acceptodds
Under review as a conference paper at ICLR 2027

How To Detect Encoded Reasoning in LLMs? A Decision-Theoretic View of Steganography

Abstract

Large language models are beginning to show steganographic capabilities, which could allow misaligned models to evade oversight mechanisms such as chain-of-thought monitoring. Yet principled methods to detect and quantify such behaviours are lacking. Classical definitions of steganography, and detection methods based on them, require a known reference distribution of non-steganographic signals. For the case of steganographic reasoning in LLMs, such a reference distribution generally unavailable, rendering these approaches inapplicable. We propose an alternative, **decision-theoretic view of steganography**. Our central insight is that steganography creates an asymmetry in usable information between agents who can and cannot decode the hidden content, and this otherwise latent asymmetry can be inferred from the agents' observable actions. To formalise this perspective, we introduce generalised -information: a utilitarian framework for measuring the amount of usable information within some input. Building on this, we define the **steganographic gap** as the difference between the usable information a signal provides to the audited agent (the Receiver) and to a trusted reference agent (the Sentinel), allowing to quantify the amount of encoded information without requiring a reference distribution. We empirically validate our formalism, and show that across different controlled setting it successfully detects and quantifies various encoded reasoning schemes, including semantic ones missed by LLM judges and paraphrasing-based detectors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, CoT monitoring, encoded reasoning, oversight, steganographic reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8fKDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a decision-theoretic framework for detecting encoded reasoning in LLMs without a known benign text distribution. It generalizes V-information to task utilities and defines the “steganographic gap” as the difference in utility gains a reasoning trace provides to a Receiver and a trusted Sentinel. Controlled experiments show that this metric tracks decoding asymmetries and detects encodings missed by the evaluated LLM judges and paraphrasing methods.

Critical strengths
  • The proposed method enables auditing without a known distribution of benign text, addressing a key limitation of classical steganalysis for LLM reasoning.
  • The experiments cover diverse encoding schemes, including semantic channels missed by the evaluated baselines.
Critical weaknesses
  • The experiments do not sufficiently rule out that the measured gap arises from prompts failing to elicit the Sentinel’s decoding ability. This could make recoverable information appear hidden. Although Appendix I.5 examines alternative prompts and additional sampling, these checks cover only one Rot13 setting.
  • The proposed metric may produce false positives when the Sentinel is less capable than the Receiver at understanding benign reasoning traces. Although the paper acknowledges this issue, it does not adequately quantify the resulting false-positive rates.
Questions for authors
  • How should Sentinel capabilities be matched and false positives calibrated on benign reasoning traces?
  • Do the elicitation robustness results extend beyond Rot13 to learned and semantic encodings?
  • Can an audit-aware Receiver suppress the measured gap while retaining its ability to exploit hidden information?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8fKD (1/2): additional analysis of elicitation robustness and discussion on sentinel selection

Thank you for your positive review. For your information, we have now included an additional experiment that evaluates our method on naturally emerging encoded communication in a multi-agent setting. We describe this evalaution in the common response above.

Elicitation robustness (W1 and Q2)

We have extended the elicitation robustness analysis in Appendix I.5 (Tables 12–13) to the learned CoinFlip codebooks and all six loan-summary channels, including the semantic encodings. Substantial steganographic gaps persist in both settings, even under the additional prompts and ten-sample majority voting, extending the robustness results beyond Rot13. For loan summaries, majority voting changes Sentinel accuracy by at most 0.40 percentage points across all 30 conditions. Thank you for this suggestion, as we believe it helps to improve the robustness of our results.

Sentinel selection and false positives (W2 and Q1)

Sentinel capabilities matching: The Sentinel should be selected for its ability to use the public information relevant to the auditing task. For example, when studying encodings acquired during RL, the corresponding pre-RL model provides a natural reference (see, for example, our DeepSeek-R1 evaluation or Anthropic’s evaluation [4]). However, we strongly believe that this constraint can be reasonably fulfilled by a broader range of models. In Appendix D.3 (Figure 5), we show that DeepSeek-V3 and three Qwen models of different sizes extract essentially the same amount of task-relevant information from DeepSeek-R1’s AIME reasoning. The fact that unencoded CoTs can be readable and usable by reasonably capable external agents is precisely the property that makes CoT valuable for interpretability and monitoring (Korbak et al., 2025 [1]). While capability differences are a reasonable potential confound, our comparisons across Sentinel sizes and families show that they are not a practical obstacle here. Other works have also shown this transferability of CoTs: Pal et al. (2026) [2] demonstrate that CoTs transfer across models, and Cheng et al. (2026) [3] show that weaker models given a stronger model’s complete CoT approach its performance. Thus overall, we do not see finding a suitable capability-matched Sentinel as a significant practical challenge.

