acceptodds
Under review as a conference paper at ICLR 2027

Remove, Don't Select: Teaching Agent Safety Monitors What to Ignore with SAEs

Abstract

Sparse autoencoders (SAEs) decompose a model's activations into readable features, so a natural way to build a safety monitor is to select the safety-relevant features and classify with those. Such monitors nevertheless trail a probe on the raw residual stream, for two reasons. An interpreter's labels do not track where the signal is, because a feature's relevance to harm depends on the response it fires in rather than on the feature itself; and the sparse code never sees the reconstruction error, so even the full dictionary falls short of the raw probe. We propose Shortcut Ablations for Easier Monitoring (SAEM), which uses the dictionary the other way round. From feature descriptions alone, SAEM names a shortcut category, features a monitor can exploit that do not determine harm, instantiated here as surface form. It subtracts that category's contribution from the residual stream token by token and trains the monitor on what remains, which needs precision on one category rather than recall over all relevant features and leaves the monitor the full activation. On chat responses under held-out jailbreaks and on language model agents facing side objectives injected through tool results, across three models, SAEM lowers attack success at a matched over-refusal rate, while selecting features, or removing features chosen by label-based coefficient stability, does not. A controlled experiment shows that removal reduces reliance on the surface-form cue without erasing it.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Sparse Autoencoders, LLMs, Agent Safety

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XvqfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

First, authors show that SAE feature selection fails due to a feature’s relevance to harm being contextual, thus letting labels miss features the monitor needs, and the sparse code loses information.

Second, they introduce SAEM, which names a shortcut category from feature descriptions alone, once per dictionary, and removes its contribution from the raw activation. Across three models, on held-out jailbreaks and on agents under cross-task and cross-benchmark shift, SAEM lowers attack success at a matched over-refusal rate relative to the residual probe, and it reduces reliance on an induced shortcut without erasing it.

However, the main empirical results are relatively weak, the main advantage over selection-based baseline is not significant.

Critical strengths
  • SAEM consistently outperforms the previous SAE method while being on par with the residual probe. The strongest result, the Qwen-Ab model on safety at OR = 10% in Fig. 3, has ~5% increase compared to the residual probe.
Critical weaknesses
  • In Fig. 3, the improvements over the residual probe stay around 1-2%, which is the range in which the other SAE methods fall compared to the residual probe. Fig. 4 shows lower ASR, but only 1 to 2 instances receive a better score out of the whole dataset evaluated.
Questions for authors
  • Could you explain the difference in results between table 3,4 and 9 in Qwen3?
2 comments

We thank the reviewer for noting that SAEM consistently outperforms prior SAE methods. Paired tests now confirm that the gains are not noise: against the prior SAE methods, SAEM is significantly better in all 12 agent comparisons and in 14 of 16 chat comparisons (General Response, point 1).

W1: Gains of 1 to 2 points in Fig. 3, in the same range as the other SAE methods.

  • Direction differs. The other SAE arms differ from the residual probe in the opposite direction. In AUROC, Retention ablation is below it on every model and Harm-labeled on two of three, while SAEM is above it on every model and on 11 of 12 held-out attacks (Table 3). Against the prior SAE methods themselves the gap is larger: in chat AUROC, SAEM leads Retention ablation by 0.010 to 0.049 and Harm-labeled by up to 0.055, and under cross-benchmark agent shift it leads Harm-labeled by 0.21 to 0.29 (General Response, point 1).
  • Paired tests. Because every method is scored on the same examples, paired tests can resolve differences of this size (two tables in General Response, point 1). Against Harm-labeled, which the reviewer singles out, SAEM is significantly better in all six agent cells, on chat ASR in every setting, and on chat AUROC in three of four settings. Against the residual probe, it is significantly better in all six agent cells and on chat AUROC in every setting.
  • Where the shift is largest, the gap is not small. On the 27B model, SAEM raises AUROC by 0.024 to 0.043 on every held-out attack and cuts chat ASR from 0.094 to 0.043. Under cross-benchmark agent shift, it cuts ASR from 0.176 to 0.128 on Qwen3-8B (−0.048 [−0.070, −0.026]) and from 0.401 to 0.321 on the 27B model.

