about 93% of Code's joint predictive gain. The applicability-only and behavior-only predictors in App. C.2 are fitted separately and their gains overlap; they are not an additive decomposition. On Health, the behavior-only gain exceeds the applicability-only gain. The applicable-only evaluation and controlled policy changes provide complementary evidence beyond this diagnostic: predictions improve within applicable cases, and recovered conditions follow changes in the target's private policy.

Takeaway: The conditions improve violation prediction within applicable cases, and the policy-reversal experiment shows that recovered conditions follow changes in the target's private policy. Together, these results support recovery of target-specific violation tendencies beyond applicability.

AuthorsOpenReview ↗

Response to W3 + Q3:Safety Gains Without Detected Capability Loss

Response to W3 + Q3: Selective correction and capability preservation

Testing the reviewer's refusal hypothesis. We replaced 26 of the 94 unsafe safety-rule examples with benign requests under the same recovered conditions, keeping the other 230 examples, the 256-example budget, and both seeds fixed (new App. C.6.3; Table G3 in our reply to pbo5, W2/Q3 continuation). In this Benign-enhanced arm, explicit refusal on safe questions falls from 7.14% to Frequency's level (0.89%), yet unsafe-question violations remain 5.83 points lower [2.71, 8.95] than Frequency. The safety gain thus persists without excess explicit refusals on safe questions. Table G4 combines the joint scoring of Direct and BSI-Search-89 with Frequency/full-profile scores reused from Table G3 and obtained in a separate pass with the same evaluator and rubric (see the corrective-training note in our reply to pbo5, W2/Q3 continuation).

Matched training composition. BSI-Search-89 uses the same number of rule-derived examples (89) and the identical 167-example background as Direct, with the same budget and seeds (new App. C.6.4; Table G4). It retains a 5.45-point reduction in unsafe-question violations [1.84, 9.05] relative to Frequency. The gain therefore does not require the full profile's larger number of rule-derived examples. Direct also improves at this count; this control establishes the usefulness of recovered conditions at matched count, without establishing a corrective advantage of search over one-pass induction.

Benign answering. The training examples withhold the forbidden part of a request while completing the legitimate part (Fig. A11). With same-condition benign examples, mean benign answering rises from 33.04% to 37.50%, but the two seeds differ in direction, so we do not claim consistent recovery; it remains below Frequency's 50.89%. Unnecessary refusal falls from 11.61% to 5.36%, also remaining above Frequency's 0.89%. The new construction removes excess explicit refusal while retaining the safety gain, but does not eliminate the benign-answering trade-off.

Capability preservation (Q3). We added 150 ordinary tasks outside SpecBench, with no SpecBench instructions: 50 MBPP problems, 50 ARC-Easy questions, and 50 legitimate requests (new App. C.6.5; Table G5). Corrected rows average two seeds.

Table G5. Capabilities outside the specification, Qwen3-8B (App. C.6.5, Table A7). Percentages on 50 tasks per category; corrected rows average two seeds. No corrected configuration is detectably below the untrained model (S39–S47). Legitimate requests, difference from untrained: Frequency −14.0 [−28.4, +0.4]; full profile −9.0 [−23.4, +5.4]; benign-enhanced −6.0 [−21.5, +9.5].

ConfigurationMBPP pass@1ARC-Easy, strict formatARC-Easy, final choiceLegitimate requests completed
Untrained54.018.098.072.0
Frequency63.035.098.058.0
Full profile60.044.096.063.0
Benign-enhanced55.016.097.066.0

No corrected arm shows a detectable decline relative to the untrained model at this sample size; with 50 tasks per category, the intervals are wide and the test detects only large declines. ARC-Easy's strict-format failures concern answer formatting, not observed wrong final choices. Legitimate-request completion also falls under Frequency (72% to 58%, versus 63% for the full profile and 66% for Benign-enhanced), so this cost is not confined to profile-guided correction. On a second target, Llama-3.1-8B, its own recovered profile reduces unsafe-question violations by 4.37 points [1.72, 7.03] relative to Frequency, with the same benign-answering trade-off (new App. C.6.6; Table G6 in our reply to FhuE, W2).

