acceptodds
Withdrawn from ICLR 2027

SafeLoop: Preserving Refusal of Harmful Prompts in Looped Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Refusing to fulfill harmful requests is a crucial safety behavior of language models. Looped language models offer a parameter-efficient approach to reasoning by reusing the same Transformer weights for additional computation before predicting each token. However, whether refusal remains reliable as recurrent depth increases is not well understood. Evaluating this reliability through attack success alone can be misleading: deeper models can stop producing harmful answers because their generation has degenerated. We identify a distinct failure in refusal representations. A behaviorally validated refusal direction can recover harmful–benign separability after erasure in one loop, yet its projection can drift far from its reference value during unperturbed inference. The ability to reconstruct this representation therefore does not ensure that additional loops preserve it. This finding motivates SafeLoop, which stores each input's projection at a reference loop and corrects subsequent deviations without updating model parameters. On frozen Ouro-1.4B, SafeLoop increases the percentage of responses that are non-degenerate refusals (effective refusal) from 9.5% to 41.0% at 16 loops and from 6.5% to 35.5% at 24 loops. It also improves effective refusal in selected recurrent models converted from dense backbones. Benefits vary across models and depths; projection correction does not resolve widespread generation degeneration and can increase benign over-refusal. These results establish cross-loop preservation as a concrete target for refusal control and show why evaluating deeper inference requires separating effective refusal from generation failure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Looped Language Models, LLM Safety, Refusal Representations

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nFAAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether safety refusal behavior is preserved when looped language models perform substantially more recurrent computation than their training depth. The paper makes an important distinction between effective refusal and generation degeneration, arguing that low attack success at large recurrent depths can be misleading when the model simply stops producing usable outputs. Through experiments on Ouro and several dense-to-loop recurrent models, the authors show that a behaviorally validated refusal direction can be rapidly reconstructed after transient erasure, while its projection nevertheless drifts substantially during ordinary inference. Based on this observation, the paper proposes SafeLoop, an inference-time intervention that caches each input's refusal-direction projection at a reference loop and repeatedly corrects later hidden states toward that reference. The method substantially improves effective refusal in some settings, e.g., from 9.5% to 41.0% at 16 loops and from 6.5% to 35.5% at 24 loops for Ouro-1.4B.

Critical strengths
  1. The paper identifies an new, open question in AI safety. The author finds looped LLMs may behave differently from traditional LLMs in refusals. The experiments demonstrate that very low attack success at large recurrent depths can coexist with severe generation degeneration. This makes a compelling case that conventional jailbreak success rates alone can substantially overestimate safety at extrapolated recurrent LLMs.

  2. The paper's writing is clear and easy to understand. The author uses figures, examples, and comprehensive expreiments to structure their paper.

Critical weaknesses
  1. The effectiveness of SafeLoop is not convincing enough, as 1) it's substantially model dependent, and 2) its impact on helpfulness isn't negliable.

The method works particularly well for Ouro-1.4B, but gives little or negative benefit for Ouro-2.6B; for example, effective refusal decreases by 2.0 and 6.2 percentage points at depths 16 and 24. Likewise, SafeLoop only marginally improves effective refusal for highly degenerate Gemma at extrapolated depth. Since looped LLM is a relatively new direction and there're not many open-source models to test on, the generalizability of the method is questionable.

In addition, a good safeguard approach should not only prevent harmful generations but also preserve the model's ability to respond appropriately to benign requests. The reported increase in over-refusal on XSTest-safe and the changes on MMLU already suggest that SafeLoop may trade off safety for helpfulness. And since the reported tasks are rather simple, SafeLoop's impact on more complex tasks are still unknown.

  1. The evidence that refusal-projection preservation is the causal mechanism behind the behavioral gains is not yet sufficient.

SafeLoop directly controls only one coordinate of the hidden state, and the paper itself reports that a random-axis control can also reduce deep Ouro degeneration. This makes it difficult to attribute the behavioral improvements specifically to preservation of the refusal direction rather than to a more general stabilization effect caused by repeatedly constraining the hidden-state dynamics. The authors acknowledge this issue, but the current experiments do not fully disentangle these explanations. A stronger set of controls could include multiple behaviorally validated directions, matched random or semantic directions, or interventions that preserve the same amount of hidden-state displacement without explicitly preserving the refusal coordinate.