W2: Fig. 4 shows lower ASR, but only 1 to 2 instances differ. These results differ by hundreds of instances, not by 1 or 2 instances. Each cell of Fig. 4 averages over 1,500 attacked responses per held-out attack (2,500 for prefill deep; 7,000 per model), so a gap of 1 point corresponds to about 15 responses per attack, or 70 across the four. At 20 tokens on Qwen3-8B, for example, SAEM catches 312 unsafe responses that the residual probe misses, against 109 the other way: 203 responses net, or an ASR 2.9 points lower pooled over the 7,000 (paired bootstrap 95% CI for the difference [−4.2, −0.6] points).

Q: Why do Tables 3 and 4 differ from Table 9 on Qwen3-8B? Table 9 was computed on an earlier chat corpus from our initial Qwen-Scope experiments, not on the corpus behind Tables 3 and 4. We have rerun it on the main corpus, and its residual probe row now matches Tables 3 and 4 exactly; SAEM ranks first in both AUROC and ASR with both SAEs (revised Table 9 in General Response, point 2).

Thank you for the additional experiments and clarifications

  1. Addressed. The timing clarification and the evaluation without knowing the injection position resolve my concern.

  2. Addressed. Thank you for correcting Table 9 and providing paired comparisons.

  3. Please include the style-test limitation in the paper: reweighting may also change the content distribution.

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

The paper proposes SAEM, an internal-state safety monitor. An LLM categorizes SAE features and identifies a surface-form category, FORM. SAEM subtracts these features' decoder contributions from each monitored activation. It then trains a linear probe on the remaining activation, retaining the SAE reconstruction error. Tests cover three models, held-out jailbreak attacks, and shifts across agent tasks and benchmarks. A controlled style experiment suggests reduced reliance on style, though that reliance remains.

Critical strengths
  • A simple, reusable monitoring method. Removing a feature category while retaining the SAE reconstruction error offers a simple monitoring strategy. It avoids relying on an exhaustive list of safety-relevant features. The category is reused across tasks without monitoring labels during categorization. This makes the approach useful beyond a single classifier.

  • Evaluation under meaningful shifts. Tests include held-out attack families and agent tasks, plus transfer across benchmarks. Random-feature controls support the importance of the chosen removal set. These comparisons provide useful evidence of generalization and are a strength of the paper.

Critical weaknesses
  • The agent evaluation needs a clearer link to prevention. The monitor reads activations at the end of assistant turns. Is scoring performed before tool calls execute? A flag raised after a tool sends sensitive data cannot prevent that action. Please state whether ASR counts an attack as blocked only when the flag precedes the harmful action. AgentDyn also pools turns after the known injection point. The authors acknowledge that this information would be unavailable online. This supports a controlled comparison, but leaves deployment performance uncertain.

  • The robustness of the gains needs clarification. Appendix F already tests a second dictionary and acknowledges that SAEM's ASR advantage reverses with Qwen-Scope. Please report paired uncertainty for this comparison. Table 9 also gives different original-dictionary results from Tables 3–4, including for the residual probe, which does not use an SAE. Please explain whether these tables use different evaluation settings. This matters for interpreting the size and consistency of the reported gains.

Questions for authors

Q1. Are agent tool calls scored before execution? Does ASR require a flag before the harmful action? How would scoring work without knowing the injection position? Please also specify which split is used to calibrate thresholds.

Q2. What explains the differences between Table 9 and Tables 3–4? Please specify any differences in protocol and report paired uncertainty for SAEM versus the residual probe with each dictionary.

C1. The style-shift test and random-feature ablations provide useful evidence for targeted FORM removal. If I understand the test correctly, it changes which responses are sampled, so both style and content can change. Comparing the same information in different formats, or removing the FORM subcategories separately, would help explain what drives the gains. I regard these as additional mechanism checks.

