acceptodds
Under review as a conference paper at ICLR 2027

Why Static Probes Fail: Context Determines Deception Feature Composition

Abstract

With increasing autonomy, Large Language Models (LLMs) may engage in deceptive behaviors, such as scheming to pursue misaligned goals or sandbagging to escape oversight. Monitoring outputs alone is insufficient, as models can produce seemingly benign responses while harboring misaligned internal objectives. While mechanistic approaches like linear probing are promising, they rely on the assumption that deception corresponds to a single direction in representation space. We demonstrate that such single-direction probes fail to generalize across diverse deception scenarios. To bridge this gap, we propose representing deception as a multi-dimensional subspace rather than a single direction. Our method utilizes iterative orthogonal decomposition to identify complementary directions that capture distinct facets of deceptive intent. To effectively leverage these directions for detection, we find that static combinations prove brittle due to strong domain dependence across different scenarios. We therefore introduce a domain-adaptive gating mechanism that dynamically composes the extracted directions through few-shot adaptation at test time. With only 10 examples, our approach achieves an average AUROC of 0.97 and 94% recall at 5% FPR across all evaluated datasets, substantially outperforming existing methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deception Detection, Representation Subspaces, Linear Probing, Domain Adaptive Gating, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Q1q3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the generalization limitations of static linear probes for detecting deceptive behavior in LLMs. The authors argue that deception cannot be adequately represented by a single universal direction in activation space because different deceptive behaviors may involve different internal representations. They propose Dynamic Deception Subspace Navigation (DDSN), which includes Iterative Orthogonal Subspace Decomposition (IOSD), where previously identified directions from the activation space are removed to extract multiple orthogonal directions associated with deception-related features. DDSN also includes Domain-Adaptive Gating, which is a lightweight MLP dynamically combines the extracted directions. The empirical results are promising, and the idea of combining multiple deception-related directions through domain-adaptive gating is interesting.

Critical strengths
  • Investigating internal representations is a promising direction for developing monitoring systems, especially for models operating in increasingly autonomous settings. Representing deception through multiple complementary directions is a reasonable extension of conventional single-direction probing.

  • The authors motivate this approach by showing that simple linear probes exhibit substantial performance degradation across heterogeneous deception benchmarks. The paper evaluates several deception scenarios across two main language model architectures and reports generally strong AUROC improvements. On Llama-3.3-70B-Instruct, DDSN improves substantially over the static linear probe on the more challenging MASK, PKU-DeceptionBench, and THU-DeceptionBench datasets.

  • The framework freezes the underlying LLM and the extracted deception directions, adapting only a small gating network using target-domain examples.

Critical weaknesses

I have some concerns about the correctness of the theoretical claims, particularly Theorem 4.2, as well as the novelty of the method, the fairness of comparisons against simpler few-shot adaptation methods, and the interpretation of false-positive rates.

  • Theorem 4.2 use a combination of linear functions to approximate a non-linear function. However, a fixed linear combination of linear projections is itself a linear function. Independently of how many orthogonal directions are extracted, this formulation cannot approximate arbitrary nonlinear functions. The appendix attempts to establish the theorem by arguing that the projected-feature matrix can achieve full row rank. However, this condition is not guaranteed by IOSD. For N samples and activation dimension D, when N>D, the proposed argument cannot establish full row rank. The paper uses this theorem to justify the capacity of IOSD to capture complex nonlinear deceptive patterns. The theorem, as stated, does not support that conclusion.

  • Theorem 4.1 claims that the extracted directions capture decreasing amounts of mutual information about deception labels. However, the implemented IOSD procedure extracts directions by minimizing a regularized logistic regression objective under orthogonality constraints. However, the proof in Appendix C reformulates IOSD as directly maximizing mutual information over successive constrained spaces. These two objectives are not generally equivalent. The proof therefore establishes a property of an idealized optimization procedure rather than the actual method.

Questions for authors
  1. How is Theorem 4.2 valid? Can the authors clarify how arbitrary nonlinear functions can be approximated and justify the full-rank assumption used in Appendix D?

  2. Theorem 4.1 is proved for mutual-information maximization, whereas IOSD is implemented using regularized logistic regression. What guarantees the claimed mutual-information ordering for the actual optimization procedure?