Questions for authors
  1. How sensitive is SafeLoop to the choice of reference loop and correction strength (beta)?

  2. What is the computational and memory overhead of token-wise persistent correction during generation? For fast inference engines like vLLM and sglang, can SafeLoop adapt to them easily?

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

This paper investigates the reliability of refusal behaviour in looped language models as recurrent inference depth increases. The authors observe that low attack success rates at deeper inference depths can be misleading because models may stop producing coherent outputs rather than reliably refusing harmful requests. To address this issue, they introduce effective refusal, defined as detected refusal without detected generation degeneration. The paper further investigates the behavior of a refusal-related direction in the residual stream across recurrent loops. The authors show that although harmful–benign separability can rapidly recover after a transient erasure of this direction, the corresponding projection can drift substantially during unperturbed inference. Motivated by this observation, they propose SafeLoop, a training-free intervention that records an input-specific refusal projection at a reference loop and repeatedly corrects deviations from this reference during subsequent loops.

Critical strengths
  1. The evaluation point about depth extrapolation is useful and well supported. Figure 1 and Tables 3–4 show that attack success near zero at deep loops coincides with 90%+ degeneration.
  2. SafeLoop provides a mathematically straightforward approach to maintaining an input-specific refusal-related projection across recurrent computation without modifying model parameters. The distinction between rapid reconstruction after transient erasure and natural projection drift is interesting.
Critical weaknesses
  1. Although SafeLoop produces substantial improvements on Ouro-1.4B, the benefits are not consistent across the evaluated architectures. Ouro-2.6B does not show comparable improvement, and correction can reduce effective refusal at certain depths.
  2. The experiments do not fully establish that the observed improvements arise specifically from preserving the refusal direction. The authors acknowledge that correction along a random axis can also substantially reduce degeneration in deep Ouro models, suggesting that part of the benefit may result from nonspecific changes to recurrent dynamics. Furthermore, effective refusal relies on a refusal-phrase matcher and heuristic degeneration detector rather than comprehensive human evaluation. The method also produces increased benign over-refusal in some settings, such as the 24-percentage-point increase on XSTest-safe at 16 loops for Ouro-1.4B.
Questions for authors
  1. With 10 or more random axes (and norm-matched non-refusal probe directions) at matched β=1, how does the distribution of ΔQ and ΔD compare with the refusal direction at T=16, 24, and 32?
  2. Is there a setting where SafeLoop reduces harmful completions (not degeneration) at a depth someone would actually use?
  3. The T=4 reference row in Table 1 (Q=71.5, J=19.5) appears closer to the α=0.5 sweep row in Table 3 than to the α=0 row (Q=71.0). Which configuration does the main SafeLoop protocol use, and is α held fixed across the off/on comparisons and the suppression experiments in Tables 6–7?
Reviewer_L4YSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors analyze refusal within looped LLMs. They find that more loops does not equal better refusal and show that this effect is largely attributed to degeneration of the output. They further show that erasure of a refusal direction is recovered in the subsequent turns and that the model will have roughly the same refusal representation compared to the same loop without intervention. This shows that the single-time intervention likely does not change anything and that the natural drift of the refusal representation occurs nevertheless. That leads to the conclusion that one needs to do an intervention at every step which is why the paper proposes SafeLoop to clamp the refusal projection at every loop to a reference value of an early iteration.

Critical strengths

The paper shows two interesting findings:

  1. that looped models do not become safer but effective refusal degrades typically and the effect was previously misattributed to safety rather than degeneration.
  2. looped models go back to the refusal representation after erasure in the previous iteration but this does not help sustain effective refusal.
Critical weaknesses

Potentially missing substance. To me the paper consists of two interesting findings and is reframed to provide a method. The method itself is not that strong since on e.g. Ouro 2.6B (or Qwen at T=4) it does not really add any benefit and even slightly hurts performance. And in the appendix Table 11 & 12 the authors barely improve over clamping a random orthogonal direction. Also recovering effective refusal seems somewhat meaningless for queries when the model wouldn't provide a non-degenerative answer without intervention so looking at the ones that only would have a meaningful harmful answer without refusal would be required to strengthen the method's usefulness.

This is just my initial assessment, please view any statement as a question you can rebut.

Questions for authors

Why does the one-shot control use the t-1 loop as a reference and not also the same as persistent? Did you try that?

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

Related papers

Discussion (0)

Sign in to comment.