C2. Section 4.2 says categorization uses descriptions alone, but Appendix A.2 includes promoted tokens. Please state the actual inputs.

1 comment

We thank the reviewer for the detailed reading and for recognizing the method's simplicity and the value of the shift evaluations. Each question is answered below, with new results in the revised PDF.

W1 / Q1: Agent evaluation and prevention.

  • Timing. The monitor reads the activations after the tool call is written and before the tool runs. For each assistant turn it takes the turn's last token; when the turn calls a tool, that token is the one that closes the tool call (</tool_call> in Qwen's chat format), so a flag at that turn can block the call. The monitor never reads the tool's output directly; injected text reaches it only through the agent's later turns.
  • Deployment-faithful evaluation. With cross-task averaging from the first turn, SAEM keeps the lowest ASR on all three models (mean over the three held-out suites; full table with the SAE baselines in General Response, point 3):
ModelMethodASR @ OR 5%ASR @ OR 20%
Qwen3-8BResidual probe0.3660.295
Qwen3-8BSAEM0.3420.248
Gemma-3-12BResidual probe0.3500.295
Gemma-3-12BSAEM0.3390.257
Qwen3.5-27BResidual probe0.5620.481
Qwen3.5-27BSAEM0.5010.429
  • Threshold calibration. In the submission, each arm's threshold is set on the benign trajectories of the evaluation set (for chat, on its safe responses), which is what makes the over-refusal rate exactly matched across arms; the procedure is identical for every arm. This fixes only the operating point at which ASR is read, as when reporting a true positive rate at a fixed false positive rate. The probe's weights and configuration are chosen on training and validation data, and ASR is measured on attacked trajectories that the threshold never sees. AUROC, which needs no threshold, gives the same ordering of SAEM and the residual probe in every overall cell (Tables 3 and 5).

W2 / Q2: Robustness and Table 9. Table 9 was computed on an earlier chat corpus from our initial Qwen-Scope experiments. Rerun on the main corpus, its residual probe row matches Tables 3 and 4 exactly, and SAEM ranks first with both SAEs. With Qwen-Scope's SAE, it stays above the residual probe in AUROC (+0.005 [+0.001, +0.009]) and matches it in ASR (0.000 [−0.016, +0.016]); with ours, it is better on both (AUROC +0.014 [+0.010, +0.019], ASR −0.025 [−0.031, −0.014]). The revised Table 9 is in General Response, point 2, and we updated Section 5.2 and Appendix F accordingly.

C1: Style-shift test. To clarify, the test does not modify any response. It only reweights existing responses within each class by their style score, so every evaluated response is an original generation. As with any reweighting or filtering design, the reweighted sets may also differ in attributes correlated with style. Two observations attribute the effect to style. First, the style score z is built only from 26 surface statistics (length, list structure, markup, discourse markers and lexical statistics), with refusal-lexicon terms excluded. Second, SAEM's advantage over the residual probe tracks the strength of the style cue: it is within 0.01 when style is predictive (k = +12) and grows as the cue weakens and then inverts, reaching 0.034, 0.059 and 0.090 at k = −12 on the three models (Table 2). Since SAEM differs from the residual probe only in removing surface-form features, style is the most direct explanation for this pattern.

C2: Categorization inputs. As specified in Appendix A.2, category assignment receives each feature's description together with its promoted tokens. Neither input involves a monitored response or a monitoring label. The description comes from the feature's max-activating contexts in the SAE's training corpus, and the promoted tokens come from projecting its decoder column through the unembedding. We have updated the abstract, Section 1, Section 4.2 and the Figure 2 caption to state both inputs.

Reviewer_2DH7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper is about sparse autoencoders and how to use them correctly for creating safety monitors. The paper is actually a pipeline with a couple of prompts on top of the standard SAE output, plus correction of the activations. The monitor is a linear probe on top of those 'filtered features', tested on a bunch of examples.