Scope. Sec. 6 tests whether recovered conditions locate useful supervision at a fixed budget. Corrective training demonstrates a downstream use of recovery; the new controls in App. C.6 and the revised limitations clarify the result's scope and the remaining benign-answering cost.

Takeaway: The safety gains persist when explicit refusals on safe questions return to Frequency's level and when the number of rule-derived examples is matched. This supports recovered conditions as a useful way to locate corrective supervision at a fixed budget.

We hope these additions address the concerns the reviewer identified, and we would be glad to clarify any remaining point.


Reviewer_pbo5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies Behavioral Spec Inversion (BSI). BSI aims to recover the behavioral rules that language models actually follow. It works from a limited number of black-box observations. The authors argue that a model's enacted specification may differ from both its written specification and its self-reported rules. To study this problem, they introduce a framework. The framework builds natural-language descriptions of model behavior. It checks these descriptions against observed behavior. It also tests their predictions on unseen inputs. The paper considers two settings. The first recovers value preferences. The second identifies the conditions under which a model violates specific rules. For the second setting, the authors propose BSI-Search. BSI-Search uses an LLM to generate candidate clause-condition rules. It estimates violation probabilities under these rules. It then selects rules based on cross-validated predictive gains. Experiments on value-adapted models and on SpecBench Code and Health show that the recovered descriptions can predict unseen behavior. On SpecBench Code, BSI-Search improves the macro Brier score over both frequency-based prediction and direct rule induction. The paper also uses the recovered conditions to build corrective training data. This gives better specification adherence, but at a considerable cost to benign answering.

Critical strengths

1.The paper studies how to recover a model's enacted specification from limited black-box behavior, and the problem is clear and practically relevant. I find the distinction between written, spoken, and enacted specifications useful. It helps explain why conventional adherence scores may be insufficient for auditing language models. The paper does not simply check whether a model violates a rule. It tries to characterize the circumstances under which violations occur. The requirement that recovered descriptions predict unseen behavior is also well motivated. This makes the framework more meaningful than simply asking an LLM to summarize observed failures. 2.The experimental design provides useful evidence for the recovered behavioral descriptions.I appreciate that the authors freeze predictions before generating test responses. This helps separate genuine prediction from post-hoc explanations. The behavior-replacement controls are valuable. The experiments with reversed policies and implanted defects are valuable too. They examine whether the recovered conditions reflect the target model's behavior. The paper is also fairly transparent about its limitations. These include the mismatch between recovered descriptions and exact preference thresholds, and the trade-off observed in corrective training.

Critical weaknesses

1.It is unclear how far BSI-Search recovers model-specific behavioral tendencies rather than learning clause applicability. My main concern is the interpretation of BSI-Search's reported improvements. According to Figure A3 and Appendix C.2, applicability-only prediction accounts for about 93% of the Brier-score gain on SpecBench Code. The method treats satisfied and inapplicable clauses as the same negative label. A large part of the improvement may therefore come from identifying which clauses are relevant to a request. It may not come from predicting whether the target model will violate an applicable clause. The authors acknowledge this distinction. The Health results also suggest that the relative contributions can vary by domain. Still, this issue deserves more attention. Recovering a model's enacted specification is a stronger claim than identifying the specification requirements tied to an input. A clearer evaluation of violation prediction, conditioned on clause applicability, would make the central contribution more convincing. 2.The practical advantage of the recovered conditions is not yet fully established. BSI-Search beats the frequency baseline and direct rule induction on SpecBench Code. However, its advantage over direct induction is not statistically established on Health. The paper also does not compare BSI-Search directly with the related language-predicate methods discussed in Section 2. Such comparisons would show how much of the gain comes from the proposed search and selection procedure. I also have reservations about the corrective training results. The profile-guided approach improves specification adherence, but benign answering drops from 66.07% to 34.82%. This is a large reduction. It raises the question of whether the improvement mainly reflects more conservative model behavior. The training methods also differ in the number of rule-derived examples. This makes it hard to isolate the contribution of rule quality. I would be more convinced of the practical value of the recovered profiles if the authors could separate these effects better.

