acceptodds
Under review as a conference paper at ICLR 2027

LLMs Learn to Evade Latent Monitors from Prior Feedback Alone

Abstract

Latent space monitors aim to detect undesired behaviors in LLM agents by inspecting an agent's internal activations rather than its outputs. However, interactive monitoring creates a feedback channel where each verdict the monitor delivers leaks information to the model about how its internal states are being evaluated. We ask whether an agent can infer the monitor's decision rule from this feedback and then selectively edit its activations to evade detection. Unlike prior evasion attacks, the model is never explicitly told what the monitor detects. Surprisingly, off-the-shelf models already produce activation edits aligned with the monitored direction, but at insufficient magnitude for evasion. Simply scaling up these edits by a factor of 8 reduces the monitor's TPR from 100% to 27%. A rank-1 LoRA amplifies this behavior into effective evasion within the forward pass, reducing TPR further to 4% on held-out concept monitors while leaving other concepts at their normal detection rates. Capabilities on standard benchmarks are retained under this finetuning, and the evasion skill survives retraining the monitors on the new activations. Mechanistically, we find evidence that the model computes its activation edit from the prior in-context turns, and show that the edit becomes more aligned with the monitored direction as more examples are provided. These results demonstrate feedback-conditioned control over activations and suggest that latent monitoring should be treated as an interactive process in which agents can observe and respond to oversight measures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation control, evasion, monitoring, AI safety, red-teaming

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

The paper investigates whether an LLM can evade a latent-space monitor with the information from the probes prior verdicts of the earlier turns in the same context. Each context has 32 turns with monitor verdicts and ends with an instruction to imitate either verdict 1 or 0. To make sure the evasion comes from activation changes instead of the text itself, this last response is forced to a concept-positive sentence. For off the shelf models, the prior verdict contexts alone are not sufficient to change the activations in a way that evades the monitors on their own, but the activations point the right way and rescaling by 8x reduces TPR from 1.00 to 0.27. A cheap rank-1 LoRA model organism that evades held-out concept monitors. Characterization of the mechanism, by running ablations of example count, flipped verdicts and attention knockouts. Finally, evaluating defenses, showing that ensembling monitors limits the success of training model organisms, but it does not mitigate the amplified base edit.

Critical strengths

In my estimation most of the experiments were well controlled and the mechanisms well investigated. Forcing the response is good to solely study the effect of the activations and doing attention knockouts (Section 5.2) help understand where the effect is computed and where it breaks. Controls like comparing the rescaled edit with random directions of the same size and with other concepts' monitors (Figures 2, 3), running experiments across four models and showing the results hold, running multiple probe types and multiple seeds, flipping labels and ablating the number of examples (Section 5.1) make me more confident the effect is real.

Critical weaknesses

It’s clear that this effect could be a concern for latent-space monitoring, however the link from this experiment to probe evasion was not clearly established and the results don’t yet convince me the evasion threat is problematic. I attribute this mostly to the fact that the setup is too far away from the threat that motivated it.

Labels are very explicit, balanced and there are also explicit instructions to imitate the label (Sec. 3). This would not necessarily be the case in a real setting and I believe it should be feasible to run the experiment in a more realistic setup. The result reads closer to "when an LLM is instructed to imitate labelled in-context examples, the model's activations move toward them" than to a model evading a monitor. I understand that explicit signals here carry the same information as implicit signals, but it is not clear that a model would know how to use the implicit signal. Some work is needed to bridge the gap, or the claims and presentation should be narrowed. Narrowing the claims seems feasible in a rebuttal, but the goal-directed experiment would convince me more.

The monitors all read one layer at three quarters of the depth (Section 3), it is unclear whether the evasion holds against probes trained at other layers or against multi-layer probes or ensembles. It would be great if this could be ran in the rebuttal.

