acceptodds
Withdrawn from ICLR 2027

Same Request, Different Answer: How Conversation History Shapes the Decision to Refuse or Comply

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The same harmful request may be refused when asked directly, yet answered after several rounds of conversation. We combine research on text-based multiturn jailbreaks with mechanistic interpretability, holding the final request fixed on Llama-3.1-8B-Instruct and Crescendo attack histories and using history controls and internal-state interventions to trace how prior exchanges change refusal or compliance. We find that the actual exchanges between both participants and the assistant’s most recent response affect attack outcomes, while retaining only length, topic, or one participant’s content has difficulty reproducing the effect of the complete history. This effect can partly transfer with the response-opening state, namely the residual states at the template positions before the answer and the first few generated positions: transferring single-turn states can restore refusal, while transferring states from the attack history can induce compliance or recover part of the effect missing from failed histories. The effective components change with depth; the refusal direction can strongly change answers, but transferring it alone has difficulty restoring the effect of the complete history. Attention-source interventions further identify a path that reads explicit refusal wording in the preceding assistant reply, while timing comparisons show that the main effective stage for restoring refusal is at the opening of the answer, with later interventions changing answer content more often. These results suggest that the model’s safety behavior is jointly regulated by multiple components that are sensitive to history and change across layers, and establish a mechanistic analysis workflow connecting history text, the response-opening state, and final behavior.

Primary area:
interpretability and explainable AI
Keywords:
multi-turn jailbreak, Refusal Direction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_z6NDOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates why the exact same final request can lead to refusal in a single-turn setting but compliance after a multi-turn conversation. Using Llama-3.1-8B-Instruct with fixed final requests and Crescendo attack histories, the authors combine history controls, residual-state interchange, activation steering, cross-layer pinning, and attention-head interventions. Their main contribution is a causal framework linking conversation history, response-opening internal state, refusal or compliance. They show that the full interaction between both user and assistant, especially the assistant’s most recent reply, materially affects behavior; much of this effect is carried by the model’s internal state at the beginning of the answer. They also find that the effective components change across layers, and that a strong refusal direction can substantially alter behavior but cannot by itself recover the full effect of conversation history.

The paper validates these claims through multiple controlled and causal experiments. Transferring single-turn response-opening states into multi-turn attack contexts substantially restores refusal, while transferring multi-turn states into single-turn contexts increases compliance. Attention interventions further identify heads that read explicit refusal wording from the previous assistant response, and timing experiments show that interventions at the start of generation are much more effective at restoring genuine refusal than later interventions, which mainly alter answer content. The authors also evaluate capability and over-refusal costs, and report preliminary similar phenomena on Gemma3-4B and Qwen3-8B, though the full quantitative analysis is limited to Llama-3.1-8B-Instruct.

Critical strengths

Strong causal mechanistic design. The paper fixes the final request and varies only the conversation history, then combines history controls, state interchange, steering, cross-layer pinning, and attention-head interventions. This allows the authors to go beyond correlation and identify internal states that causally influence refusal versus compliance, making the mechanistic contribution substantially stronger than simply demonstrating a multi-turn jailbreak effect.

Multiple complementary experiments support the main claims. Reversal, induction, rescue, attention-source, and intervention-timing experiments consistently point to the response-opening state as a key decision stage, while also showing that a single refusal direction cannot explain the full history effect. This convergence of evidence makes the central conclusions more convincing and consequential for understanding multi-turn safety behavior.

Critical weaknesses

Limited external validity. The full quantitative mechanistic analysis is conducted almost entirely on Llama-3.1-8B-Instruct with Crescendo. The results on Gemma3-4B and Qwen3-8B are only preliminary and lack systematic quantitative validation. It is therefore unclear whether the identified mechanisms generalize across model families, scales, or other multi-turn attack settings.

The main mechanistic results rely heavily on a selected subset. Many of the central state-interchange, rescue, and cross-layer analyses are conducted on 64 trajectories selected because their behavior changed with conversation history. Although the paper provides larger-scale replication for some directions and attention interventions, the key state-transfer results are not validated on an unselected evaluation set. This makes it difficult to determine how sensitive the reported effect sizes and mechanistic pathways are to the selection procedure.

Questions for authors

The 64 trajectories used for the main mechanistic analyses were selected because their behavior changed with conversation history. Do the core response-opening state interchange results remain qualitatively similar on the full 219 evaluation trajectories, or on another unselected subset? This would materially affect my assessment of how selection-sensitive the central mechanistic claims are.