  3. Could the authors clarify the orthogonal projection formula in Eq. (7), including the matrix dimensions, direction normalization, and exact residual-update implementation?

  4. How much of the performance improvement comes from IOSD versus target-domain few-shot adaptation? Can the authors compare with an adapted linear classifier over the same projections and a matched-capacity MLP trained on the original activations?

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

The paper proposes DDSN, combining orthogonal probing directions with few-shot adaptive gating to improve cross-domain detection of LLM deception. The empirical results are promising, although the sources of improvement and claims about internal representations require further validation.

Critical strengths
  1. Relevant problem: The paper addresses cross-domain generalization and false positives, extending evaluation beyond in-distribution accuracy.
  2. Reasonable design: Fixing a shared subspace while adapting its composition provides a plausible inductive bias for few-shot detection.
Critical weaknesses
  1. Insufficient isolation of contributions: The method changes source data, representations, and classifier architecture simultaneously. Matched comparisons with standard MLPs and fixed-weight combinations are needed to identify the sources of improvement.
  2. Substantive theoretical gaps: The mutual-information proof does not match the implemented logistic regression objective, while fixed combinations of linear projections do not establish general nonlinear approximation.
Questions for authors
  1. Under matched training data, target-domain examples, and tuning budgets, how much does each component—orthogonal decomposition, dynamic gating, and target-domain adaptation—contribute?
  2. Can the theoretical claims be revised to match the implemented algorithm and explicitly state the additional assumptions required?
  3. How much improvement remains after controlling for shortcut cues such as topic and wording, and fixing the target-domain false-positive rate at 5%?
Reviewer_LEy9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies cross-context generalization in LLM deception detection and proposes Dynamic Deception Subspace Navigation (DDSN). The method uses iterative orthogonal decomposition to extract complementary deception-related directions from model activations, then adapts a gating network with 10 labeled examples per target domain to combine their scores dynamically for each input. The authors evaluate DDSN on seven deception benchmarks using Llama-3.3-70B-Instruct and Qwen3-32B, comparing it with linear probes, statistical methods, and black-box judges. They report average AUROCs of 0.97 and 0.93, respectively. Cross-dataset transfer experiments and ablations on the number of directions, activation extraction layer, and gating function further examine the method’s adaptability and robustness.

Critical strengths

The paper addresses an important practical problem: deception detectors can fail when transferred across contexts. Results across several benchmarks show that a single probe can have high AUROC yet produce many false positives at the operating threshold. This makes the paper’s focus on detection reliability valuable.

Combining multiple directions with few-shot adaptation is a promising approach. On the evaluated benchmarks, adapting the gate with just 10 labeled examples per target domain improves detection and substantially reduces the false positives of static linear probes in some settings.

Critical weaknesses

The two main theoretical claims are not adequately supported. The proof of Theorem 4.1 assumes mutual-information maximization, whereas the algorithm optimizes logistic-regression loss. The proof of Theorem 4.2 relies on an unguaranteed full-row-rank condition and does not establish the stated general nonlinear approximation claim. These gaps weaken the theoretical justification for IOSD and require substantive revision.

The experiments do not adequately isolate the benefit of dynamic gating. DDSN uses two source datasets, while the primary linear-probe baseline uses one, and a static multi-direction baseline trained with the same data and adaptation budget is missing. The reported gains therefore cannot be clearly attributed to dynamic gating rather than multiple features or additional source supervision, leaving the central methodological contribution insufficiently validated.

Questions for authors

What connects the mutual-information maximization assumed in the proof of Theorem 4.1 to the logistic-regression objective used by the algorithm? Does IOSD guarantee the full-row-rank condition required by the proof of Theorem 4.2? If not, how should the theorem be qualified?

With identical source data and the same 10 labeled target-domain examples, does DDSN still substantially outperform a refitted linear probe and a learned fixed-weight combination of the same IOSD directions?

Which adaptation loss was used in the experiments: the squared error in Equation (10) or BCEWithLogits in the appendix? Also, how was the average recall of 0.69 in Table 9 calculated? The seven listed values average to approximately 0.53.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.