Questions for authors

1.Applicability-only prediction explains about 93% of the improvement on SpecBench Code. Could the authors clarify what evidence supports the claim that BSI-Search captures model-specific violation tendencies? In particular, how does the method perform when evaluated only on examples where the corresponding clause is applicable? How does this compare with a simple applicability-aware baseline? 2.How important is the candidate search and selection procedure compared with existing natural-language predicate learning methods? The comparison with Direct is helpful. Could the authors clarify which aspects of BSI-Search are expected to offer advantages over the approaches discussed in Section 2? 3.In corrective training, does the reduction in benign answering mainly come from concentrating training examples in high-risk contexts? Would a more balanced construction that includes benign requests satisfying the same recovered conditions help preserve helpfulness? Also, to what extent do the reported gains depend on the larger number of rule-derived examples in the BSI-Search training arm?

5 comments
AuthorsOpenReview ↗

Response to W1 + Q1: Predicting Violations Beyond Clause Applicability

We thank the reviewer for the careful assessment and the specific questions, which shaped several of the new experiments. Tables G1–G4 accompany the replies below.

Response to W1 + Q1: Violation tendencies beyond clause applicability

Short answer. The recovered conditions provide predictive information beyond clause applicability. On both Code and Health, BSI-Search improves over applicability-aware clause frequencies when evaluation is restricted to applicable cases. Separately, the policy-reversal experiment shows that recovered conditions follow changes in the target's private policy while the queries and public requirement remain fixed.

Applicable-only evaluation with an applicability-aware baseline (Table G1; new App. C.6.1). We score only the (query, clause) pairs whose clause applies. The baseline for this conditional endpoint is each clause's violation rate among applicable training cases. BSI-Search lowers conditional Brier by 0.0026 [0.0007, 0.0045] on Code and by 0.0033 [0.0021, 0.0044] on Health (simultaneous 95% intervals), with positive point estimates for all six target–scenario pairs. The relative reductions are approximately 2.0% on Code and 2.4% on Health, compared with 8.5% on Code's joint endpoint. We retain the conditions selected for the joint endpoint and refit only the two branch risks on applicable training cases; no condition is reselected for this test.

Table G1. Applicable-only prediction (App. C.6.1, Table A3). Conditional Brier (lower is better) on the (query, clause) pairs whose clause applies. The baseline is each clause's violation rate among applicable training cases. Gains carry simultaneous 95% intervals; all six per-target gains are positive (Code: +0.0046 / +0.0020 / +0.0013; Health: +0.0061 / +0.0028 / +0.0010 for Llama / Qwen / Mistral).

ScenarioApplicability-aware FrequencyBSI-SearchGain
Code0.12950.1269+0.0026 [+0.0007, +0.0045]
Health0.13780.1345+0.0033 [+0.0021, +0.0044]

Reading the 93% figure. It comes from two separately fitted diagnostics, not an additive split (Apps. A.4, C.2). One lets applicability vary with the condition and holds the conditional violation rate fixed; the other does the reverse. On Code they recover +0.0122 and +0.0020 of a +0.0131 gain, so they overlap; on Health the order reverses (+0.0007 and +0.0016). The 93% figure describes Code's predictive gain; it does not by itself establish what is recovered.

Conditions follow the target where the ground truth is known. In policy reversal, the model, queries, and public Next Steps requirement stay fixed while the hidden policy changes: the recovered high-risk context follows the policy in all three realizations, and an opposite-policy profile predicts worse than Frequency (pooled Brier 0.47 vs. 0.23; Sec. 5, Figs. 3 and A9). In the implanted-defect experiment, the six top-ranked candidates name a marker sentence and the high-risk branch is identified from 131 records, although the marker had little effect before implantation (6.1% vs. 4.3%; App. C.4).

