acceptodds
Under review as a conference paper at ICLR 2027

What Is a Conversation Made Of? A Reusable Semantic–Relational State Across Safety Tasks

Abstract

Multi-turn safety prediction tasks in an agentic system all rely on the same underlying state that accumulates over a conversation: what has been introduced and remains salient, what the participants are trying to accomplish, what restricts them, and how these relate. Systems for these tasks are typically trained and evaluated one task at a time, even though multi-task and frozen-backbone architectures are common more broadly. We show that a single lightweight conversational state can be induced once, frozen, and reused across independent safety tasks and across the distribution shifts each task encounters, without retraining the representation. A shared mention encoder (M1) and a graph inducer (M2) together extract open-vocabulary entity, action, goal, and constraint nodes with persistent identity and typed relations; this M1+M2 inducer is trained once (419M parameters), frozen, and read by three small (sub-million-parameter) readers at the granularity of a whole trajectory, a single proposed action, and a single turn boundary. Our central evidence is controlled transfer under distribution shift rather than a single benchmark score. For multi-turn jailbreak detection we show that the standard generic-benign split admits a large topical shortcut (a first-turn topic probe alone reaches 0.87 AUROC), and then hold the reader to topic-matched benign twins, naturally occurring sensitive conversations, over-refusal cases, and three unseen attack families; on the twins the frozen structured state reaches 0.84 AUROC against 0.71 for a capacity-matched dense reader over the same jointly trained encoder. For agent action authorization, a single operating point selected only on InjecAgent transfers to a live same-harness AgentDojo evaluation at 6.1% targeted attack success and 74.2% benign utility (and 1.9% / 70.9% at a high-security threshold), and remains effective across suites, attacks, and a change of agent model. The representation is frozen throughout: across held-out hard negatives, attack families, live agent environments, and agent-model changes, the same state is reused without representation retraining. Controls rule out shortcut and encoder-only explanations, and matched occlusions and retrains show each task draws on the semantic and relational channels of the state to a different degree.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
multi-turn jailbreak detection, conversational state representation, agent action authorization, dialogue topic segmentation, representation transfer, frozen representations, semantic-relational graph induction, llm safety

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_sn7dOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that past methods developed for safety tasks (e.g., jailbreak detection, agent action authorization) start from the same base conversation (natural-language) input. Instead, this paper motivates representing conversations with a universal, semantic state. This state can take the form of a graph-like structure, where entities and cross-entity relationships are established as a conversation proceeds. The work proposes a learned Markovian updater to continually revise the state (turn-by-turn). Once the state representation is trained, it is frozen and task-specific readers accept this state representation for use in some downstream task.

Critical strengths
  • (S1) Compelling motivation. The motivation behind a universal, Markovian state representation of a conversation is compelling and has many applications (e.g., addressing the long-context problem in LLMs). The graph-like representation introduced in the method also seems compelling.
Critical weaknesses
  • (W1) Largely unreadable. I feel pessimistic that any human reader will gain benefit from this paper. While I feel that the motivation for the paper is strong, as a reader, I am left with many doubts and unanswered questions on the method and the experimental results. Even after spending time manually reading the main paper, I am unable to understand how the method works (specifically, the training), and comprehending the experimental setup and results is difficult. I put a non-exhaustive list of examples below to illustrate:
    1. Reliance-versus-value decomposition in line 151 is introduced without context. Not sure what this means.
    2. Is line 152 grammatically correct? I’m unable to understand what this means.
    3. Matched occlusions (line 154) is not defined.
    4. It is unclear which parts in 2.1 and 2.2 are M1 and M2. How do M1+M2 differ from the complete inducer F_theta (referenced in line 219)
    5. Topical shortcuts are never defined. It makes the experimental setup in lines 240-250 hard to understand. What are topic-matched negatives? Please define AUROC, TPR, FPR.
    6. Table 1 is introduced on page 2, and only referenced in page 9 (related works). Table 2 is never referenced in the main text. It will be difficult for readers to understand these tables without any context.
    7. What are cross-paper defense numbers (line 256)?
    8. How are the primary metrics defined in line 263 defined> Specifically, benign utility & utility under attack.