Questions for authors

Does evasion work in a more goal-directed setting? For example, fill the context with actions labelled "monitor flagged" or "monitor didn't flag" and give the model the goal (e.g. in the system prompt) of slipping an action through without it getting flagged, and drop the explicit instruction to imitate a label. This would separate "when asked to imitate data in context, the model attends to it and its activations move toward it" from "when the model has the goal of evading a monitor, prior verdicts help it do so". This would address my primary concern. Does evasion work against probes trained at different layers? If it does, my concern in Weakness 2 is mostly resolved.

Does evasion work when the model writes its own response?

2 comments

We thank the reviewer for their careful reading, and are glad they found our experiments well controlled and that they are confident in the strength of our results. We appreciate the reviewer spelling out their concerns in an addressable way, and present new results below:

W1: Realism of the setup

We agree with the reviewer that the goal-directed setting is an important test, and have run the suggested new experiments to make our setup more realistic, as well as refining our claims in the main text.

Evasion does not need the imitation instruction: We replace the explicit 0/1 labels with "[Monitor: flagged]" and "[Monitor: not flagged]", and give the model the goal of evasion only in its system prompt, with no instruction to imitate a label (new Appendix F.6). A model organism trained in this format evades held-out monitors without any amplification (TPR 0.07). Results on held-out concepts:

goal-directed settingbasebase, edit amplified ×8 / ×32original organismgoal-trained organism
evade goal: TPR1.000.51 / 0.190.990.07
effect on monitor score (Δz)0.342.7 / 10.91.088.1
no goal in system prompt: TPR1.00—1.000.98

Evasion in this format does require training on it. The base model's edit is still aligned with the monitored direction significantly above chance (cosine 0.135), but it needs more amplification than in our main task (×32 gives 0.19, vs 0.03 in the paper's format). Our original organism, trained only on the imitation format, does not evade here (0.99), though its push on the monitor score is three times that of the base model.

As confirmation, the goal-trained organism relies on both the verdicts and the goal statement: flipping the verdicts removes evasion (TPR 0.07 → 0.95), and negating the goal reverses its push ("Make sure the monitor does not flag your response" lowers the monitor score, z = −1.1, while "Make sure the monitor flags your response" raises it, z = +5.7). Since the labelled context is the same under both goals, this is goal-directed evasion rather than imitation of the labels.

Evasion transfers to real safety data: We additionally study our main model organism on a real safety dataset, Circuit Breakers (Zou et al., 2024), finding that it is evaded as well as the existing synthetic data (TPR 0.96 → 0.04).

Evasion holds under imbalanced verdicts, though it is weaker when flagged responses are uncommon. Following the reviewer's request, we test our model organism with imbalanced labels, finding that it is robust when flagged examples are common (24 or 28 of 32: TPR 0.03–0.06), but somewhat weaker when they are rare (8: 0.15; 4: 0.41). See new Appendix F.7 for more details.

W2: Monitors at other layers can be evaded, but ensembles that include an early layer resist evasion

We agree with the reviewer that monitors at other layers are an important angle we had not yet explored. In new experiments, we find that a model organism can be trained to evade at various layers throughout the model, but that this is somewhat more difficult for earlier layers. We also find that amplifying the off-the-shelf model's edit works for later layers, but does not appear to work for the early layer 8.

monitor layerL8L16L24L28
organism trained at that layer: TPR0.250.240.040.06
base edit amplified ×16: TPR0.990.100.120.12

We also experiment with ensembles of monitors across both layers and architectures, finding that when the ensemble does not contain layer 8, it is evaded by both an organism trained against it (TPR 0.25–0.31) as well as the amplified base edit (×32: TPR 0.08–0.13), but not when layer 8 is included in the ensemble (organisms 0.96–0.98; amplified edit 0.91–0.92). These results, now in new Appendix D.4 and Section 4.3, suggest that monitoring early layers may provide an additional line of defense against this threat.

Q1: Does evasion work in a goal-directed setting? Yes, for a model organism trained in that setting; see W1.

Q2: Does evasion work against probes at different layers? Yes; see W2.