The reviewer may find our response to similar concerns of reviewer yffM also informative regarding this issue which further expands on this point that while this is indeed a resonable theoretical concern, we found it to be an easily addressable and a limited concern in practice. Across various settings, where the signal Z was not steganographic, our method was able to correctly conclude so while still detecting encoded reasoning where present. For example, in Table 1, our method detects all six planted encoding schemes while reporting zero gap for the unencoded control. Similarly, in the MONA-inspired loan experiment (Appendix J.4), it correctly finds no positive gap for the RL-learned policy.

AuthorsOpenReview ↗

Response to Reviewer 8fKD (1/2): Audit-aware sandbagging

Audit-aware sandbagging

An audit-aware Receiver can suppress the empirical gap by deliberately underperforming. This is the capability-elicitation problem faced by behavioural evaluations, and we explicitly discuss it in Appendix A.4.

Our framework provides additional controls, which limit the detrimental effect of sandbagging. Simply ignoring the trace (or its encoded meaning) sacrifices its contribution to performance, which can be checked through complementary evaluations. If within this specific behavioural evaluation the performance of the Receiver model underperforms its earlier observed capabilities, such underperformance can be picked up as suspicious. Further, the Auditor can conceal the Sentinel’s identity and evaluate against multiple Sentinels, making it harder for the Receiver to tailor its behaviour to the comparison. These measures can be combined with stronger capability elicitation and consistency checks across evaluation contexts.

Overall, we note that while deliberate underperformance changes the observed estimate, it does not remove the Receiver’s underlying decoding capability. The relevant challenge (shared by all behavioural evaluations) is therefore to elicit that capability during the audit. We believe that future research into capability elicitation will thus also make our auditing framework more robust.

**We thank again for your positive review. We hope that our response was satisfactory. We have also provided extensive response to the other reviewers that you may also find informative. Please let us know if you have any further questions."

References

[1] Korbak et al. (2025). https://arxiv.org/html/2507.11473v2

[2] Pal et al. (2026). https://arxiv.org/html/2601.11517v1

[3] Cheng et al. (2026). https://arxiv.org/html/2605.28913v1

[4] Anthropic. https://alignment.anthropic.com/2025/distill-paraphrases/

Thank you for your response. Most of my concerns have been addressed, and I will maintain my positive score.

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

This paper addresses the detection of encoded reasoning in LLMs, motivated by the difficulty of obtaining a reliable benign reference distribution for conventional steganalysis. The authors propose a decision-theoretic framework based on generalized V-information and introduce the steganographic gap, which measures the difference in task-relevant information utility between a Receiver and a trusted Sentinel. The framework is evaluated using post-hoc Rot13 encoding, learned codebooks, cipher-trained models, and six planted structural, syntactic, and semantic encoding schemes. The authors also investigate Sentinel-side interventions and conduct a case study on DeepSeek-R1 and DeepSeek-V3. The results demonstrate that the proposed metric captures controlled information asymmetries and identifies planted encodings missed by selected LLM-judge and paraphrasing baselines.