I am not trying to be annoying. I merely want to provide evidence to convey my point. I strongly encourage authors to consider a substantial re-write of the paper so that readers can benefit from the main ideas and intuitions presented in this work.

  • (W2) Training objectives & loss functions are fully unspecified in the main text. I deem this to be a crucial methodological component. It's wholly unclear how training is done on M1+M2 (e.g., loss function, source of supervision, etc) and there are similarly no details on how fine-tuning of the reader is done. There seems to be some text about this in the Appendix, but these sections are never referenced in the main text (as far as I was able to tell), and the Appendix is ~35 pages long, so it is unreasonable to expect the reader to go hunting for important methodological details there.
Questions for authors

No questions.

6 comments

Part 1/2

We thank the reviewer for the careful reading and for finding the motivation compelling. We agree that the main text assumed some familiarity with both the terminology and the experimental setup. We have added definitions, clarified the methodological components, and made the connections between the experiments and their motivation explicit. Below we address each readability concern. Part 2 summarizes the training recipe now included in the main text, with implementation details in the appendices.

W1. Readability: point by point

  1. “Reliance-versus-value decomposition.” Reliance measures how much a trained reader's performance changes when a state channel (e.g., edges) is removed at inference. Value measures its contribution by comparing matched readers trained with and without that channel. A reader may rely on edges at inference even if a reader retrained without them performs similarly. Both terms are now defined at first use, with results reported side by side in Section 4.

  2. The contributions sentence. We replaced it with: “We separate what transfers from what inflates transfer using three controls: reader-level versus representation-level exposure, shortcut probes, and position-matched evaluation.”

  3. “Matched occlusions” and “matched retrains.” An occlusion removes a channel (edges, node semantics, or node types) at inference from an already trained reader. A matched retrain trains a reader of the same size and with the same procedure, but without that channel from the start. These are now defined where first mentioned.

  4. M1, M2, and the inducer. M1 is the mention encoder (Section 2.1); M2 is the graph inducer (Section 2.2). Together they form the jointly trained 419M-parameter inducer, F_θ, which is subsequently frozen. A notation paragraph at the start of Section 2 and the description of Eq. (2) now make this explicit.

  5. Topical shortcut, topic-matched negatives, AUROC, TPR, and FPR. A topical shortcut is a feature that distinguishes the evaluated classes without capturing the intended attack behavior. Here the first user turn's topic is predictive because attack conversations concern harmful topics while MT-Bench-101 benign conversations often concern homework or coding. A first-turn TF-IDF probe scores 0.87 AUROC on that generic split but only 0.53 against topic-matched benign twins: conversations on similar topics without the harmful objective. AUROC measures how often a random positive is ranked above a random negative (0.5 is chance). TPR at 1% FPR is the fraction of attacks detected when 1% of benign conversations are flagged. A new “Metrics” paragraph in Section 3 defines these measures.

  6. Tables 1 and 2. Table 1 (positioning against published systems) is now cited and introduced in Section 1. Table 2 (same-backbone comparison) is cited there and discussed in Section 4. Every main-text table now has an explicit in-text reference.

  7. “Cross-paper defense numbers.” These refer to results reported in other papers under potentially different AgentDojo versions, task and injection sets, agent snapshots, and thresholds; they are therefore not directly comparable. The revised text explains why we instead rerun competing defenses in the same live harness (Table 20).

  8. Benign utility, utility under attack, and targeted ASR. Benign utility is the fraction of legitimate tasks completed with the defense enabled and no attack present; utility under attack is the fraction completed when injected instructions are present. Targeted ASR is the fraction of runs in which the attacker's injected objective succeeds. These definitions now appear in the Section 3 “Metrics” paragraph rather than only in Appendix G.

Each task-setup paragraph in Section 3 now points to the appendix sections covering its data, reader, and evaluation protocol. We have revised Sections 1–3 accordingly.

Part 2/2

W2. Training objectives and loss functions

We agree that the training procedure is central and should not have to be reconstructed from the appendices. A new “Training” paragraph in Section 2 summarizes the supervision, joint objective, optimization, and reader training, with implementation detail in Appendices C (M1/M2) and E (readers).