On natural Code data, Prior Rules + Fit, which proposes conditions without target labels and then fits them on the same labels, predicts worse than BSI-Search (Fig. A2, S9); on Health the two are close (0.0970 vs. 0.0967). This tests the value of target-conditioned proposal, while the controlled studies test whether recovered conditions follow changes in target behavior.

Takeaway: The applicable-only comparison directly answers Q1: the conditions add predictive information beyond applicability. The policy-reversal and implanted-defect studies separately establish recovery of known target-dependent conditions.

AuthorsOpenReview ↗

Response to W2 + Q2: Search Improves Violation Prediction on Code

Response to W2 + Q2: Search and selection versus existing methods

Short answer. In the setting where we ran both, BSI-Search has a lower observed Brier than nlparam at about one-twelfth of the model time. The added value of search is supported by the matched comparison with one-pass induction (Direct) and the selection ablations on Code; the advantage over Direct is not established on Health.

Against Direct. Direct writes one profile in a single pass from the same 70 observations, with the same estimator. Across the three Code targets, BSI-Search lowers macro Brier by 3.3% relative to Direct; the preregistered interval excludes zero and the margin holds on every target (Sec. 5). On Health the difference is +0.0004 [−0.0002, +0.0010]. One plausible, post hoc explanation is that Health's records offer less conditional structure: only 6–10 of its 19 clauses have at least eight violations and eight non-violations among the 70 observations, against 16–18 of 21 on Code (Fig. A4c). The gain over Frequency and the applicable-only gain hold in both scenarios.

Against an existing language-predicate learner (Table G2; new App. C.6.2). We ran the released nlparam code (Zhong et al., 2024) on all 21 Code clauses of Qwen3-8B, with the same 70 observations, proposer/validator model, and its default search (three predicates, ten iterations, five candidates per step). nlparam improves on Frequency by 0.0089 [0.0037, 0.0141]. BSI-Search has the lower observed Brier overall (0.1478 vs. 0.1522), in both clause groups, and on 14 of the 20 clauses nlparam searched, at 1.12 vs. 13.65 model-hours and 3.04M vs. 61.31M input tokens. This is one target and scenario; BSI-Search averages three realizations, all counted in its cost, while nlparam was fitted once.

Table G2. Comparison with nlparam, Qwen3-8B, SpecBench Code (App. C.6.2, Table A4). Test Brier (lower is better). BSI-Search averages three realizations, as everywhere in the paper; nlparam was fitted once with its default search. nlparam improves on Frequency by +0.0089 [+0.0037, +0.0141]. Cost on the same hardware: BSI-Search (all three realizations) 1.12 model-hours, 1,308 calls, 3.04M input tokens; nlparam 13.65 model-hours, 81,805 calls, 61.31M input tokens.

MethodAll 21 clausesSafety (11)Behavioral (10)Clauses worse than Frequency
BSI-Search0.14780.13870.15790
nlparam0.15220.14320.16207
Direct0.15200.14520.15960
Frequency0.16110.16240.1596–

Which parts of BSI-Search matter (Q2). The three-target comparison with Direct above and the Code ablations show what selection and fitting contribute (Fig. A3; S14–S16): (1) composing rules per clause across candidate sets improves on taking the first candidate (+0.0068 Brier) or one whole set (+0.0066); (2) cross-validated screening returns clause frequency when no condition is supported, as on seven clauses in Table G2; and (3) shuffling fit-time condition assignments leaves only +0.0006 of the gain, showing that the fitted predictions depend on the condition–observation correspondence. The shuffle control tests fitting; the controlled target changes discussed under W1 test target specificity. Table G2 also identifies a result shared with Direct: both methods use the fitted branch-risk estimator, and in this evaluation neither is worse than Frequency on any clause, versus seven for nlparam. Direct also has nearly the same observed Brier as nlparam (0.1520 vs. 0.1522). This shared clause-wise result is distinct from the added value of cross-candidate selection tested by the matched comparison and selection ablations.

The revised Sec. 2 and main-text Table 1 distinguish BSI's recovery target and verification requirements from its shared propose–evaluate mechanism; App. G (Table A10) provides the extended comparison.