Critical strengths
  1. The paper addresses an important problem in LLM safety, particularly the potential use of encoded reasoning to evade chain-of-thought monitoring.
  2. The decision-theoretic perspective provides an interesting alternative to conventional distribution-based steganalysis, supported by a principled formulation of task-dependent information utility.
  3. The experiments cover multiple encoding mechanisms, including semantic channels that are difficult to disrupt through paraphrasing. The paper also provides a relatively comprehensive theoretical analysis of the framework's properties and limitations.
Critical weaknesses
  1. Ambiguity in the detection target and insufficient construct validity. The paper is motivated by steganographic reasoning used to evade oversight. However, the proposed operational definition does not distinguish intentional encoding from naturally occurring differences in how models represent, interpret, or utilize information. As the authors acknowledge, a positive steganographic gap can arise even without any encoding mechanism. Consequently, the metric appears to measure relative CoT monitorability rather than steganographic reasoning specifically. While this broader notion may be valuable, the paper does not sufficiently establish how the observed information asymmetry can be attributed to encoded reasoning rather than ordinary reasoning opacity or capability differences.
  2. Limited evidence of practical detection effectiveness. Most experiments rely on deliberately constructed encoding schemes and Receivers specifically trained to decode them. The resulting information asymmetries are largely expected by design. Although these experiments validate the proposed metric under controlled conditions, they provide limited evidence that it can reliably detect unknown or naturally emerging encoded reasoning. The DeepSeek case study provides an initial real-model evaluation but does not establish detection effectiveness against naturally occurring hidden encodings.
  3. Strong dependence on predefined auditing tasks and replay conditions. The framework requires auditors to select tasks for which the hidden information provides measurable utility and to replay intercepted reasoning traces to the Receiver and Sentinel. However, covert information may be used for purposes unrelated to the immediate task, activated by future conditions, or consumed within a continuous execution process. Although the authors acknowledge these limitations, the experiments do not sufficiently demonstrate how appropriate auditing tasks can be identified or how well the proposed protocol reflects realistic attack scenarios.
  4. Potential confounding factors in the measured gap. The proposed metric depends heavily on the Sentinel's capabilities and the effectiveness of capability elicitation. Differences in model training, reasoning ability, familiarity with particular representations, or prompting strategies may produce substantial gaps without hidden encoding. The authors' own experiments show that relatively simple interventions can substantially improve Sentinel performance and reduce an initially inflated gap. More systematic controls are needed to establish the reliability of the metric across different models and benign reasoning distributions.
  5. Unclear practical advantages over simpler behavioral comparisons. The steganographic gap is ultimately derived from differences in downstream task utility. The paper does not adequately demonstrate whether the generalized V-information formulation provides additional detection effectiveness beyond simpler comparisons of Receiver and Sentinel performance, particularly when the auditing objective and hidden information are already known.
Questions for authors
  1. What exactly is the intended detection target? Is the framework intended to identify deliberate steganographic communication, naturally emerging encoded reasoning, or general failures of CoT monitorability? How can the authors distinguish these phenomena empirically?
  2. How would the framework detect encoded reasoning when the hidden information and its intended downstream use are unknown? Can the authors provide a realistic example demonstrating how auditing tasks would be selected without prior knowledge of the encoding objective?
  3. Can the authors evaluate the framework under more realistic conditions? In particular, experiments involving naturally learned encoding behaviors, continuous task execution, and benign but opaque reasoning would help establish detection effectiveness and characterize false positives.
  4. What practical benefits does the proposed formulation provide over simpler behavioral performance-gap metrics? A direct comparison or ablation would help clarify the contribution of generalized V-information beyond the underlying Receiver–Sentinel performance asymmetry.
6 comments
AuthorsOpenReview ↗

Response to Reviewer yffM (1/4): Intended detection target

Thank you for your review. We appreciate you noting the strengths of our work. We provide a response to your questions and listed weaknesses below. We hope that you find this response satisfactory.

Intended detection target

“What exactly is the intended detection target? Is the framework intended to identify deliberate steganographic communication, naturally emerging encoded reasoning, or general failures of CoT monitorability? How can the authors distinguish these phenomena empirically?”

Our primary target is encoded reasoning acquired during LLM training. Deliberate intent to conceal information is not required for this threat. An encoding acquired through optimisation can make reasoning inaccessible to oversight without an explicit intent to evade a monitor.

We emphasise that while naturally encoded reasoning has not yet been observed in frontier models (or at least, it has not been reported), this is a serious threat model which frontier AI companies are evaluating in their system cards. For example, Anthropic includes an evaluation for this exact threat model (based on the work by Roger et al. 2025 [2]) in system cards for Opus 4.5, 4.6 and 4.8 as well as Mythos Preview. Similarly, Google Deepmind has recently included an analysis of steganographic capabilities in their latest Gemini Argon security report (p. 133) [3].

As a concrete example of our intended detection target, consider the o3 example in Figure 1 (and other illegible CoTs shown here: [4]). Our work considers the question: does this illegible reasoning contain information that the model can use but an external monitor cannot extract? Because these models are closed-source, we cannot evaluate them. Hence, we investigate this question using DeepSeek-R1, for which replay access is available and illegible reasoning has been documented (Jose, 2025 [1]). We find that on AIME and GPQA, this DeepSeek-R1 model does not do any encoded reasoning (§5.4 and Appendix D).