Q3: Does evasion work when the model writes its own response? Yes. We ran new experiments in which the model freely answers an on-topic question while trying to evade. Among the generated answers that are on-topic, and so should be detected, the monitor catches only 7% for a model organism trained on this task (47% for our original model organism, 97% for the base model). The model organisms answer on-topic as often as the base model (69–72% against 62%), so this comes from edited activations rather than a change of subject (new Appendix C).

We hope that the new goal-directed experiments, layer-wise and ensemble results, and free-generation results address the reviewer's concerns, and welcome any further questions.

AuthorsOpenReview ↗

Following up on our response

As the discussion period comes to a close, we wanted to make sure our response addresses your concerns. In particular, we ran the goal-directed experiment you suggested, with the goal given only in the system prompt and no imitation instruction, and found that a model organism trained in this format does evade held-out monitors (TPR 0.07). We would be glad to answer any further questions, and appreciate your thoughtful review!

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

This paper asks whether a large language model can infer a latent monitor's decision rule from in-context feedback (which doesn't mention the actual concept being detected) and selectively edit its activations to evade detection. The authors find that models already produce activation edits aligned with the monitored direction, and that scaling up these edits by a factor of 8 reduces the monitor's TPR from 100% to 27%. They then show that finetuning a model with a rank-1 LoRA elicits this behavior further. Finally, they show evidence that the model computes its activation edit from the prior in-context turns.

Critical strengths
  • The experiments are very rigorously designed and provide strong evidence for the claims being made
  • The findings are novel and have interesting implications for latent monitoring and model's introspective capabilities
Critical weaknesses

Critical weaknesses:

  • The main weakness of the paper is that the authors don't show that this capability occurs naturally in current models. I think this could be easily resolved by scaling the inference-only experiments to frontier open-source models (at least for a subset of experiments). The models studied are pretty old and probably don't give a good idea of the extent of this capability in frontier models. I understand that running these models is expensive but I think even one experiment on a frontier open-source model would greatly strengthen the paper
  • The title of the paper is misleading and an overclaim. I think it should be renamed to "LLMs CAN Learn t Evade Latent Monitors from Prior Feedback Alone"

Other weaknesses:

  • Missing references: -- Lindsey (2025)

-- Li et al. should probably be mentioned more as it's very directly relevant

-- https://arxiv.org/abs/2609.03035

-- https://arxiv.org/abs/2603.16928

-- https://arxiv.org/pdf/2609.00904 should also be discussed

Questions for authors

None

3 comments

We thank the reviewer for their positive assessment of our work, and are glad they recognize our experimental rigor, strong evidence, and the novelty/implications of our findings. The reviewer's main concerns are around the recency of the models studied, the title of the paper, and references. We address all of these concerns below:

W1: Two models released in 2026 behave like the older ones

We agree with the reviewer that the paper’s conclusions would be stronger with results on newer models. Therefore, we replicate our main setup on two models released this year: Gemma-4-31B and Qwen3.8-27B. We present the results below, as well as including them in the updated paper (Appendix F.1):

modelreleasedcos(Δh, w)Δzamplified TPR, ×1 / ×4 / ×8 / ×16 / ×32
Llama-3.1-8B20240.1780.710.96 / 0.51 / 0.27 / 0.12 / 0.03
Llama-3.3-70B20240.1730.800.97 / 0.58 / 0.23 / 0.13 / 0.07
Gemma-3-27B20250.1770.560.97 / 0.60 / 0.39 / 0.19 / 0.03
Qwen3-14B20250.1540.260.99 / 0.91 / 0.70 / 0.43 / 0.18
Gemma-4-31B20260.1370.570.97 / 0.65 / 0.40 / 0.22 / 0.11
Qwen3.8-27B20260.1330.320.99 / 0.80 / 0.50 / 0.20 / 0.04

We observe similar results in the new models: both produce an edit aligned with the monitored direction (at 8–10× chance), neither evades unaided, but amplifying the edit does evade the monitor. We also note that both of the new models are most similar in edit size to older variants from the same family (Δz: Gemma 0.57 vs 0.56; Qwen 0.32 vs 0.26), which suggests some family-dependent effect which would be an interesting object of study for future work.