Supervision (upstream training without safety labels). M1+M2 is distilled from a structured teacher that labels typed mentions, persistent-node assignments, and typed relations (Appendix C.1). The 64,888-conversation corpus combines Chatbot Arena dialogue and Claude-generated synthetic conversations (Appendix C.2). No jailbreak, injection, or topic-shift labels are used to train M1+M2. After the inducer is frozen, the three readers are trained separately on their respective task labels. Teacher-label noise is analyzed in Appendix C.7.

Joint objective. The heads are jointly optimized, with gradients flowing into the shared ModernBERT-large encoder:

�0��0�

  • LbioL_{bio} (M1): focal token classification (γ = 2) over BIO tags for the four mention types; class weights O = 1, B = 2, I = 1.5, with inverse-square-root frequency weighting across types.
  • LattL_{att} (M2, identity): cross-entropy over legal existing-node candidates, NEW, and UNRESOLVED, supervised by the teacher's persistent graph.
  • LrelL_{rel} (M2, relations): multi-label binary loss over nine relation types, using a learned adaptive threshold per pair (following ATLOP) and type-conditioned expert scoring.
  • LcompatL_{compat}: an auxiliary loss encouraging agreement between attachment and relation predictions.

Optimization. AdamW; learning rates 2×10⁻⁵ (encoder) and 1×10⁻⁴ (heads); six epochs; checkpoint selected by validation mention-plus-relation F1 (Appendix C.4). M1+M2 is then frozen before training any task reader.

Reader training (Appendix E). Each reader uses the frozen induced state. A node's 1,034-dimensional input comprises its 1,024-dimensional span embedding and 10 structural features, followed by an MLP to 128 dimensions and two R-GCN layers over nine relation types. Readers use focal cross-entropy (γ = 2) on task labels and a learning rate of 3×10⁻⁴, with task-specific readouts and schedules:

  • Jailbreak (929K parameters; Appendix E.2): three-view pooling, causal GRU, a Pre head for user turns and a Screen head for assistant turns; four epochs, validation-F1 checkpoint; trained only on Crescendo.
  • Action Guard (24.8K parameters; Appendix E.3): action node, mean goal node, and mean constraint node passed to a two-layer head; six epochs, trained only on InjecAgent. The checkpoint and 3% false-block target threshold are chosen on InjecAgent validation, without using AgentDojo.
  • Topic shift (929K parameters; Appendix E.4): jailbreak-style pooling and GRU; six epochs, trained on SuperDialseg, selected on DialSeg711, and evaluated on TIAGE.

Inference. For each new turn, M1 extracts typed mentions; M2 links each to an existing persistent node, creates a new node, or marks the mention UNRESOLVED, and predicts typed relations. The task-specific reader makes its prediction from the updated state without modifying the frozen inducer.

Revisions. Section 2 now includes notation and training summaries; Section 3 defines the metrics and points to task-specific appendices; and every table is cited at first mention. We thank the reviewer again and hope these changes make the method easier to understand and evaluate. All revisions added in the latest uploaded version.

I thank the authors for the efforts.

The introduction has certainly improved. However, I’m still concerned about the paper's readability. It feels like the paper conducts a large array of experiments, but seldom connects / curates / condenses the findings. The result is that a reader needs to continuously jump across many sections when reading the paper to understand what's going on. Even if the method is strong, this makes it hard for readers to parse the paper and take away insights for their own problems.

At this point, I have independently read the paper 4 times, and while the re-written manuscript helps clarifies previously raised concerns, it has also surfaced new issues:

Significant

  1. Reproducability concern: It is not known which frontier-LLM interface is used to generate the teacher labels. The prompts used to generate teacher labels is not provided. No link to the training dataset is provided. I’m doubtful that any independent researcher would be able to reproduce this training dataset, which seems to be a critical part of the training.

  2. Concerns about Table 2: it seems possible that the fine-tune baseline is overfitting, which may explain the lower results in Table 2. Especially since all weights in the model are trained. The paper itself has stronger backbone-controlled baselines in Table 43 of the Appendix (K3) which would be better to place in Table 2 (frozen backbone + trained “reader”-head. I am confused why the authors did not use these results in Table 2.

  3. Presentation: Section 2.1 should probably explain that per-token BIO heads for each token type help identify the span over each extracted mention m. The BIO has are mentioned only in the training text (line 235).