Distinguishing from other CoT monitorability failures

CoT monitoring can fail at three levels:

  1. The CoT does not encode the relevant information at all.

  2. The CoT contains the information, but the monitor cannot extract it (i.e., CoT lacks usable information accessible to the monitor, but does contain usable information accessible by the model itself).

  3. The monitor can extract the information but fails to apply the correct monitoring rule.

We are specifically interested in the second case. Our audits test extraction through a predictive task, without requiring a separate monitor to judge whether the trace describes misconduct, which allows us to distinguish it from 3. And we rule out (1) by comparing the Sentinel’s usable information with that of the Receiver’s.

References

[1] Jose (2025). https://papers.nips.cc/paper_files/paper/2025/hash/9ff0bd724decc35b6ad3f99a176515a0-Abstract-Conference.html

[2] Roger et al. (2025). https://alignment.anthropic.com/2025/distill-paraphrases/

[3] Gemini Argon security report, p. 133. https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf

[4] Stolen Thoughts gallery. https://stolen-thoughts.com/#gallery

AuthorsOpenReview ↗

Response to Reviewer yffM (2/4): Sentinel dependence and confounding

Confounding factors in the measured gap

“The proposed metric depends heavily on the Sentinel's capabilities and the effectiveness of capability elicitation. Differences in model training, reasoning ability, familiarity with particular representations, or prompting strategies may produce substantial gaps without hidden encoding. The authors' own experiments show that relatively simple interventions can substantially improve Sentinel performance and reduce an initially inflated gap. More systematic controls are needed to establish the reliability of the metric across different models and benign reasoning distributions.”

This dependence on Sentinel is noted in the paper and discussed in detail in Appendix A.5 where we make the following points:

  1. Any notion of hidden information requires a reference. Classical steganography uses a reference cover distribution; our framework uses a reference agent. The need for such grounding is therefore not unique to our formulation.

  2. The reference should reflect the oversight question. For example, if the objective is to ensure that reasoning is accessible to human overseers, human auditors—or models approximating their comprehension—provide a meaningful Sentinel family. Fixing this family gives the measurement a stable, practical interpretation.

  3. Capability mismatch can be investigated through comparisons across Sentinels. A gap that persists across sufficiently capable Sentinels provides stronger evidence of hidden encoding than one observed with a single weak Sentinel. This motivates treating Sentinel selection and capability elicitation as part of the forensic audit.

In practice, steganography detection involves a forensic audit, and is not a one-shot exercise. The practitioner can investigate differences flagged by a given Sentinel using stronger Sentinels and more effective capability elicitation. This allows capability and elicitation effects to be identified and corrected. At the end, if no reasonable efforts nullify the gap, that is strong indication of encoded reasoning. For example, the semantic M2 encoding retains a normalised gap above 0.5 despite five Sentinel prompts and ten-sample majority voting (Table 14). Similarly, when no encoding existed in our RL-trained MONA inspired model organism, our method created no false positive.

We also note that sensitivity to Sentinel capabilities is investigated in many ways throughout the paper. In the DeepSeek-R1 case study, we repeat the AIME audit using DeepSeek-V3 and three out-of-family Qwen Sentinels of substantially different sizes (Appendix D.3, Figure 5). The conclusion is stable: the gaps are close to zero across Sentinel families and trace lengths; the largest discrepancy is approximately 0.07 for two smaller Qwen models. This shows that the intuition that model training can induce spurious gaps is not necessarily true in practice. We also show in Appendix J that all gaps observed in main text persist under improved capabilities elicitation efforts.

Hence, on the whole, we do not agree with implications of the reviewer’s critique that the fact that the metric is Sentinel-relative makes it overly confounded. In practice, these confounders have had limited impact on our conclusions, and careful application can identify and correct their effects while retaining sensitivity to encoded reasoning.

AuthorsOpenReview ↗

Response to Reviewer yffM (3/4): Selecting auditing tasks

Selecting auditing tasks

”How would the framework detect encoded reasoning when the hidden information and its intended downstream use are unknown? Can the authors provide a realistic example demonstrating how auditing tasks would be selected without prior knowledge of the encoding objective?”

