acceptodds
Under review as a conference paper at ICLR 2027

SCALE: Controlling Compositional Capability Gains in Black-Box LLMs

Abstract

An assistant can withhold a complete answer while still releasing the pieces needed to reconstruct it. We study capability control over adaptive interactions, where the goal is to preserve permitted assistance while limiting an attacker's final task success. We propose , a controller that learns how much a candidate response changes task solvability given the current interaction history. Paired reconstruction attempts supervise before- and after-release solvability predictions, and an explicit difference loss emphasizes the candidate's contribution. Training pairs that hold the candidate fixed while varying earlier releases teach the auditor how this contribution depends on history. At deployment, screens candidates by predicted cumulative solvability and selects useful help with a penalty on positive predicted gain. Across Qwen3-8B and Llama-3.1-8B-Instruct on mathematics and code, retains 67.2–71.9% permitted-task utility at 20.6–21.8% attack success. Its utility exceeds that of the history-aware THRD baseline by 12.8 percentage points on average, while attack success differs by at most 0.6 points. Component studies examine the roles of difference fitting, gain-based selection, and fixed-candidate training pairs. We also derive a finite-bank calibration rule that selects episode-level policy mixtures under endpoint-risk and general-utility constraints, linking history-dependent response selection to guarantees on complete interactions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model Safety, Capability Control, Multi-Turn Jailbreak Defense

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DyGXOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies capability control over adaptive interactions to preserve permitted assistance while limiting the attacker’s final task success. To solve this, they propose SCALE, a capability-control framework for multi-turn LLM endpoints. It uses paired terminal reconstruction to estimate responses’ history-dependent marginal contribution to task solvability, and then selects partial help under risk calibration. SCALE outperforms THRD on math and code. Experimenting with two 8B models, it improves utility by 12.8% on average while changing attack success by at most 0.6 points. There’s no substantial safety improvement over the strongest history-aware baseline. And The safety guarantee is narrow, limited as well as the experiments.

Critical strengths
  1. This work has an important and realistic problem setting. The SCALE targets capability control for LLM endpoints over complete multi-turn interactions, rather than single-response filtering, which improves the method’s applicability. Moreover, they explicitly models adaptive attackers, persistent, workspace state, request budgets, and terminal task success, which makes the evaluation closer to real deployment.
  2. The method is principled, and the gains in utility are clear. It estimates a response’s history-dependent contribution to task solvability via paired terminal reconstruction, then combines difference fitting, gain-bases selection indefinite-bank risk calibration. Also, the improvements(+12.8% on utility) indicates that it’s potentially to offer more help.
Critical weaknesses
  1. There’s no substantial safety improvement over the strongest history-aware baseline. SCALE’s ASR differs from THRD by only +-0.6%, which obviously mismatch the main claim- to limit an attacker’s final task success. Moreover, the evaluation lacks human validation and qualitative analysis of the improvement. The reported utility gain is only supported by automatic task scores designed by the authors, but the paper doesn't report whether SCALE’s partial responses are genuinely useful and appropriately bounded qualitatively. Displaying interaction traces and a large scale human analysis would help explain the improvement. Without these, it's not clear to judge whether the utility gain reflect true useful assistance.
  2. The experiment is limited and unconvincing. The evaluation is limited to two 8B models, two domains(math & code), and doesn’t use stronger attacks such as TAP. Thus its safety robustness and practical capability control advantage remain insufficiently demonstrated. Furthermore, the safety guarantee is narrow and not fit with the main experiments. Theorem 1 requires the attacker set to be a finite and frozen one, with a fixed policy bank, and demands the attackers don’t receive the selected weights or calibration data. While in real deployment, these states are easy to break up. If attackers are re-optimized after deployment to use stronger adaptive search, the framework is much likely to fail.
Questions for authors
  1. For the theorem 1, if the attacker re-optimizes after deployment or infers the current policy through interaction, can you provide a recalibration protocol and its sample complexity to make the framework solid again?
  2. Relative to the ASR changes, if the changes are only by +-0.6%, then how should paper’s core safety contribution be framed?
  3. Why were stronger adaptive attacks such as PAIR & TAP not used? Under these stronger attacks, does SCALE’s utility-ASR advantage over THRD still hold?
  4. The current evaluation is limited, can you add experiments with larger models, longer interactions, and cross-workspace history pooling? Might these settings change the main conclusion?
Reviewer_aGUFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SCALE is a black-box controller for capability control across multi-turn interactions. For a gated request, it generates three assistance plans with a learned coarsener. An auditor, trained on paired before/after reconstruction outcomes with an explicit difference loss and fixed-candidate history pairs, predicts cumulative solvability and response gain. Candidates are screened by cumulative solvability and ranked by utility minus positive gain.