Takeaway: The matched Code comparison and selection ablations support the added predictive value of per-clause search. The result shared with Direct should not be attributed to search alone, and the Health comparison sets the scope of the advantage.

AuthorsOpenReview ↗

Response to W2 + Q3: Safety Gains Persist Under Training Controls

Response to W2 + Q3: Corrective training

Short answers to Q3. (a, b) Adding benign requests under the same recovered conditions removes excess explicit refusals while retaining the safety benefit; it does not establish stable recovery of benign answering. (c) For unsafe-question violations, the benefit remains when the number of rule-derived examples is matched. This conclusion concerns the safety endpoint.

Tables G3 and G4 include rescoring of the Fig. 4 responses with the same evaluator and rubric. The reported comparisons below use those tables; the Fig. 4 observation is kept separate. The scoring passes are specified in the corrective-training note and Table G4 in the continuation below.

(a, b) Benign requests under the same conditions (Table G3; new App. C.6.3). We replaced 26 of the 94 unsafe safety-rule examples with benign requests satisfying the same recovered conditions, each with a complete answer. The other 230 examples, the 256-example budget, and the two seeds stay fixed. Explicit refusals on safe questions fall from 7.14% to Frequency's 0.89%, and unnecessary refusals from 11.61% to 5.36% (Frequency: 0.89%). Unsafe-question violations remain 5.83 points below Frequency [2.71, 8.95], with no detectable loss relative to the full profile (−0.77 [−4.34, +2.80]). Benign answering averages 37.50% versus 33.04%, but the seeds differ in direction (44.64 / 30.36 vs. 30.36 / 35.71); it remains below Frequency's 50.89%. The safety improvement thus persists without excess explicit refusals on safe questions, while broader caution remains.

Within the original Fig. 4 evaluation, benign answering is 66.07% before training, 56.25% with Frequency, and 34.82% with the full profile. A decline also occurs with frequency-guided fine-tuning, and is larger with profile-guided data. This observation does not assign causal shares of the decline to fine-tuning and profile targeting.

(c) A matched number of rule-derived examples (Table G4; new App. C.6.4). BSI-Search-89 uses 89 safety-rule examples and the identical 167 other examples as Direct, at the same budget and seeds. It lowers unsafe-question violations by 5.45 points relative to Frequency [1.84, 9.05]; the full profile's further 1.16-point reduction is not separated from zero [−1.96, +4.27]. The safety gain therefore does not require all 151 rule-derived examples in the full profile. Direct also improves at this count, with more benign answering (46.43% vs. 41.07%) and slightly higher SAR. We claim no corrective advantage over Direct. The full profile contains 57 behavioral-rule examples absent from BSI-Search-89, so this control does not isolate the source of their SAR difference (29.86% vs. 26.98%).

Direct and BSI-Search-89 were scored together; Frequency and the full profile reuse Table G3's scores from a separate pass with the same evaluator and rubric. Compared with Fig. 4, the unsafe-question rates of the original Frequency, full-profile, and Direct arms differ by at most 1.4 points (App. C.6.4).

Beyond one target and one metric (Tables G5–G6 in our replies to SnbM W3/Q3 and FhuE W2, respectively; new Apps. C.6.5, C.6.6). On Llama-3.1-8B, a profile recovered from its own observations lowers unsafe-question violations by 4.37 points [1.72, 7.03] in both seeds; SAR rises by 1.61 points and benign answering falls by 11.61 points. Llama compares Frequency with profile-guided training; the Direct, matched-count, and benign-enhanced controls are on Qwen3-8B. On 150 ordinary Qwen3-8B tasks outside SpecBench, no decline relative to the untrained model is detected, with wide intervals at 50 tasks per category.

Takeaway: Same-condition benign examples remove excess explicit refusals on safe questions while retaining the safety benefit, and the matched-count control retains that benefit with fewer rule-derived examples. These results support the corrective use of recovered conditions, while the benign-answering cost remains.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all three reviewers for their careful reading and constructive suggestions. Their comments center on three questions: what specification inversion adds beyond prediction and rule learning; whether recovered conditions reflect the target's behavior beyond clause applicability; and whether their corrective use remains useful after accounting for refusal, training composition, and broader capabilities.