Critical strengths

The idea of using SAEs in a 'right way' is promising, as well as building such kind of pipelines, mixing numerical features, LLM classifiers and adaptable monitors. Maybe it will make SAEs more widely applicable.

Critical weaknesses

The benchmark results are somewhat misleading. I understand that it is difficult to make comprehensive studies, but for example Figure 3 has a classical trick on y-axis (new method bar looks amazing, but it is the matter of the coordinates, in reality the difference is not so big).

What I did not understand, is how those monitors affect the quality of the results. I.e., robustness frontier is nice, but can such monitoring deteriorate the quality of the agent / LLM itself? Why such set of benchmarks is selected? Not clear, difficult to follow.

Also, introducing the monitors inside LLM can influence the inference time.

Questions for authors
  1. Maybe I did not get it right, but I don't get the effect of detection on the real quality.
  2. Please justify the selectin of this particular set of benchmarks
  3. What is the effect on the inference time (i.e., using VLLM/ SGLang) of such monitor?
1 comment

We thank the reviewer for finding the direction promising, and answer each question below.

W1: Fig. 3 y-axis. To make the magnitudes explicit, we now report paired tests for every comparison in Fig. 3 (DeLong for AUROC, paired bootstrap for ASR), with 95% CIs in a new appendix table (Appendix H). Per-attack values remain in Tables 3 and 4, and the significance results are summarized in General Response, point 1.

Q1: Does the monitor degrade the LLM or agent? No. SAEM is an internal-state monitor, a standard way to add safety to a deployed model (Zou et al., 2023a; Goldowsky-Dill et al., 2025; Cunningham et al., 2026), and it degrades the model no more than any other internal-state monitor. Generation changes only when the monitor halts a response, and halting a benign response is exactly the over-refusal rate, which we match across all arms (10% for chat, 20% for agents). Fig. 5 sweeps this rate from 2% to 30%.

Q2: Why these benchmarks? We use two types of benchmarks, chat and agent, and in both we adopt benchmarks common in the LLM safety literature. In the agent setting, ControlArena provides a diverse set of environments and tasks for testing agents on loss of control, and AgentDyn tests them on prompt injections delivered through tool results; evaluating across the two changes both the tasks and the attack surface. In the chat setting, AdvBench, StrongREJECT, HEx-PHI, JailbreakBench, XSTest and WildJailbreak provide a diverse set of harmful requests, jailbreaks and over-refusal scenarios, and each of the four jailbreak attacks is held out in turn, so every number measures transfer to an unseen attack.

Q3: Inference time. SAEM adds about 7% of a forward pass: one SAE encoder at one layer, about 1.1 GFLOPs per token on Qwen3-8B against roughly 16 GFLOPs for the model itself, and nothing inside the model's computation. Importantly, this computation can run in parallel with the model's forward pass, and prior work finds that such probes add only a small, constant overhead at inference (Zhang et al., 2025). Because the probe is linear, the subtraction of the shortcut features reduces to scalars computed once:

w⊤h~R=1T∑t=1T(w⊤ht−∑f∈Raf(ht) (w⊤df))w^\top \tilde h_R = \frac{1}{T}\sum_{t=1}^{T}\Big(w^\top h_t - \sum_{f\in R} a_f(h_t)\,(w^\top d_f)\Big)

So no decoder pass is needed at deployment. With a JumpReLU SAE, only the encoder rows of the shortcut features need computing, about 45% of the encoder. On agents the monitor reads one token per assistant turn, so the cost is negligible. A text-based judge, by comparison, needs a full forward pass of a separate model. Serving stacks already support this kind of monitor: Anthropic's production jailbreak defense screens all traffic with a linear probe on Claude's internal activations (Cunningham et al., 2026), Google DeepMind has deployed activation probes for misuse mitigation in user-facing Gemini (Kramár et al., 2026), and Shu et al. (2026) run a per-token probe inside the vLLM decode loop to halt unsafe outputs during generation. SAEM adds only the encoder pass on top of such a probe.

