Chat models that refuse a harmful request asked outright often comply when the same request is spread over several turns. The attacker wraps it in a story, splits it into harmless-looking sub-questions, and then asks the model to expand the one piece they actually wanted. Safety monitors fail here for a specific reason. When they read a long transcript, the harmful payload is buried among framing, role-play, and decoy questions. When they read one turn at a time, every turn looks benign. This proposal borrows an old idea from social-media summarization: what readers reply to tells you which parts of a text matter. We apply it to the dialogue itself. In a multi-turn attack, the user's later turns are replies to the model's earlier answers. The parts of those answers that the user picks up, quotes, and pushes on ("expand step 3", "more detail on the second method") point to the real objective. The framing that the user never returns to is chaff.
The core claim is that the user's follow-up behavior gives a monitor a cheap, explicit signal that full-transcript judges and per-turn classifiers both ignore. We will build a reply-weighted filter that scores each segment of the assistant's past answers by how much subsequent user turns draw on it. The filter then hands an off-the-shelf guardrail or judge a short "pressed payload" view instead of the raw transcript. If the claim holds, this view should recover the latent harmful objective more often and block attacks earlier than current monitors, without raising false refusals on benign technical conversations.
Where the idea comes from and what it is not
The supplying method comes from reply-based salience in collective-intelligence filtering. Generating Overview Summaries of Ongoing Email Thread Discussions builds a "comparison vector" from the replies to an email and uses cosine similarity against it to find the central issue sentence. This lifts recall from 24.3% for a lead-sentence baseline to 62.2% with a centroid of replies, and to 70.3% with a hybrid vote. VSoLSCSum supplies the lexical premise: readers reuse source vocabulary when they respond. Abstractive Text Summarization by Incorporating Reader Comments supplies the denoising lesson: raw reactions are noisy and must be down-weighted before they guide salience. Its ablation shows that removing the reader-focus tracker drops ROUGE-L from 27.16 to 20.80. Socially-Informed Timeline Generation for Complex Events supplies the two-way view: source importance and reaction relevance are estimated jointly. We use only the cheap one-pass version of that idea, not its bipartite optimization.
The motivating evidence comes from the safety side. ObjexMT shows that LLM judges can flag a dialogue as unsafe but cannot state what the attacker was after. Extraction accuracy is about 81% on human-written multi-turn jailbreaks and falls to about 24% on automated ones, and the best models score only around 0.60 overall. LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet shows that persistent humans reach 51.7–87.9% attack success against defenses that automated attacks barely dent, and that over 91% of their successes need more than one turn. Persistence, meaning repeated pushing on a fragment, is exactly the behavior our filter reads. SCCD gives a parallel from human harassment detection: stripping the reply context drops GPT-4's recall from 89% to 64.5%. Reactions carry information that the source text alone does not.
The closest existing work differs in what it does with dialogue structure. WildTeaming at Scale mines crowd jailbreak tactics but only for single-turn training, and it names multi-turn as future work. JailbreakHunter surfaces multi-turn persistence in public logs through human visual auditing, not an automated monitor. One-Shot is Enough compresses multi-turn attacks into single structured prompts. It compresses the user side for offense; we compress the assistant side by what the user extracted, for defense. One Turn Too Late (TURNGATE) reads the candidate response together with the history to find the earliest harm-enabling turn, but it treats the history as undifferentiated context rather than as a reply structure. Cross-Session Decomposition Attacks retrieves a user's sessions by shared latent intent. That work is complementary: it gathers the pieces, whereas we rank which pieces the user actually pressed on.
We are honest about scale. Inside one dialogue the "crowd" is a single attacker, so the borrowed principle is reply-based salience, not crowd aggregation. That is the part of collective-intelligence filtering whose evidence transfers. The crowd-scale version, which would aggregate follow-up patterns across many logged conversations, is a later extension and is not the claim tested here.