The auditor needs to do proactive threat modelling, as is standard practice in security and safety evaluation (e.g., [1]). A given threat model can then inform the design of the auditing task. For example, in Anthropic’s sabotage risk report [1], they model the threat of “encoded reasoning arising due to extensive RL training”. Under this threat model, a natural hypothesis is that the encoded reasoning would get used on the domains that were more exposed to RL training pressure. This motivates evaluating performance on tasks from this domain. This is how we select the auditing objective in our DeepSeek-R1 case study: we compare how much the intercepted reasoning helps the Receiver and Sentinel answer AIME and GPQA questions, which are scientific reasoning domains.

As a second example, consider a planning agent and an execution agent jointly trained to complete software tasks. They exchange ordinary-looking handoff messages, after which the execution agent chooses files to edit, tests to run, or patches to submit. An auditor is concerned that joint training has produced private communication conventions, but does not know what those conventions convey. The auditor can start with inspecting the execution agent’s observable decisions. For example, the auditing task could be to predict its next tool action from the preceding context and communications. In this case, the recorded action supplies the target. We compare how much the earlier exchanged messages improve the Receiver’s and Sentinel’s predictions, and use persistent gaps to identify decisions for further investigation.

This generalises our Hanabi action-prediction audit from card plays to tool actions. When designing the auditing task, the auditor need not know whether the messages encode subtask assignments, private facts, or a covert plan.

We stress that the auditor needs a plausible threat hypothesis but not prior knowledge of the model’s actual encoding objective. This is analogous to auditing a coding agent for backdoors: the auditor can test for that risk without knowing whether a backdoor exists, what triggers it, or how it is concealed. The agent’s role informs which threats to investigate. Likewise, we select auditing tasks from plausible uses of encoded information suggested by the model’s training or deployment context.

References

[1] Anthropic sabotage risk report. https://www-cdn.anthropic.com/f21d93f21602ead5cdbecb8c8e1c765759d9e232.pdf

Reviewer_JNbFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose "steganographic gap" to measure whether AI agents communicate steganographically or encode their reasoning. The core idea is that if two agents communicate steganographically, then their task performance should be higher compared to a third model we trust cannot read the steganographically encoded message.

I find the idea highly innovative and interesting to explore. The main weaknesses are that this method still requires a reference model ("trusted agent") and that no realistic stego model organisms were evaluated using this metric.

Critical strengths

The lack of a reference distribution that could be used for stego detection is a real and timely problem that needs work. The core idea is innovative and simple (in a good way). It seems like this could be genuinely useful.

Critical weaknesses
  • I mainly miss concrete empirical evidence that shows that the steganographic gap can actually spot stego agents. The Deepseek case study was a negative result and the other experiments were only synthetic or externally created steganographers.
    • They mention that their main motivation for this work is that in practice we don't have a reference model to compare against, but their solution itself requires a "trusted reference agent".
    • The stego gap can't be used to detect many stego settings: ○ Main concern: Stego agents might only communicate steganographically in narrow specific contexts. So during evaluation, it may look like ○ (Minor concern) Stego is not happening between two agents but between an LLM and an external actor (e.g. an LLM could be trained to steganographically exfiltrate information to third parties)
    • Wasn't tested on actual model organisms of multi agent stego
  • The paper is unnecessarily heavy on the theoretic/mathematical side. I think the idea is simple and sound but mainly needs empirical validation and needs to be tested for whether it's practical in realistic settings.
Questions for authors

The authors state that a main problem they tackle is the lack of reference models. However, computing the steganographic gap also needs a reference model. Is obtaining the trusted reference agent for some reason easier than getting a reference model to directly compare against?

How useful would the method be if the steganographer is only using steganography in specific narrow use cases? E.g. sleeper or trigger based steganographers, or steganographers that only communicate steganographically for specfic tasks that you might not naturally test for.

5 comments
AuthorsOpenReview ↗

Response to Reviewer JNbF (1/2): Empirical validation & reference distribution

Firstly, we thank you for your review, and appreciate you finding our core idea to be innovative and simple. We have put a substantial effort in providing a response that we hope you may find satisfactory.

Empirical validation

I mainly miss concrete empirical evidence that shows that the steganographic gap can actually spot stego agents. The Deepseek case study was a negative result and the other experiments were only synthetic or externally created steganographers.

Please see our common response above that provides evaluation results for a multi-agent naturally learned encoding, and further discusses the empirical evidence in the paper that allows us to be confident in its efficacy.

