- 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
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].
| Arm | SAR ↑ | Unsafe-question violations ↓ (seed 1 / seed 2) | Benign answering ↑ |
|---|---|---|---|
| Frequency | 19.56 | 29.25 (29.67 / 28.82) | 38.39 |
| Profile-guided | 21.17 | 24.87 (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
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 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).