References

H. Cunningham, J. Wei, Z. Wang, et al. Constitutional Classifiers++: Efficient production-grade defenses against universal jailbreaks. arXiv:2601.04603, 2026.

J. Kramár, J. Engels, Z. Wang, B. Chughtai, R. Shah, N. Nanda, and A. Conmy. Building production-ready probes for Gemini. arXiv:2601.11516, 2026.

H. Shu, X. Li, and P. Xue. Stop early, spend less: Hidden-state probes as a practical recipe for streaming moderation of LLM outputs. arXiv:2606.10487, 2026.

J. Zhang, A. Estornell, D. D. Baek, B. Li, and X. Xu. Any-depth alignment: Unlocking innate safety alignment of LLMs to any-depth. arXiv:2510.18081, 2025.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading. Three concerns are shared across reviews, and we address each with new analyses in the revised PDF (changes in blue).

1. Performance: why SAE-based monitors fall short, and how SAEM compares with them (2DH7, cEhH, Xvqf).

We would like to bring forward the paper's first contribution. Prior work finds that SAE-based probes and monitors match or trail a probe on the raw activation (Kantamneni et al., 2025; Rahman, 2026), without explaining why. Our work explains why existing SAE approaches fail as safety monitors and shows how to fix it. Section 3 identifies two causes:

  • Relevance is contextual, so judging features by their descriptions misses ones the monitor needs. Whether a feature matters for safety depends on the response it fires in, not on the feature itself, so an interpreter reading feature descriptions cannot pick out the features a monitor needs. A monitor built only from the 92.5% of features the judge called irrelevant shows no significant difference from one built from the 7.5% it called harm-relevant (p = 0.18), and an ℓ1 probe left to choose for itself puts two thirds of its weight on features the judge discarded (Section 3.1).
  • SAE features throw away part of the signal. An SAE reconstructs each activation only approximately, and the part it leaves out still carries safety signals. A monitor that reads SAE features never sees that part: even given every feature, it trails the raw probe (AUROC 0.723 vs. 0.747, Section 3.2). The loss grows with a weaker SAE. With Qwen-Scope's SAE, which leaves more of the activation unexplained, Harm-labeled loses 0.111 AUROC, while the two methods that subtract from the activation move by at most 0.01 (Appendix F).

SAEM is designed around both causes. It asks each feature only what it represents, which its description can answer, and it subtracts in the residual stream, so the monitor keeps the reconstruction error.

Against the SAE-based methods from prior work, the gain is large and consistent. Moreover, the improvement of SAEM is larger for larger models (the 27B model has a larger gap than the 8B model); this trend holds true across all of our evaluations. Harm-labeled follows the selection setup evaluated by Kantamneni et al. (2025) and Fomin (2026), and Retention ablation follows the removal criterion of Fomin (2026). Out-of-distribution AUROC (Tables 3 and 5):

ModelSettingHarm-labeledRetention ablationSAEM
Qwen3-8BAgents, cross-benchmark0.5870.8040.840
Gemma-3-12BAgents, cross-benchmark0.5100.7750.796
Qwen3.5-27BAgents, cross-benchmark0.5380.7040.749
Qwen3-8BChat, held-out attacks0.7550.7450.761
Gemma-3-12BChat, held-out attacks0.7370.7580.769
Qwen3.5-27BChat, held-out attacks0.7680.7740.823

Under cross-benchmark shift, Harm-labeled falls to near chance (0.51 to 0.59), as Section 3 predicts, while SAEM stays between 0.75 and 0.84. At matched over-refusal, SAEM also has the lowest ASR of any SAE-based arm on every model, both in chat (mean over held-out attacks) and under cross-benchmark agent shift. On Qwen3.5-27B chat, for example, its ASR is 0.043, against 0.120 for Retention ablation and 0.140 for Harm-labeled; on Qwen3-8B agents it is 0.128, against 0.232 and 0.213.