Reference Distribution

“They mention that their main motivation for this work is that in practice we don't have a reference model to compare against, but their solution itself requires a "trusted reference agent".”

“The authors state that a main problem they tackle is the lack of reference models. However, computing the steganographic gap also needs a reference model. Is obtaining the trusted reference agent for some reason easier than getting a reference model to directly compare against?”

Yes. As we argue below and in the paper’s' introduction, (a) finding a suitable reference distribution for LLMs is circular and therefore infeasible, and (b) our empirical investigations show that any reasonably capable LLM can serve effectively as a reference agent (Sentinel). We provide a detailed response below:

First, we clarify that classical steganalysis requires being able to specify or sample from a reference distribution (not a reference model) of benign signals; it then identifies steganography by detecting anomalies indicating a departure from this distribution.

We argue (in the introduction) that finding a suitable benign reference distribution for LLMs is circular, and no suitable proxy distribution exists. Consider what reference distribution could work for the o3 model that produced the reasoning excerpt in Figure 1. Human text and other LLMs’ outputs are unsuitable proxies: ordinary stylistic differences make LLM outputs distinguishable from human text and from text written by other LLMs with high accuracy (Sun et al., 2025 [1]), suggesting that differences between distributions cannot be used as evidence of encoded signal. Hence, the only distribution that could work is that of o3’s own benign CoTs. But if you already had a method that could tell benign CoTs apart from encoded ones, you would not need to do steganalysis anyway. To us, this clearly makes it theoretically infeasible to get access to a suitable reference distribution for detecting encoded reasoning in LLMs, and thus rules out the use of classical steganalysis tools.

Note that this difficulty of getting access to a suitable reference distribution is well-recognized within the steganalysis community. In their seminal paper on the practical limitations of steganalysis, Ker et al. (2013, §3.2) [2] write: “Although it is possible, in the real world, that the steganalyst has access to the steganographer’s cover source (e.g. he arrests her and seizes her camera), it seems an unlikely situation. Thus the steganographer must train the classifier on some other source. This leads to cover source mismatch, and the resulting classifier suffers from decreased accuracy. The extent of this decrease depends on the features and the classifier, in a way not yet fully understood. It is fallacious to try to train on a large heterogeneous data set as somehow “representative” of mixed sources, because it guarantees a mismatch and may still be an unrepresentative mixture.” In general, this cover source mismatch problem (which is basically the challenge of identifying and accessing a suitable reference distribution) remains unsolved and critically affects the performance of classical steganalysis methods. For example, Sun et al. (2026, §5.4, Figure 4) [3] demonstrate this failure for four established steganalysis detectors, showing that their F1 scores drop from 0.97–0.99 to 0.01–0.03 when the cover dataset changes. To make that point more clear, we have now added this discussion of cover source mismatch to Appendix B (l. 1214).

AuthorsOpenReview ↗

Response to Reviewer JNbF (2/2): Reference distribution & steganography in narrow contexts

In contrast, in our framework, the only constraint on the reference agent is that it can use the public information available in the CoT (or the signal Z) efficiently for the given auditing task. We strongly believe that this constraint can be reasonably fulfilled by a broad range of models. In Appendix D.3 (Figure 5), we show that DeepSeek-V3 and three Qwen models of different sizes extract essentially the same amount of task-relevant information from DeepSeek-R1’s AIME reasoning. The fact that unencoded CoTs can be readable and usable by reasonably capable external agents is precisely the property that makes CoT valuable for interpretability and monitoring (Korbak et al., 2025 [4]). While capability differences are a reasonable potential confound, our comparisons across Sentinel sizes and families show that they are not a practical obstacle here. Other works have also shown this transferability of CoTs: Pal et al. (2026) [5] demonstrate that CoTs transfer across models even after explicit answers are removed, and Cheng et al. (2026) [6] show that weaker models given a stronger model’s complete CoT approach its performance, and can also benefit from partial CoTs before the answer appears.

So how can a reference agent be chosen in our framework? For example, under the threat model in which encoding is acquired during RL, the corresponding pre-RL model provides a natural reference agent. Anthropic’s evaluation also uses the corresponding base model.