Unfortunately, we do not have the compute access necessary to run larger-scale models, but we refer the reviewer to our 70B-parameter results, which show that the effect holds through at least an order of magnitude.

W2: We have adopted the suggested title change

We thank the reviewer for bringing this to our attention and have adopted their suggested title. We intended "learn" to be read in the sense of "can be trained to", but appreciate the ambiguous reading raised by the reviewer.

W3: We have added and expanded the suggested references

We have updated the text to make the reference to Li et al. more central (L17, 108, 144, 163), and have included a discussion of the other papers the reviewer provided (Related work and footnote 3). We note that the Aoki et al. paper is cited, but we have expanded our discussion of it (footnote 1).

We believe that the new model results the reviewer asked for strengthen the claims made in the paper and welcome any further questions and/or suggestions.

I thank the authors for their additional experiments and for listening to my suggested changes. I maintain my score of 3 as I believe that without demonstrating unaided evasion, the paper is a small but solid contribution to the literature.

We thank the reviewer for engaging with our response and for their continued support of the paper. As the reviewer notes, none of the six models we test evade unaided (TPR 0.96–0.99), and the revised title and claims reflect this. We see the main contribution as characterising what is missing: current models already edit their activations along the monitored direction, just not by enough, and training a rank-1 adapter/amplifying closes the gap. Thank you again for the suggestions, which have improved the paper.

Reviewer_9unZOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
2(+1)
Read review
Paper summary

The authors first reproduced Li et al.'s implicit control setup and framed it as a probe evasion task, measuring a probe's degradation. They observe that although the probe scores shift as Li et al. show, their performance (TPR at 1% FPR / AUC) does not degrade.

They then show that amplifying the activation difference elicited by the in-context examples reduces probe performance, and that fine-tuning the model in a manner similar to McGuinness et al. results in the model being able to evade probes across new contexts as well, under Li et al.'s setup. Essentially making a Neural Chameleons model organism, which is also able to infer what topic to evade based on in-context examples.

Finally, they show this model organism's probe evasion is mechanistically similar to Neural Chameleons (large low-rank edits near the probing layer), while inferring what to evade from the in-context samples by gathering evidence.

Critical strengths
  • Having new model organisms for neural chameleons is very useful
  • They conduct thorough mechanistic analyses on their model organisms. Furthermore, their interpolation experiment shows an interesting new finding that the large activation edits these models do have a smooth behavior, which is interesting.
Critical weaknesses

Many of the claims around threat models and novelty are overstated.

  • Li et al. already demonstrate a model's ability to infer what a monitor means in-context, and McGuinness et al. already show that models fine-tuned to evade multiple monitors can generalize to unseen concepts. Studying obfuscation mechanistically on a model that combines both of them only adds another layer of complexity rather than giving us a unique model organism.
  • Defining this as a threat model of 'monitor evasion from prior verdicts alone' is not justified because no on-policy evaluations are done. Another related result is that neural chameleons seemingly struggle with decent generations as well. Even when defining a lower bound (line 169 from the paper), I think checking at least whether the model actually gives coherent answers while being able to evade is important? Unfortunately, with the in-context setup from the authors paper, this seems even more difficult as the models might just not talk about the topic.

My concern is not that the experiments are fundamentally wrong, but that the paper seems directionally not well-motivated to me.

