Chat models that refuse a dangerous request when it is asked directly often comply when the same request is spread across several turns. Interpretability work offers a clean signal for this. One can measure how strongly the model's hidden state points along a "refusal direction", a single vector in activation space that lights up when the model is about to decline. During multi-turn jailbreaks that signal steadily fades. The obvious defense is to watch the signal and intervene when it drops. The authors who found the fade say this does not work in practice. Ordinary technical, medical, or scientific conversations also show low refusal activation, so a single global threshold either misses attacks or flags legitimate expert users (State-Dependent Safety Failures in Multi-Turn Language Model Interaction).
We think this is a domain-generalization problem rather than a dead end. What counts as a "normal" refusal level depends on the conversational domain, and also on how far into the conversation one is. A monitor tuned on some domains then misfires on others, the same way dialogue-state confidence estimators lose calibration when the schema changes (Confidence Estimation for LLM-Based Dialogue State Tracking). We propose to measure refusal drift relative to what benign conversations in the same domain, at the same turn depth, look like. The reference statistics are estimated from unlabeled benign traffic in that domain. The claim is easy to explain to a colleague. Safety monitors fail across domains because "low refusal" is domain-relative. If you re-center the signal on each domain's own benign baseline, a monitor trained on a few domains transfers to unseen ones, including conversations that move from one domain into another mid-dialogue.
Core Claim
Let $h_t$ be a mid-layer hidden state of the target model at the last token of user turn $t$. Let $r$ be a refusal direction extracted from contrastive single-turn harmful and harmless prompts. The raw monitor reads $s_t = r^\top h_t$. Suppose we have a small pool of unlabeled benign conversations from domain $d$. From it we estimate a turn-indexed mean $\mu_d(t)$ and a covariance $\Sigma_d(t)$ of hidden states. The monitor then reads the whitened, re-centered state:
$$\tilde h_t = \Sigma_{\hat d_t}(t)^{-1/2}\,\big(h_t - \mu_{\hat d_t}(t)\big),$$
Here $\hat d_t$ is a soft domain assignment computed from the conversation so far. It is a nearest-centroid mixture over domain reference pools, so a conversation that drifts from, say, hotel booking into pharmacology is scored against a reference that also shifts. A linear probe, or simply the projection $r^\top \tilde h_t$, is trained on labeled source domains and applied unchanged to held-out domains. The only new-domain input is benign, unlabeled conversations.
The prediction is specific. Under leave-domain-out evaluation, an unnormalized probe loses much of its separation between attacks and benign expert conversations. The domain-relative probe keeps most of it. The gap should be largest in the technical domains that STAR singles out as problematic.
Relation to Prior Work
Several papers motivate the gap. STAR projects hidden states onto a refusal direction, documents its monotonic decay across turns in Llama-3, and explicitly rejects refusal monitoring because benign technical dialogues also read low (State-Dependent Safety Failures). We take that objection as the problem to solve, not as a closing verdict. SafeDialBench evaluates models across 22 application scenarios and finds guardrails collapsing after a few turns. It reports scenario scores as static bins and never tests transfer between them (SafeDialBench). We reuse its scenarios as domains for a leave-one-out split. Two attack papers show that professional framing is the strongest attack lever. STAR does this with query-aware roles, and BLUEPRINT with consultation-style worldview simulation reaching jailbreaks in 2.46 queries on average (Before the Script, Set the Stage). That framing is exactly what pushes the raw refusal signal into the range where benign experts also sit.
Other work supplies the method. The idea that multi-turn history should be compared against a domain-specific notion of "in-distribution" comes from out-of-domain intent detection. Caro separates domain-discriminative signal from multi-turn conversational noise and mines pseudo-out-of-domain examples from unlabeled corpora (Out-of-Domain Intent Detection Considering Multi-Turn Dialogue Contexts). We borrow the use of unlabeled in-domain data but aim it at safety drift rather than functional intents. Sun et al. show that calibration learned on one dialogue schema does not carry to another (Confidence Estimation). That result is the domain-generalization analogue of our hypothesis.