Reviewer_4V7MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how conversation history changes an LLM's decision to refuse or comply with harmful requests. Using Crescendo multi-turn jailbreaks on Llama 3.1 8B, the authors hold the final request fixed and examine the effects of different conversation histories. They use activation interchange, representation steering, and attention interventions to identify internal states associated with refusal and compliance. The main findings suggest that conversation history influences the response-opening state, that the refusal direction alone cannot explain the full history effect, and that early-generation interventions are particularly effective at restoring refusal.

Critical strengths
  • The paper addresses an interesting question about how conversational context influences model safety behavior. Holding the final request fixed and manipulating the preceding history provides a useful experimental setup.

  • The authors conduct several causal interventions rather than relying solely on observational analysis. The distinction between changing refusal behavior and recovering the effects of conversation history is interesting.

Critical weaknesses
  1. Limited generalizability of the findings.

The complete quantitative mechanistic analysis is conducted on a single model, Llama 3.1 8B using one jailbreak strategy, Crescendo. Given the diversity of contemporary models, architectures, and safety-training approaches, it is unclear whether the identified representations and mechanisms are specific to this model or generalizable properties of LLM safety behavior.

Although the authors report preliminary observations on Gemma and Qwen, systematic cross-model validation is missing. Furthermore, the detailed mechanistic analysis uses only 64 selected trajectories in which behavior changed with conversation history. This selection is useful for studying the phenomenon but further limits the generality of the conclusions.

  1. The mechanistic insights remain incomplete.

The experiments demonstrate that transferring internal activation states can substantially alter refusal behavior. However, this does not by itself explain how conversation history changes the model's safety decision.

The authors show that the refusal direction is an effective intervention coordinate but is insufficient to recover the full history effect. While this is an interesting observation, the additional components and computations responsible for the effect remain largely unidentified.

Consequently, the paper provides an extensive collection of intervention results, but the resulting mechanistic understanding remains somewhat limited. It is unclear what generalizable insight emerges beyond the observation that conversation history alters internal states and that these states causally influence subsequent behavior.

Questions for authors
  • Do the main mechanistic findings generalize to newer and larger models, particularly models with different safety-training procedures? How consistent are the identified directions and intervention effects across model families?

  • Do similar mechanisms arise under other multi-turn jailbreak strategies or naturally occurring conversations, rather than only selected successful Crescendo attacks?

  • Beyond demonstrating that activation interchange can change refusal behavior, what specific computational mechanism have the authors identified that explains how conversational history changes the decision to refuse or comply?

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

The paper combines multi-turn jailbreak analysis with mechanistic interpretability to investigate how conversation history changes an LLM's decision to refuse or comply with the same harmful request. Particularly, the authors analyze the internal representations of Llama-3.1-8B-Instruct under Crescendo-generated attack histories.

Critical strengths
  1. The paper introduces a detailed experimental framework for investigating the internal mechanisms responsible for refusal and compliance under multi-turn attacks.

  2. The authors investigate not only attack success rates but also different response outcomes.

Critical weaknesses
  1. Main concern is the limited generability due to evaluation of only one local LLM (Llama-3.1-8B-Instruct) and one multi-turn attack approach (Crescendo).

  2. The formal methodology and experimental results are difficult to follow. The paper introduces numerous internal representations, intervention strategies, mathematical formulations, and evaluation metrics without sufficiently clear explanations of their relationships.

  3. The paper understanding requires a deep understanding of domain. For example, the authors select layer 11 for the main residual-interchange experiments (Table 1), but the motivation for selecting this particular layer is not sufficiently explained. Moreover, the ASR presented in Figure 7 shows the reduction from layer 7 to 11 but after layer 11 the ASR was increased. This phenome isn't explained.

  4. Insufficient cross-model validation using only preliminary experiments (Appendix C) with Gemma3-4B and Qwen3-8B. Only qualitative observations are provided without detailed numerical results or direct comparisons with the main experiments.

  5. It feels like the authors did an extensive use of genAI tools during the research as acknowledged by authors (see AI Use Statement) " generative AI was employed for research ideation, hypothesis refinement, experiment design and implementation, interpretation of results, mathematical proofs, literature review, and manuscript preparation" In addition, at least one reference may be wrong and was not verified: "Measuring massive multitask language understanding"

Questions for authors
  1. Given the extensive use of generative AI for research development and mathematical reasoning, what independent verification procedures were performed.
  2. More clear explanations of methodology, especially selection of layers, and generability is required.
  3. Could the authors provide reproducible implementation artifacts, experimental configurations, and validation scripts.

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

Related papers

Discussion (0)

Sign in to comment.