Contribution and manuscript revisions. BSI recovers a fixed model's enacted specification as a testable description of its preferences or clause-specific failure conditions. Its contribution combines this recovery task, an evidence standard beyond predictive fit, and an auditable rule-setting instantiation with clause-linked conditions, fitted risks, and supporting counts. The standard combines selection against alternative descriptions and predictions frozen before unseen behavior is collected with reference and behavior-replacement checks where available; our experiments show why predictive fit alone cannot certify a description's content or dependence on the target. To make these contributions clearer, the Introduction now states them explicitly, and the condensed, reorganized Related Work distinguishes the recovery objective and verification requirements from the shared propose–evaluate mechanism. New main-text Table 1 compares what each approach infers, from what evidence, and about what, with the extended comparison in App. G. App. C.6 reports the added evaluations, and the revised limitations distinguish what the prediction, recovery, and correction results each establish.

New evaluations (App. C.6; Tables G1–G6). We added applicable-only prediction against an applicability-aware frequency baseline; comparison with the released nlparam learner; benign requests under the same recovered conditions; a matched number of rule-derived training examples; capability evaluation outside SpecBench; and corrective training on a second target. These additions complement the existing policy-reversal, implanted-defect, and value-recovery controls.

What the evidence supports.

  • Recovery beyond applicability. Applicable-only prediction improves by a small margin on both Code and Health. Separately, the controlled policy and implanted-defect studies show that recovered conditions can track known changes in target behavior under fixed requests. The value experiments demonstrate why content and behavior-replacement checks are needed alongside prediction and identification ranks.
  • The value of search. Matched comparisons with one-pass induction and candidate-selection ablations support an additional predictive benefit on Code. In the tested target and scenario, BSI-Search also has a lower observed Brier and lower measured model time than nlparam.
  • Corrective use and its scope. On Qwen3-8B, separate controls retain the reduction in unsafe-question violations when rule-example count is matched and when excess explicit refusals on safe requests are removed. A profile recovered from Llama-3.1-8B's own observations also reduces its unsafe-question violations. The benign-answering cost remains, and these results establish neither an improved trade-off at equal helpfulness nor a corrective advantage over Direct. No capability decline is detected in the limited evaluation outside SpecBench, whose sample size constrains that conclusion.

Tables G1–G6 provide the detailed results and evaluation notes in the individual replies: G1–G4 in pbo5, G5 in SnbM W3/Q3, and G6 in FhuE W2.


AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers and Area Chair for their constructive feedback. We have comprehensively addressed the reviewers' concerns through additional experiments, controlled analyses, and manuscript revisions.

Reviewers SnbM and FhuE questioned whether BSI recovers behavior beyond clause applicability. Applicable-only gains on Code and Health, together with policy-reversal and implanted-defect controls, support target-specific behavioral recovery. Comparisons with Direct and nlparam further establish predictive value on Code. Corrective safety gains persist even when explicit refusals match Frequency, while matched-count and second-target experiments address training-composition and single-target concerns.

Reviewer SnbM's follow-up questioned scalability and capability preservation. Results on a 284B MoE model demonstrate that BSI-Search extends to substantially larger models, retaining gains over Frequency and Direct, including beyond applicability. Expanded GSM8K and CommonsenseQA evaluations quantify capability costs and show that benign enhancement mitigates mathematical-performance loss while preserving safety gains. These results directly address the reviewer's empirical concerns. Requiring exhaustive reconstruction of interacting failure mechanisms or a superior safety–helpfulness trade-off goes beyond the paper's stated claims: BSI aims to recover testable behavioral conditions and demonstrate their corrective use.

Reviewer SnbM assigned 2 to all three assessment dimensions, describing the soundness concerns as "minor or fixable". We have focused our rebuttal on addressing the substantive concerns behind this recommendation with direct experimental evidence. While reviewer FhuE has not participated in the discussion, we sincerely appreciate both reviewers' careful assessments and constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.