Some work constrains the design. CoSafe shows that attackers hide the harmful referent behind pronouns whose antecedents sit in earlier turns (CoSafe). Any monitor that compresses away history, as Caro's information bottleneck does, risks discarding exactly that antecedent. For this reason we read the full-context hidden state rather than a compressed summary. MTSA and CoSafe both report that hardening increases over-refusal; MTSA, for example, reports a 5.62% rise on XSTest (MTSA). So false positives on benign expert traffic are a primary metric, not an afterthought.
Research Program
The first phase is diagnostic and stands on its own even if the method fails. We ask two questions. Does the raw refusal projection's benign distribution shift by domain and by turn depth? Do conversations that transition across domains erode faster than conversations that stay in one domain?
We build domain-matched benign multi-turn conversations for each SafeDialBench scenario. They use the same register and terminology as the attack dialogues and are checked by annotators to contain no harmful goal. We add MultiWOZ 2.2 and Schema-Guided Dialogue sessions as task-oriented benign references. We then construct transition conversations. Each starts as a genuine task-oriented session in one domain and pivots, through a decomposed request in the style of Speak Out of Turn, into a harmful goal from MultiBreak's 26 categories (MultiBreak). Matched controls pursue the same goal within a single domain.
The second phase builds the domain-relative monitor on open-weight models. We use Llama-3-8B-Instruct, where STAR located the refusal bottleneck, and one other instruct model. We try two monitor variants: the projection-only form, and a full linear probe on whitened states. We also try two domain-assignment variants: oracle labels and inferred soft labels.
The third phase tests whether flagging is useful. On a flag, the system prepends an explicit safety reminder in the manner of STREAM's warnings, without retraining the model. We then measure attack success and benign utility end to end.
Baselines
The naive reference point is the undefended target model on each benchmark. That number anchors how much any defense helps. SafeDialBench reports, for example, a 69.60% attack success rate for its most vulnerable model (SafeDialBench). We re-measure our own target models rather than borrow that figure.
The first internal baseline is the global-threshold refusal monitor that STAR describes and dismisses. It uses the same refusal direction and the same layer, with one threshold tuned on pooled source domains. Our method must beat it at equal recall on false positives in held-out technical domains. Otherwise the core claim fails.
The strongest internal competitor is a supervised linear probe trained on unnormalized hidden states from the same labeled source-domain data. It gets the same layer choices and the same hyperparameter search budget. It can also absorb the benign target-domain pool as extra negatives, which gives it access to the same unlabeled data our method uses. Beating this probe on leave-domain-out area under the ROC curve, and on false-positive rate at 90% attack recall, is the decisive comparison.
The strongest existing external method is STREAM, a distilled 1.5B reasoning moderator that reads the dialogue history and injects warnings. It reports an average attack success rate of 57.3% on LLaMA-3.1, against 75.3% for Llama-Guard-3 and 72.0% for Granite-Guardian-3.2 (SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues). We run the released moderator, or a retrained one if weights are unavailable, on the same held-out domains with the same warning-injection intervention. Our method counts as beating STREAM if it achieves lower attack success at equal or lower over-refusal on our domain-matched benign set and on XSTest. A tie at a fraction of the inference cost would be reported as a tie, not a win.
We also include periodic grounding reminders every three turns. This is the cheap intervention Liu et al. found reduced belief confirmation by 0.55 points (A Prompt-Masked Pilot for History-Dependent Safety Degradation).
Ablations isolate the key ingredient. One replaces domain-specific statistics with global benign statistics, so normalization is kept but domain is removed. Another drops the turn index from $\mu_d(t)$. A third swaps oracle domain labels for inferred ones. A fourth varies the size of the unlabeled benign pool from tens to thousands of conversations. If global normalization performs as well as domain-specific normalization, the domain claim is false and we will say so.
Why not just…?