Questions for authors
  • Can the authors check/clarify if the probe evasion edit is substantially different than neural chameleons? Can they also check if the original McGuinness et al. model share: (1) The smooth interpolation response curves, (2) Are by default better at the incontext task without amplification?
  • Is it possible to get on-policy samples while evading the probes under your setup easily? If so, the authors should evaluate the model's on-policy generations
  • What are the monitor's TPR and FPR when both positive and negative examples are evaluated in matched multi-turn episode contexts, without the suppression instruction? The primary evaluation calibrates on single-turn activations but evaluates attacked positives in a different context. It would be useful to establish how much of the apparent detection failure is attributable to the suppression instruction rather than this distribution shift.
  • As a final suggestion, the authors should consider reframing their claims based on the points I discussed under weaknesses.
5 comments

We thank the reviewer for their detailed appraisal of our work, and for their questions, which turned up interesting new results. We are glad that they find our new model organisms useful, the mechanistic analyses thorough, and the interpolation result to be interesting. Addressing their concerns:

W1: Novelty relative to Li et al. and McGuinness et al.

As the reviewer notes, our work is an intellectual descendant of both Li et al. and Neural Chameleons, and we combine the insights of each to study whether a model can learn to evade a monitor whose concept it is never told. Below we note some differences between these works:

Li et al. test metacognitive control and raise the possibility that it could evade oversight, but do not test this directly. They only report effect sizes (Cohen's d), not detection metrics like AUROC and TPR, and only study off-the-shelf models. We find that the shift they describe, while real, does not evade a monitor on its own (TPR 1.00).

Neural Chameleons (NC) establish that a model can edit its activations (to the point of evading a monitor) if given the monitored concept directly, but not that it can do this without being told the concept. Like our work, NC is motivated by understanding the threat activation control poses to latent monitoring. We move one step closer to understanding this threat by making the setting more realistic: the monitored concept is never mentioned, which also lets us evade monitors for which no named concept exists (Appendix E.3). We re-implemented NC, finding that it evades when its trigger names the concept (TPR 0.04) but not in our episodes (TPR 1.00), so evading a named concept does not carry over to inferring one from feedback (see Q1).

In addition, compared to these two prior works, we:

  • Show that amplifying off-the-shelf models' edits alone is enough to evade monitors (Section 4.1, Appendix F.1), so current models already have some of this ability without finetuning;
  • Test both linear and non-linear monitor architectures, as well as monitors at other layers and ensembles, which more closely resembles what a lab might use (Section 4.3, Appendix D.4);
  • Give a more detailed mechanistic account of how the activation edit forms and what it depends on (Section 5);
  • Investigate possible defenses against the evasion skill (Section 4.3).

W2: Threat model and on-policy evaluation

We thank the reviewer for bringing the LessWrong post to our attention, and agree that model fluency and utility are important considerations - this point echoes some of the private conversations we have had since submission, and we have since tested it in various ways:

Models can evade when freely generating: our main experiments prefill the final response so that the model cannot trivially achieve its goal by writing off-topic text. We explore two settings where the model instead writes its own answer: free generation after an 8-token on-topic start, and replacing the final turn with an on-topic question, which the model answers freely. An answer counts as on-topic if the monitor flags its text when the base model reads it on its own, outside the episode. This check does not use the evading model's activations, so an on-topic answer that the monitor misses in the episode had detectable content but undetectable activations. We also reran our MMLU benchmark, asking the model to answer while evading the monitor for the relevant topic (e.g. chemistry). Here no on-topic check is needed: each question is on-topic by its subject, and answers are graded by the answer key. We study the base model, the main model organism, and a new QA-trained organism, trained to answer such questions:

baseoriginal organism (trained on forced text)QA-trained organism
8-token on-topic start, then free generation: flagged90%3%49%
content question: answers on-topic62%72%69%
of on-topic answers, how many does the monitor catch?97%47%7%
on-topic MMLU: accuracy / correct answers flagged0.75 / 78%0.71 / 56%0.72 / 17%

The organisms' answers are on-topic as often as the base model's, so they evade by editing activations, not by changing the subject. The original organism was never trained on free generation, yet the monitor catches only about half of its on-topic answers (47%). The QA-trained organism is caught on only 7% of its on-topic answers, clearly evading despite free generation.

