acceptodds
Under review as a conference paper at ICLR 2027

EvoSteer: Restoring Safety in Experience-Driven Agentic Evolution via Activation Steering

11 views

Abstract

Experience-driven evolution allows agents to improve by accumulating and reusing experience while keeping the backbone model fixed, but recent studies show that the resulting memory can also degrade agent safety. Existing mitigations mainly modify the evolution mechanism, alter the prompt, or introduce additional agent components. We instead ask whether this degradation can be corrected directly inside the model while leaving the evolved memory and evolution process unchanged. We find that retrieved experience shifts the internal representations of harmful requests toward execution along a safety-relevant direction, providing a representation-level signal for correcting the degradation. Based on this observation, we introduce , an activation-steering method that uses the memory-free forward pass of the same request as a per-request reference and selectively removes the execution-ward component of the memory-induced representation shift. requires no modification to the model weights, prompt, evolved memory, or evolution algorithm. Experiments show that our method effectively mitigates the safety degradation induced by evolved memory while largely preserving benign performance. Our results demonstrate that representation-level correction provides a viable defense for experience-driven agent evolution without redesigning the evolution process.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent, evolution, safety, activation steering

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 4.00 over 3 reviews, above 99% of the others.
Reviewer_K5ZLOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces EvoSteer, a method for steering the refusal behavior of an LLM-based agent (in particular, memory-based ones that retrieves memories and incorporates them into its context). Prior work has shown that such memory augmentation can reduce refusal rates compared with memory-free baselines. EvoSteer first estimates a refusal-related direction ulu_l at each layer using a fit set. At inference time, it requires two forward passes, one with memory and one without. It projects the difference between their representations at selected token positions onto ulu_l, then counteracts the execution-ward component during the memory-augmented forward pass. The results suggest that EvoSteer achieves a better balance between safety-related refusal and task performance. EvoSteer is evaluated on Mind2Web-SC and SafeAgentBench, covering two memory methods and four base LLMs across the experiments. It shows promising recovery of refusal behavior on Qwen3-14B and Qwen3-32B.

Critical strengths
  • The paper addresses under-refusal in memory-based agents, a pressing concern given their utility and increasing adoption.
  • The motivation for EvoSteer is clear, and the methodology is reasonable. The paper is well organized, and the technical details are clearly explained. The analysis in Figure 2 provides useful support for the method’s design.
Critical weaknesses
  • The ablation study in Section 4.4 is not convincing. EvoSteer has a relatively small effect on Qwen3.6-27B, and the AWM memory baseline does not reduce the aggregate refusal rate compared with the memory-free baseline. This makes the setting less informative for evaluating recovery from memory-induced under-refusal. Moreover, several differences in Figure 3 are only 2.5 percentage points, corresponding to a single task, and may reflect evaluation variability. These results provide limited support for the conclusions drawn from the ablations.

  • Important baselines are missing:

  1. As discussed around line 122, a fixed steering vector can be applied to adjust the model’s refusal boundary. In this setting, a natural baseline would directly steer along the same ulu_l estimated from the fit set, with a fixed steering strength tuned on that set. This comparison would help establish the benefit of EvoSteer’s adaptive correction over simpler steering, particularly given the additional reference forward pass it requires.
  2. Another possible baseline is a lightweight probing classifier trained on representations from the fit set to predict whether a request should be refused. Would such an approach be applicable to this setting, and how would it compare with EvoSteer?
Questions for authors

See weaknesses above.

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

This paper investigates safety degradation in experience-driven LLM agents, where memories accumulated from benign tasks can weaken refusal behavior. It proposes EvoSteer, an activation-steering method that estimates layer-wise safety directions from paired benign and rule-violating requests. During inference, EvoSteer uses the memory-free representations of the same request as references and selectively corrects memory-induced shifts toward execution. The method requires no changes to model weights, prompts, memories, or evolution procedures. Experiments across multiple models, benchmarks, and memory evolution mechanisms demonstrate improved safety while largely preserving benign-task performance.

Critical strengths

The paper addresses safety degradation caused by benign experience accumulation, an important challenge for self-evolving agents. Its representation-level approach restores safety without modifying model weights, memories, or evolution procedures, providing a meaningful alternative to existing memory-level defenses.
Experiments across multiple models, benchmarks, and memory evolution mechanisms demonstrate improved refusal behavior while largely preserving benign-task performance. The random and orthogonal direction controls further support the effectiveness of the proposed safety-direction correction.

Critical weaknesses

The safety direction is estimated from benign and violating requests, but their representation differences may reflect task semantics rather than refusal behavior specifically. The current analysis does not fully establish that the observed shifts are the causal mechanism behind memory-induced safety degradation.
EvoSteer treats negative shifts along the estimated safety direction as undesirable and corrects them regardless of whether the request is harmful. However, such shifts may also reflect legitimate execution-oriented benefits from memory. The paper does not sufficiently establish that this criterion reliably distinguishes safety degradation from beneficial memory effects.
EvoSteer repeatedly adjusts hidden states across multiple layers using a fixed memory-free reference, but later deviations are also affected by preceding interventions. The paper does not sufficiently analyze these interactions or establish why repeated layer-wise corrections are preferable to simpler interventions at selected layers.

Questions for authors

How can the authors distinguish the proposed refusal-related representation shift from general semantic differences or context-induced activation changes?
What evidence supports using negative projection shifts as the correction criterion, particularly when similar shifts may occur on benign requests?
How do corrections at earlier layers influence the deviations measured at later layers? Is repeated layer-wise correction necessary compared with interventions at selected layers?

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

This paper proposes EvoSteer, an activation steering method to mitigate safety degradation caused by evolved memory in LLM agents. It estimates layer wise safety directions from violating benign request pairs and uses memory free activations as references to correct execution ward shifts induced by retrieved memory. Experiments across four models, two evolution methods, and two benchmarks show that EvoSteer improves refusal rates while largely preserving benign task performance, without modifying model weights or memory.

Critical strengths

The paper addresses an important safety issue in experience driven LLM agents. EvoSteer offers a simple and training free approach that requires no modifications to model weights or evolved memory. The experiments cover multiple models and evolution methods, with results showing improved refusal rates while largely preserving benign task performance. The ablation studies also provide useful evidence supporting the effectiveness of the proposed steering direction and correction strategy.

Critical weaknesses

Weaknesses

W1. Insufficient justification for asymmetric correction.
If memory should not alter safety decisions, its induced shift should ideally have minimal projection onto the safety direction. However, EvoSteer only corrects negative shifts and allows overcorrection when β>1\beta>1. The rationale for this asymmetric design remains insufficiently justified.

W2. Potential over refusal on benign tasks.
EvoSteer may suppress beneficial execution ward shifts, leading to unnecessary refusals. Table 2 shows increased false refusal rates in several settings, raising concerns about whether the method restores safety or simply amplifies refusal behavior.

W3. Inability to explicitly correct memory induced false refusals.
Memory can also increase false refusals, as shown by Qwen3-32B with MemP in Table 2. However, EvoSteer leaves positive shifts uncorrected, providing no explicit mechanism to address such errors.

W4. Unverified consistency of the safety direction.
EvoSteer estimates a single direction per layer by averaging activation differences across request pairs. If individual directions vary substantially, the mean may not reliably represent the safety direction for each request. The paper lacks an analysis of this variability.

W5. Additional inference overhead.
EvoSteer requires an extra memory free prefill for every request, followed by layer wise activation corrections during the memory augmented prefill. This introduces additional computational and latency overhead, but the paper does not report inference time or memory cost measurements.

Questions for authors

I have raised all my main concerns in the weaknesses section, please refer to that part.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.