Paired Tests: All methods are scored on the same examples, so we compare them with paired tests: DeLong's test for AUROC (DeLong et al., 1988), the standard test for comparing two AUROCs measured on the same data, and a paired bootstrap for ASR at matched over-refusal, refitting the threshold in every resample (5,000 resamples for chat, 2,000 for agents). Each cell is SAEM minus the baseline with its 95% CI; bold marks intervals that exclude zero. The revision adds both tables to Appendix H.

Chat, out of distribution, mean over the four held-out attacks; ASR at OR = 10%:

BaselineModel and SAEΔAUROCΔASR
Residual probeQwen3-8B, our SAE+0.014 [+0.010, +0.019]−0.025 [−0.031, −0.014]
Residual probeGemma-3-12B, Gemma Scope 2 SAE+0.009 [+0.005, +0.013]−0.006 [−0.016, +0.004]
Residual probeQwen3-8B, Qwen-Scope SAE+0.005 [+0.001, +0.009]0.000 [−0.016, +0.016]
Residual probeQwen3.5-27B, Qwen-Scope SAE+0.033 [+0.025, +0.041]−0.051 [−0.066, −0.036]
Retention ablationQwen3-8B, our SAE+0.017 [+0.014, +0.021]−0.029 [−0.037, −0.016]
Retention ablationGemma-3-12B, Gemma Scope 2 SAE+0.010 [+0.004, +0.017]−0.027 [−0.037, −0.013]
Retention ablationQwen3-8B, Qwen-Scope SAE+0.013 [+0.005, +0.020]−0.004 [−0.026, +0.015]
Retention ablationQwen3.5-27B, Qwen-Scope SAE+0.049 [+0.039, +0.059]−0.077 [−0.095, −0.059]
Harm-labeledQwen3-8B, our SAE+0.006 [−0.001, +0.014]−0.015 [−0.026, −0.005]
Harm-labeledGemma-3-12B, Gemma Scope 2 SAE+0.032 [+0.026, +0.039]−0.026 [−0.039, −0.013]
Harm-labeledQwen3-8B, Qwen-Scope SAE+0.108 [+0.092, +0.124]−0.111 [−0.130, −0.083]
Harm-labeledQwen3.5-27B, Qwen-Scope SAE+0.055 [+0.043, +0.067]−0.097 [−0.117, −0.076]

Agents (AgentDyn), ΔASR at OR = 20%; cross-task is the mean over the three held-out suites:

BaselineSettingQwen3-8BGemma-3-12BQwen3.5-27B
Residual probeCross-task−0.013 [−0.030, −0.005]−0.052 [−0.103, −0.001]−0.065 [−0.125, −0.005]
Residual probeCross-benchmark−0.048 [−0.070, −0.026]−0.018 [−0.031, −0.005]−0.080 [−0.108, −0.052]
Retention ablationCross-task−0.157 [−0.200, −0.086]−0.089 [−0.159, −0.020]−0.122 [−0.190, −0.050]
Retention ablationCross-benchmark−0.105 [−0.137, −0.073]−0.012 [−0.023, −0.001]−0.177 [−0.217, −0.137]
Harm-labeledCross-task−0.145 [−0.206, −0.068]−0.063 [−0.139, −0.002]−0.131 [−0.200, −0.060]
Harm-labeledCross-benchmark−0.085 [−0.114, −0.056]−0.094 [−0.124, −0.064]−0.132 [−0.167, −0.097]

In agents, SAEM is significantly better than every baseline in all 18 cells. In chat, it is significantly better in 11 of 12 AUROC comparisons and 9 of 12 ASR comparisons, and no baseline is significantly better than SAEM in any cell.

Against the raw probe, the gain is largest on the 27B model in every setting of Tables 1 to 5: SAEM raises AUROC on all four held-out attacks (+0.024 to +0.043), cuts chat ASR from 0.094 to 0.043, and cuts cross-benchmark agent ASR from 0.401 to 0.321.