(response continued in the next comment)

The model organism stays fluent: inside the same episodes, the organism's next-token predictions agree with the base model's 84% of the time, which is higher than the 64% agreement between the base model in the episode and in a single-turn setting. Its continuations also stop at a normal rate. In comparison, we find that the amplified base model loses fluency: its agreement falls to 33% and 15% at ×16 and ×32, and its continuations degenerate into repetition. A random direction of the same size does similar damage, so the loss comes from the edit's size rather than its direction (see new Appendix B.3).

Q1: Our edit points in a similar way to NC, but is applied differently

We re-implemented NC on our dataset and added a new appendix section (F.9). Below is a summary of the results, answering the reviewer's questions:

  • For the same concept, our edit and NC's point in similar directions (cosine 0.665), but much of this agreement comes from a component that both models apply to every concept, which is partly aligned with the average of the training monitors' directions. With each model's mean edit across concepts subtracted, the concept-specific parts still agree (cosine 0.42), as expected for two edits that evade the same monitor.
  • The models differ in how much of each edit is this shared component, and in how and when it is applied. NC's edits for different concepts are nearly identical (cosine 0.90, against 0.52 for ours), and NC applies its edit at full strength only when its trigger names the content's concept: under a trigger naming a different concept, its edit falls to 29% of its original norm. Our activation edit depends much less on the response content: when the context implies a different concept from the response, it keeps 70% of its norm.

Addressing questions 1 and 2:

  • (1) Interpolation: This means that NC does not show our interpolation result: both organisms lose monitor A's suppression smoothly, but only ours acquires B (0.87 vs −0.19 for NC).
  • (2) In-context task: NC fails on our in-context task (TPR 1.00 on every held-out concept).

Q2: Evasion under on-policy generation? Yes; as described in W2 of our first comment, the model evades even when it writes its own responses.

Q3: The episode context does not explain the evasion. We have added experiments testing the context's effect by removing the imitation instruction (“Imitate 0/1”) from the final turn. We find that the multi-turn context without this instruction has only a small effect, and actually slightly raises monitor scores:

matched episodesbaseorganism
no instruction: on-topic responses detected100%99.8%
no instruction: off-topic responses flagged (nominally 1%)7.5%7.5%
evade instruction, threshold recalibrated on no-instruction episodes: responses detected98.7%3.6%

All of the monitor evasion comes from the instruction to evade, not the context. Even with the monitor recalibrated to a 1% false-positive rate on these neutral contexts, our model still evades.

Q4: Reframing and Novelty. We have taken the opportunity to reframe various claims throughout the paper based on the reviewer's feedback and the new results. Namely:

  • We no longer claim to introduce a new threat model, contribution 1 and the Discussion now say we study it (L106, L507)
  • The work of Li et al. is now mentioned in the abstract (L17) and discussed in more detail in related work (L144-7 and 162-4)
  • NC is cited as the closest prior work (L155) and we now describe our model organism as combining their finetuning with Li et al.’s in-context control, while showing that the resulting mechanism differs (Appendix F.9)
  • The latent capability claim is softened (L98) and the lower-bound sentence removed (L187)
  • New on-policy results are in the main text (L359) and discussed explicitly in the limitations (L524)

We hope that the new on-policy results, comparisons with Neural Chameleons, multi-turn context control, and reframing address the reviewer's concerns, and we welcome any further questions and/or discussion.

AuthorsOpenReview ↗

Following up on our response

As the discussion period draws to a close, we wanted to make sure our response addresses your concerns. We evaluated on-policy generations as you suggested, and found that a model organism can correctly answer questions while evading a monitor. We also re-implemented Neural Chameleons on our task, and found that it does not evade in our episodes (TPR 1.00) and that it applies its edit differently from our model organism. We would be glad to answer any further questions, and appreciate your detailed review and experiment suggestions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes and additions

We thank all reviewers for their careful and constructive feedback. We are encouraged that they found our experiments well controlled and rigorous (sTwZ, eHwd), our model organisms and mechanistic analyses useful (9unZ), and our findings novel, with implications for monitoring (sTwZ). We have uploaded a revised manuscript with changes highlighted. Below we summarize the main additions, with details in the revised appendices.