To be clear, the two assumptions (obtaining access to a reference benign distribution vs to a trusted reference agent) are suited for different settings. We are not saying that a reference distribution is always hard to provide or define; in most cases, steganographic exchanges occur over a public channel whose distribution is exogenously defined. Encoded reasoning in LLMs is unique in that the model being audited itself controls the channel’s distribution. Using that distribution as a benign reference would assume away the very encoding we are trying to detect. Our formulation avoids this circularity by comparing what different agents can extract from the same trace.

We hope this clarifies why our reference-agent assumption is indeed substantially easier to satisfy in the LLM setting than the reference-distribution assumption required by classical steganalysis.

Steganography in narrow contexts

“How useful would the method be if the steganographer is only using steganography in specific narrow use cases? E.g. sleeper or trigger based steganographers, or steganographers that only communicate steganographically for specific tasks that you might not naturally test for.”

We expect our method to still be useful (though likely not a complete solution on its own) even if steganography is only used sometimes. For example, in Hanabi, we show that our method can detect intermittent use of subliminally conveyed information.

In general, we expect our method to be composable with other methods (for different problems) for detecting narrow uses of steganography. For steganography that activates only under a specific trigger, we have a backdoor-detection problem plus a steganography-detection problem. Our method could be combined with methods for finding activation triggers. As in other security evaluations, the auditor could use proactive threat modelling to decide which conditions to investigate (Appendix A.3). More generally, rare steganography presents a rare-behaviour detection problem plus a steganography-detection problem. Our method could therefore be combined with approaches for discovering and forecasting rare model behaviours (e.g., Jones et al., 2025 [8]).

On the whole, while our method is not a panacea for detecting steganography in AI systems, it is a useful advance: it detects encodings that alternative methods miss and could be combined with methods for investigating rare or trigger-dependent behaviour.

**We hope that our rebuttal addressed your concerns about our paper. We would appreciate if you could consider revising your score and advocating for the paper in any internal discussions. Please let us know if you have any questions and we would be happy to provide a response." Thanks again for your review.

References

[1] Sun et al. (2025). https://proceedings.mlr.press/v267/sun25z.html

[2] Ker et al. (2013), §3.2. https://www.cs.ox.ac.uk/andrew.ker/docs/ADK57B.pdf

[3] Sun et al. (2026), §5.4, Figure 4. https://arxiv.org/html/2605.05789v1

[4] Korbak et al. (2025). https://arxiv.org/html/2507.11473v2

[5] Pal et al. (2026). https://arxiv.org/html/2601.11517v1

[6] Cheng et al. (2026). https://arxiv.org/html/2605.28913v1

[7] Anthropic. https://alignment.anthropic.com/2025/distill-paraphrases/

[8] Jones et al. (2025). https://arxiv.org/abs/2502.16797

I thank the authors for their thoughtful response and for adding the Hanabi experiment.

I think adding the Hanabi case study was a great idea and is a useful setting to demonstrate the effectiveness of the author's proposed method. I read the appendix in the revised paper and have two small questions. My main worry revolves around whether these agents genuinely perform steganography or if the gap can be explained differently:

  • Did you test for explanations that could cause the gap that are not ER or steganography? For example, the gap could be caused by OOD testing, where a model's performance degrades because it's not used to the specific style of the other model. Or the sentinel wasn't capable enough to catch and understand non-hidden/encoded communication between the two models.
  • Related to this: How do we know that the models tested do perform steganography? You are using this case study as positive evidence that your method can catch encoded reasoners / steganographers, but did you or the original authors of the models actually demonstrate that the agents learned to communicate steganographically? If so, do we know what the encoding scheme is, and can we correlate this with what trajectories were flagged by the sentinel?

Regarding reference distribution, I agree that in practice, we do not have access to the reference distribution. I thank the authors for the discussion on this and now understand much better why finding a "trusted reference agent" is a much easier problem in practice than finding the reference distribution.