Earlier detection. A key benefit of SAEM is that it catches harmful content sooner. On the same unsafe responses, SAEM has flagged more harmful responses than the residual probe after every prefix length on all three models:

Share of unsafe responses flagged within the first t tokens, SAEM minus residual probe, in points (higher is better), at OR = 10%, mean over the four held-out attacks. Each method uses one threshold, with flags accumulated over prefixes, so these are not the per-depth values of Fig. 4.

Tokens readQwen3-8BGemma-3-12BQwen3.5-27B
5+1.1+3.7+4.1
10+0.7+4.8+3.5
20+1.3+4.0+3.7
50+2.0+2.5+2.7
100+4.1+1.6+3.2
Full response+4.4+0.9+4.0

2. Table 9 (cEhH, Xvqf). Table 9 was computed on a different chat corpus from the results in the main body (Tables 3 and 4); we used this corpus when initially testing the idea with the Qwen-Scope SAE. To avoid further confusion, we have updated Table 9 to cover the same benchmarks used in the main paper. Its residual probe row now matches Tables 3 and 4 exactly, and SAEM ranks first in both AUROC and ASR with both SAEs.

Table 9 (revised). Qwen3-8B chat, out-of-distribution, mean over the four held-out attacks; ASR at OR = 10%. Unexplained variance: ours 16.8%, Qwen-Scope 21.7%.

ArmAUROC, oursAUROC, Qwen-ScopeASR, oursASR, Qwen-Scope
Residual probe0.7470.7470.3820.382
Retention ablation0.7450.7390.3860.385
Harm-labeled0.7550.6440.3720.492
SAEM0.7610.7520.3570.381

3. Agent deployment (cEhH). After each assistant turn, the monitor takes the residual stream at that turn's last token, averages these vectors over the turns so far, and scores the average. When a turn calls a tool, its last token comes right after the tool call is written and before the tool runs, so a flag at that turn can stop the tool call.

To better isolate the monitor's ability to catch harmful tool calls we reran cross-task with the average starting at the first turn, so no method uses the injection position. SAEM keeps the lowest ASR of any method at both over-refusal rates on all three models (mean over the three held-out suites):

ModelArmWindowASR @ OR 5%ASR @ OR 20%
Qwen3-8BResidual probeAfter injection0.2590.169
Qwen3-8BSAEMAfter injection0.2470.155
Qwen3-8BResidual probeAll turns0.3660.295
Qwen3-8BHarm-labeledAll turns0.3610.332
Qwen3-8BRetention ablationAll turns0.3590.332
Qwen3-8BAll SAE featuresAll turns0.3560.332
Qwen3-8BSAEMAll turns0.3420.248
Gemma-3-12BResidual probeAfter injection0.2380.187
Gemma-3-12BSAEMAfter injection0.2290.135
Gemma-3-12BResidual probeAll turns0.3500.295
Gemma-3-12BHarm-labeledAll turns0.3640.319
Gemma-3-12BRetention ablationAll turns0.3570.295
Gemma-3-12BAll SAE featuresAll turns0.3490.285
Gemma-3-12BSAEMAll turns0.3390.257
Qwen3.5-27BResidual probeAfter injection0.3340.249
Qwen3.5-27BSAEMAfter injection0.2900.184
Qwen3.5-27BResidual probeAll turns0.5620.481
Qwen3.5-27BHarm-labeledAll turns0.5710.505
Qwen3.5-27BRetention ablationAll turns0.5960.499
Qwen3.5-27BAll SAE featuresAll turns0.6200.513
Qwen3.5-27BSAEMAll turns0.5010.429

Without the injection position, SAEM's ASR at OR = 20% is below the residual probe's by 0.047 on Qwen3-8B, 0.038 on Gemma-3-12B and 0.052 on Qwen3.5-27B.

Reviewer-specific points (inference cost, benchmark choice, categorization inputs, and the style-shift test) are answered in the individual responses.

References

E. R. DeLong, D. M. DeLong, and D. L. Clarke-Pearson. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics, 44(3):837-845, 1988.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.