Core Claim
Split each past assistant turn into segments $a_i$, such as sentences or numbered steps. Let $u_j$ be the user turns that come after segment $a_i$. The follow-up salience of a segment is
$$s(a_i) = \sum_{j > i} w_j \, \mathrm{pickup}(u_j, a_i),$$
where $\mathrm{pickup}$ measures how much $u_j$ draws on $a_i$, and $w_j$ is a denoising weight that suppresses content-free follow-ups such as "thanks" or "continue". Pickup combines three signals: lexical reuse, explicit pointers such as ordinals, quotes, or "the second one", and embedding similarity. A small model is trained to resolve which segment a follow-up refers to.
The monitor receives the top-scoring segments, plus the user turns that pressed on them, as a compact view under a fixed token budget. That view is then classified by an unchanged guardrail or judge.
The testable prediction is that in successful jailbreaks, follow-up salience concentrates on the harmful fragment. The benign framing that the attack front-loads should receive little pickup. If this holds, the distilled view should raise objective-extraction accuracy and move blocking earlier.
Research Program
The work has three stages.
The first stage is a diagnostic that can kill the project before anything is built. Using human-written jailbreaks from the Multi-Turn Human Jailbreaks (MHJ) dataset (537 conversations), plus the automated SafeMTData sets inside ObjexMT, annotators mark which assistant segments contain the harmful payload. We then measure whether follow-up salience ranks those segments above two cheaper signals: position (recency) and whole-dialogue centrality. Whole-dialogue centrality is a centroid of all turns. It is the "aspect bias" baseline from summarization, because it favors frequent content rather than demanded content. If follow-up salience does not beat recency here, the project stops.
The second stage builds the filter. Lexical and pointer features are computed in one pass, and the reference resolver is a small encoder. Its supervision comes from synthetic dialogues in which a follow-up is generated to target a known segment, so no harmful labels are needed. Denoising weights are fit on benign multi-turn chats, where many follow-ups are generic.
The third stage plugs the filter into existing monitors as an input transformation. The monitors are a guardrail classifier, an LLM judge, and the TURNGATE turn-stopping policy. Because the filter changes only what the monitor reads, any gain can be attributed to the view and not to a new classifier.
The obvious attack on this defense is decoy pressing: the attacker asks follow-ups about harmless segments to dilute the salience mass. We treat this as a first-class experiment. An adaptive attacker gets several extra turns and full knowledge of the filter. The expected trade-off is that decoy pressing costs turns and still requires at least one pointed follow-up on the payload, which receives nonzero salience. Whether that holds empirically is an open question, not an assumption.
Baselines
The naive reference point is a per-turn input/output guardrail run independently on each user turn and response, with no history. This is what most deployments do. The file gives no number for this setting on MTID. On M2S-consolidated attacks, One-Shot is Enough reports a 71% bypass rate for Llama Guard 3. That is a different input format, so we cite it only as context. Beating this baseline is necessary but not interesting.
The second baseline feeds the full transcript to an LLM judge, which is the setting ObjexMT evaluates. The reported extraction accuracies are about 0.61 for the best models overall, about 81% on MHJ, and about 24% on SafeMTData_Attack600. We run the same judge models with the full transcript and with our distilled view, scored by the same similarity protocol and the same $\tau^* = 0.66$ threshold. Beating this baseline means a statistically significant extraction gain on the automated sets, where the gap is largest, with no loss on MHJ.
The strongest existing method is TURNGATE from One Turn Too Late. On its MTID benchmark it reports an F1 of 0.699, a harmful score $\phi_2$ of 0.602, and an early-block miss rate of 17.7%. We use its released setup and backbone, and we retrain it with the same data and tuning budget in two versions: one whose history input is replaced by our distilled view, and one where the view is added alongside the history. Beating it means higher F1 and a lower early-miss rate on MTID at matched benign false-block rate, measured on MTID's benign hard negatives.
The ablations isolate the key ingredient, the direction of replies. Each ablation produces a distilled view of the same token length from a different rule: the last $k$ turns, whole-dialogue centrality without reply direction, a random selection, and an LLM-written summary of "what the user wants". Every ablation is classified by the same monitor. If whole-dialogue centrality or recency matches follow-up salience, the reply idea contributes nothing and we will report that.
Why not just…?