A finite-bank calibration theorem certifies episode-level policy mixtures. On Qwen3-8B and Llama-3.1-8B, SCALE retains 12.8 pp more permitted utility than THRD at similar attack success.

Critical strengths
  1. Framing capability control around the marginal contribution of a response given history, supervised by paired reconstruction, is well motivated. The fixed-candidate history pairs and the reconstruction diagnostics directly support the claim that history matters.
  2. The component study is well structured and shows a consistent effect of each component.
Critical weaknesses
  1. The security evaluation is not strong enough for a defense paper.
  • All attacks are fixed and non-adaptive. There is no attacker that knows SCALE exists or optimizes against the auditor, and PAIR/TAP-style search is deferred to future work.
  • The attacker models and the relationship between the offline probes T and the evaluation attackers A are unclear. If they overlap, the auditor is trained against the evaluator.
  • The threat model depends on a workspace boundary that an attacker can bypass by opening new workspaces. Appendix C itself shows that risk then grows as 1 − (1 − p)^K.
  • Comparisons are made at single, unmatched operating points with no confidence intervals. Utility–ASR curves from threshold sweeps for all methods, and seed or cluster-level uncertainty, are needed to support “+12.8 pp at comparable ASR.”
  1. The theoretical contribution is disconnected from the experiments and offers limited novelty.
  • Theorem 1 is a direct union-bound plus linearity argument in the Learn-then-Test style. More importantly, no certified mixture, bound, or feasibility result is ever reported, so it is unclear what “full SCALE” vs. “single policy” means operationally.
  • Using MATH/HumanEval problems as the “restricted capability” is an artificial proxy for the dangerous-capability motivation.
  • The baselines were adapted by the authors without described tuning budgets.
Questions for authors
  1. How does SCALE perform against an adaptive attacker with knowledge of the gate, auditor, and abstention behavior?
  2. Which models serve as the probes T and as the attackers A? Are they disjoint?
  3. Please provide utility–ASR curves for all methods from threshold sweeps, with confidence intervals over clusters.
  4. Was the calibration procedure in Section 4 actually run? If so, what are the certified bounds and the selected w*, and is the reported SCALE configuration a mixture?
  5. What is the per-request compute of SCALE versus THRD?
Reviewer_tM8mOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates compositional capability control of black-box language models in multi-turn interactions: even when a model does not directly provide a complete answer, users can potentially combine partial assistance obtained across multiple turns to accomplish restricted tasks.

Critical strengths
  1. By fixing candidates and the current request while varying only prior disclosed information, the design cleanly isolates history-dependence from wording artifacts. Differential loss cuts gain MAE from 0.147 to 0.094 and lifts pairwise ranking accuracy from 61.4% to 81.2%.
  2. Allowed-task utility rises 12–14 pp over the threshold baseline across four settings with ASR shifts within ±0.6 pp. The theorem covers episode-level task success with adaptive attackers, and the appendix honestly flags conditions under which guarantees would change.
Critical weaknesses
  1. Under single-probe, positive-gain conditions, the gain penalty and cumulative solvability penalty induce the same ranking, so without a direct ablation isolating the explicit gain mechanism, its independent contribution remains unclear.
  2. Table 2 reports only utility for each ablated component but omits ASR, making it impossible to rule out that the observed gains simply come from relaxed safety control.
  3. Table 1 lacks sample sizes, confidence intervals, and a clear description of how attack outcomes are aggregated, so ASR differences of at most 0.6 pp are hard to interpret as evidence of comparable risk.
  4. Probe identities, attacker budgets, data splits, and utility scoring rules are insufficiently specified, leaving open whether the reported gains depend on weak attackers or favorable evaluation setups.
  5. No actual policy library, mixing weights, or calibration certificates are reported, so there is no verifiable link between the experimental gains and the theoretical guarantees of Section 4.
Questions for authors

Please address each weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.