Why not just train a better supervised probe on more domains? This is the most serious threat. If labeled harmful and benign data from many domains is available, a pooled probe may learn domain invariance on its own. Our position rests on the realistic constraint that new deployment domains arrive with benign traffic but without labeled attacks. The leave-domain-out comparison against the probe given the same unlabeled pool settles it directly. We will also report a scaling curve over the number of source domains. If the gap closes once five or six source domains are available, the contribution shrinks to a low-data regime, and we will state that.
Why not just use an external multi-turn moderator like STREAM? STREAM already beats per-turn guards. However, it reads surface text, and its own authors leave open whether cross-domain vocabulary shifts fool it. Human red-teaming also broke defenses that looked perfect against automated attacks. CYGNET went from 0% to 70.4% attack success (LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet). We compare against STREAM on held-out domains and on the human-written MHJ conversations. We also test a hybrid that feeds our flag to STREAM, in case the two signals are complementary rather than competing.
Why not just harden the system prompt or add periodic reminders? CogSafe reports safety swings of 3% to 35% from system-prompt changes (Dynamic Evaluation with Cognitive Reasoning for Multi-turn Safety). Liu et al. found that one-time framing still collapses around turns 5 to 7. Periodic reminders did help, though, so they are run as a baseline at matched reminder frequency. Our monitor must show it achieves the same safety with fewer interventions, or with less utility loss on benign conversations.
Why not just retrain the model with multi-turn reinforcement learning? MTSA's future-reward training cut Zephyr-7B attack success from 74.50% to 23.50% but raised over-refusal (MTSA). It also requires retraining for each model. Our proposal is a training-free monitor for deployed models, so it is complementary. We include MTSA-style training only as a reference, if compute allows, and evaluate its over-refusal on our task-oriented benign set, where backward credit assignment may penalize legitimate multi-turn slot filling.
Proposed Experiments and Decisive Checks
All monitors are evaluated by leave-one-scenario-out over SafeDialBench's 22 scenarios. They are further tested out of distribution on CoSafe, MultiBreak, the MHJ human jailbreaks, and our constructed transition conversations. Benign false positives are measured on the domain-matched benign sets, MultiWOZ 2.2, SGD, MT-Bench, and XSTest. Detection metrics are area under the ROC curve, false-positive rate at fixed attack recall, and the turn at which an attack is first flagged relative to the first harmful output. End-to-end metrics are attack success rate under the warning intervention and over-refusal. Harmfulness is judged by an LLM judge spot-checked by humans, following SafeDialBench's reported above-80% judge–human agreement.
Three results would decide the proposal. First, the diagnostic must show that benign refusal projections differ substantially by domain. If they do not, STAR's objection has a different cause and we stop. Second, the domain-specific normalization must beat both the global-threshold monitor and the pooled supervised probe on held-out domains, with the gap concentrated in technical scenarios. Third, on transition conversations, the shifting soft-domain reference must catch pivots that a fixed single-domain reference misses.
We will also run an adaptive attacker that knows the monitor and tries to keep hidden states near the target domain's benign mean while pursuing a harmful goal. Its success rate tells us how much of any gain survives an informed adversary.
References
- State-Dependent Safety Failures in Multi-Turn Language Model Interaction
- Confidence Estimation For LLM-Based Dialogue State Tracking
- SafeDialBench: A Fine-Grained Safety Evaluation Benchmark for Large Language Models in Multi-Turn Dialogues with Diverse Jailbreak Attacks
- Before the Script, Set the Stage: How Worldview Simulation Amplifies Psychologically Grounded Persuasion in Multi-Turn Jailbreaking
- Out-of-Domain Intent Detection Considering Multi-Turn Dialogue Contexts
- CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference
- MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming
- Speak Out of Turn: Safety Vulnerability of Large Language Models in Multi-turn Dialogue
- MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety
- SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues
- A Prompt-Masked Pilot for History-Dependent Safety Degradation in Multi-Turn Conversational Agents
- Uncovering Model Vulnerabilities With Multi-Turn Red Teaming
- Dynamic Evaluation with Cognitive Reasoning for Multi-turn Safety of Large Language Models