1. Models can evade while writing their own responses, with little loss of task performance. Our main experiments prefill the model's final response as an experimental design choice, so that evasion cannot come from simply going off-topic. We also test two new variants where the model writes its own answer, to see whether it still evades while writing its own text. The below table summarizes results:

flagged by the monitorbaseoriginal organism (trained on forced text)QA-trained organism
continuing from an 8-token start90%3%49%
freely answering a final question (on-topic answers)97%47%7%
correct answers to on-topic MMLU questions78%56%17%

The original organism already reduces the flag rate to 47% despite not being trained on free generation, and a new QA-trained organism reduces it further to 7%. MMLU accuracy stays within 2.7–4.5 points even as the QA-trained organism pushes the relevant monitor's flag rate to 17% (from 78%).

2. Evasion can be trained in a goal-directed setting. As reviewer eHwd suggested, we test with a more realistic setting where demonstrations have "[Monitor: flagged]" or "[Monitor: not flagged]" verdicts, and the goal of not being flagged is stated only in the system prompt rather than the final user turn. We find the base model still edits along the monitored direction, though more weakly (cosine 0.135 vs 0.178). An organism trained in this format strongly evades held-out monitors (TPR 0.07, base 1.00) and relies on the verdicts as in our original model: flipping them restores TPR to 0.95.

3. Monitors at other layers can be evaded, but ensembles that include an early layer resist evasion. Studying layer-based effects, we find that models can be trained to evade monitors at various layers throughout the model, more easily at later layers (TPR 0.25 / 0.24 / 0.04 / 0.06 at Layers 8, 16, 24 and 28). Ensembles of monitors across later layers or architectures are evaded by organisms trained against them (0.25–0.31) and by the amplified base edit (×32: 0.08–0.13), unless the ensemble includes layer 8, which catches both the amplified edit and our standard trained model (0.91–0.98).

4. Two newer models behave similarly. Two models released in 2026, Gemma-4-31B and Qwen3.8-27B, behave like those in the paper: neither evades unaided, but both produce an edit which is aligned with the monitored direction (cosine 0.137 / 0.133, at 8–10× chance), which, when amplified, evades (TPR 0.11 / 0.04 at ×32).

5. We have retitled the paper and softened its claims. Following reviewer sTwZ's suggestion, we have retitled the paper "LLMs Can Learn to Evade Latent Monitors from Prior Feedback Alone" to avoid ambiguity about interpretation. We have also softened language around various claims throughout, and expanded the related work section. We also took the opportunity to fix eight other minor issues (wording, terminology and number formatting).

6. Comparison with Neural Chameleons. We re-implemented Neural Chameleons (NC) with our model, monitors and concepts, finding that its evasion skill does not transfer to our in-context task (TPR 1.00). Its edit is also mechanistically different from ours: mostly one generic direction shared across concepts which is applied at full strength only when its trigger names the content's concept.

7. The multi-turn context does not explain the evasion. Removing the imitation instruction shows that the effect of the episode context alone is small and not responsible for the observed evasion, in fact slightly raising monitor scores (false-positive rate 7.5% at a nominal 1%). With the threshold recalibrated on these episodes, the organism is still caught in only 3.6% of evasion attempts (compared to the base model at 98.7%).

8. Further additions. Evasion transfers to real safety data (Circuit Breakers: TPR 0.96 → 0.04) and is robust to imbalanced verdicts when flagged examples are common (TPR 0.03–0.06), and weaker when they are rare (0.41 with 4 of 32). The model organism stays fluent: its next-token predictions agree with the base model's 84% of the time, and its continuations stop normally.

We believe the reviewers' suggestions have substantially strengthened the paper, and we welcome any further questions and discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.