Why not just give a strong LLM judge the whole transcript and tell it to watch what the user keeps asking about? This is the most serious threat, because a capable model might perform reply-tracking implicitly from a prompt. ObjexMT's roughly 24% extraction on automated attacks suggests full-transcript reading does not do this reliably. SafeDialBench's observation that reasoning models notice risk and then reason into compliance for coherence (SafeDialBench) suggests that long context actively pulls judges toward the attacker's framing. Still, we include a "focus on follow-ups" prompted judge as an explicit baseline. If it matches the trained filter, the honest conclusion is that the contribution is a monitoring prompt, and we will say so.
Why not just use TURNGATE or another response-aware turn monitor? TURNGATE already reads responses, and it is the bar to beat. Our filter is designed to be added to it, not to replace it. The experiment that settles the question is the TURNGATE-with-view versus TURNGATE-alone comparison on MTID at equal tuning budget. If the view adds nothing there, the reply signal is redundant with what response-aware training already learns.
Why not just flatten or summarize the dialogue and run a single-turn classifier? One-Shot is Enough shows that flattening multi-turn attacks into structured single prompts makes them more effective against guardrails, not less. A summary that preserves the attacker's framing therefore inherits the attack. An LLM-written summary of user intent is also an objective-extraction step, which ObjexMT shows is where judges fail. The LLM-summary ablation tests this directly against follow-up salience at equal length.
Why not just align the model for multi-turn safety with trajectory-level reinforcement learning? MTSA does this and reduces Zephyr-7B's attack success from 74.50% to 23.50%, with a 5.62% rise in XSTest over-refusal. That approach changes the model; ours is a monitor usable on closed models and is orthogonal. We do not claim to beat MTSA. We do report whether the monitor adds protection on top of an MTSA-style aligned model where one is available.
Proposed Experiments
Objective extraction runs on ObjexMT's 2,817 instances, with accuracy computed under its protocol and calibration reported (ECE and Brier). Detection and timing run on MTID, measuring F1, $\phi_2$, and early-block miss rate. Attack success is measured by replaying MHJ and MultiBreak conversations (MultiBreak, 10,389 prompts) against a guarded target. Over-refusal is measured on XSTest (XSTest), where Speculative Safety Honeypot reports 0% false refusals and is the reference to stay near, and on MTID's benign technical hard negatives. Those hard negatives are the real risk case, because benign experts also press on technical details. SafeDialBench supplies a consistency check across its 3–10-turn dialogues and seven jailbreak types.
Three results would decide the project. First, in the diagnostic, follow-up salience must place the annotated payload in the top segments significantly more often than recency or whole-dialogue centrality, on both human and automated attacks. Second, on ObjexMT's automated subsets, the distilled view must raise extraction accuracy above the full-transcript judge, and on MTID it must lower TURNGATE's 17.7% early-miss rate at matched benign false-block rate. Third, under the adaptive decoy-pressing attacker, we measure how much salience mass the payload retains and how much the attacker's success rate recovers. A defense that collapses once attackers know about it is a finding, not a method, and we would report it as such.
References
- Generating Overview Summaries of Ongoing Email Thread Discussions
- VSoLSCSum: Building a Vietnamese Sentence-Comment Dataset for Social Context Summarization
- Abstractive Text Summarization by Incorporating Reader Comments
- Socially-Informed Timeline Generation for Complex Events
- ObjexMT: Objective Extraction and Metacognitive Calibration for LLM-as-a-Judge under Multi-Turn Jailbreaks
- Uncovering Model Vulnerabilities With Multi-Turn Red Teaming
- SCCD: A Session-based Dataset for Chinese Cyberbullying Detection
- WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models
- JailbreakHunter: A Visual Analytics Approach for Jailbreak Prompts Discovery From Large-Scale Human-LLM Conversational Datasets
- One-Shot is Enough: Consolidating Multi-Turn Attacks into Efficient Single-Turn Prompts for LLMs
- One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue
- Cross-Session Decomposition Attacks: Scaling Risk and Intent-Aligned Retrieval Defense
- SafeDialBench: A Fine-Grained Safety Evaluation Benchmark for Large Language Models in Multi-Turn Dialogues with Diverse Jailbreak Attacks
- MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming
- MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety
- XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models
- Speculative Safety Honeypot: Toward Proactive Defense Against Multi-turn Agent Attacks