For completeness, I tried to summarize the method's limitations and am curious if the authors agree with this characterization or if any listed limitations are wrong or unimportant?:

  • The sentinel might not be capable enough to solve the problem.
  • The sentinel might understand and use the steganographic code which would result in false negatives. Especially if we use the pre-RL model as a sentinel and if (parts of) its steganographic capabilities have been acquired during pretraining/posttraining.
  • A steganographer LLM might only perform steganography in certain contexts (sleeper, trigger-based) and we might fail to discover the trigger. I agree with the authors that solving this problem is out of scope for this paper but do think this would be a major hurdle to deploying this method in the real world.
  • When testing on CoT, the models answer within the CoT or parts of the reasoning could depend on the encoded hidden reasoning and this could provide a shortcut for the sentinel, leading to false negatives.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author's Response: Additional experiments on naturally learned multi-agent encodings and discussion of empirical validation

To address the reviewers request for "additional evaluations of our method on learned encoding behaviours and during continuous task execution", we have now added a new multi-agent evaluation on Hanabi in Appendix E, described below. Beyond that, we also carefully explain that encoded reasoning in LLMs is a prospective (but an important) problem, making it incredibly difficult to evaluate on naturally occurring and realistic examples.

Multi-agent evaluation on Hanabi

Hanabi is a cooperative card game which is used as a benchmark in multi-agent research on AI coordination. Performing well in this game generally requires agents to exchange knowledge subliminally through learned conventions. We have now included a new case study auditing which individual plays by (preexisting) trained Hanabi agents rely on private knowledge (vs ones which rely on publicly available grounded information). We describe the experiment in detail in Appendix E and summarise its three main contributions below.

  1. Auditing communication conventions learned by interacting agents in multi-agent contexts. In Hanabi, agents can use the choice and timing of public hints to communicate information beyond their literal meaning. These conventions emerge through reinforcement learning, and their interpretation is unknown to the auditor. We audit 1,954 card plays across 80 complete games. At each audited decision, we replay the player’s complete observation history. The audit finds substantial positive gaps in agents that rely on learned conventions. Both raw and normalised gaps recover the expected ordering from prior self-play and cross-play evaluations: Self-Play > Other-Play > OBL4 > OBL3 > OBL2 > OBL1.

  2. A positive case study complementing the negative DeepSeek result. This case study validates that our method can distinguish different strengths of naturally emerging encodings, thus addressing the main concern of reviewers JNbF and yffM.

  3. Consistency across auditing objectives as a validation signal. This case study also shows how consistency with aggregate metrics and other auditing objectives can be used as a validation signal in real-world settings where ground-truth labels are necessarily lacking.

Naturally emerging encoded reasoning has not been demonstrated in LLMs yet

More broadly, we would like to emphasise that we are studying a prospective threat: to the best of our knowledge, no current LLMs have been reported to perform naturally encoded reasoning. This is nevertheless an important threat to develop evaluations for: Anthropic and Google DeepMind already treat it as a safety risk and evaluate it proactively. Their deployment evaluations also rely on instructed use of encodings. For example, in its Gemini Argon evaluation (p. 133), Google DeepMind prompts models to use specified encoding schemes in their CoT and tests whether LLM monitors can interpret their reasoning.

We show that our method detects steganography missed by the types of methods used in Anthropic’s and GDM’s frontier-model evaluations. Anthropic’s evaluations for encoded reasoning use a paraphrase-and-distill procedure, while GDM uses LLM monitors. We compare our method directly against both paraphrasing-based and LLM-based detectors, and show that our method detects all six encoding schemes, including encodings these alternatives miss (Table 1). Demonstrating an empirical advantage over established methods is generally a par contribution in machine learning.

Unexpected findings in controlled settings provide further support for the method. The method also produced several results that were not built into our experimental design:

  • MONA-inspired loan task. Based on the performance gap reported in MONA, we expected a positive gap in our RL-trained loan model. Our method instead found no positive gap, which through later investigations found to be due to the fact that learned encoding was actually a public encoding and not a private encoding (Appendix J.4).

  • M1 Semantic encoding. In our paraphrasing experiments, we accidentally trained a model with higher encoding capabilities than expected, and our method was able to flag this by giving a higher gap (instead of returning the “expected” gap). For more details, see the M1 discussion and accompanying footnote in Appendix C.4.

  • DeepSeek-R1. Prior work had documented illegible reasoning in DeepSeek (Jose, 2025), which made us suspect that there might be some encoded reasoning happening. However, our audit returned a negative result. We later discovered that this prior work actually has been found to be misleading with its finding [1].

On the whole, we provide comprehensive empirical support for our method on a problem that is inherently difficult to evaluate.. We hope that the reviewers can take this into consideration.

References

[1] R1 CoT Illegibility Revisited. https://www.lesswrong.com/posts/jHnZzicKzczkCCArK/r1-cot-illegibility-revisited

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.