  4. Presentation: An image showing architecture + training would be helpful in clarifying the method described in Section 2.

Minor

  1. What are the “10 structural features” in line 246? That’s the first time the term “structural features” is used.

  2. The role of Pre/Screen heads in line 248 are not clearly defined.

  3. “InjecAgent” not defined when it is first introduced in line 249

  4. Table 1: The AUROC label is in the wrong task

  5. The x1..xkx_1..x_k in conversation prefix (line 166) are not defined. Is it different than t1,t2,t3t_1, t_2, t_3 used in Fig. 1? If they are the same, please unify the notation.

  6. Unclear what LcompatL_{compat} is, even after checking Appendix.

At this point in the review phase, I still recommend the paper for rejection, and I strongly suggest the authors to rewrite their paper. The readability issues continue to interfere with my ability to assess the methodological & technical strengths of the work. The authors may benefit from using an outline like so:

  • Introduction (this section is largely fine, though Table 1 and 2 should probably be moved to results)

    • Motivate a persistent state
    • Introduce and name the three tasks this paper focuses on (multi-turn jailbreaks, agent action authorization, topic-shift detection).
    • Fig. 1
    • Contributions
  • Method

    • Notation
    • Mention Encoder (M1)
    • Graph Inducer (M2)
    • Training Dataset Construction
    • Training Loss Functions
    • Design of Readers (probably better to defer per-task specific details)
  • Main Benchmark

    • Task Description + Metrics
    • Main Results (Table 1 and 2)
  • Transfer Evaluation

  • Ablations & other analysis

    • intrinsic quality of M1+M2 inducer (Table 3)
    • Occlusion vs. matched retrains
    • deployment cost (can probably be reduced to one paragraph)
  • Related Work + Limitations + Conclusion

If the authors are able to respond in the remainder of the discussion period, I will make my best effort to review their response and revise my assessment, but time is limited and re-reviewing the paper takes significant effort on my part. Rather than rushing, it may be better for the authors to take the time to work on the manuscript and re-submit at a future venue.

Reviewer_zmpNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper induces a causal, turn-by-turn conversational graph once and reuses it, frozen, across safety tasks. Nodes are ENTITY, ACTION, GOAL and CONSTRAINT mentions with persistent identity, joined by nine typed relations. Small readers on top handle multi-turn jailbreak detection, agent action authorization and topic-shift detection. The paper also shows that the standard ActorAttack-vs-MT-Bench-101 split is topic-confounded and builds topic-matched hard negatives. Against these, the structured state outperforms a capacity-matched dense reader on the same frozen encoder (0.84 vs. 0.71 AUROC).

Critical strengths

Well-motivated idea. Multi-turn attacks spread intent across individually benign turns, so a persistent, accumulated state is a natural fit for defending against them.

The first-turn topic probe finding. A TF-IDF probe on the first user turn alone reaches 0.87 AUROC on the standard ActorAttack-vs-MT-Bench-101 split but drops to chance (0.53) on topic-matched negatives. This exposes a real confound in a common evaluation setup, and the hard-negative sets are useful to the community. The finding also suggests attack and defense directions worth exploring further.

Critical weaknesses
  1. No comparison to alternative structured representations. The central claim is that this specific semantic–relational state is what transfers. Yet the only non-LLM representation baselines are structure-free: mean-pooled turn embeddings with a GRU, or a full fine-tune. Two kinds of alternatives are missing: prior graph-based or stateful multi-turn detectors the paper itself cites (G-Guard, DeepContext); less rigid graph constructions, such as an open-schema LLM-extracted graph (GraphRAG- or OpenIE-style). Without these, it is unclear whether the fixed four-type, nine-relation ontology matters, or whether any reasonable entity graph would do. This is especially relevant for jailbreak. There, disabling edges costs only 0.029 AUROC while removing node semantics costs 0.144 (Table 32). That suggests the gain may come from pooling over extracted mentions rather than from the relational structure.

  2. The jailbreak threat model is non-adaptive. The reader is trained on Crescendo only and evaluated on fixed public attack datasets. No learned attacker (e.g., SEMA, X-Teaming) is used, and no attack is optimized against the detector itself. The paper’s own Action Guard results show the risk: a goal-aligned adaptive attack roughly doubles ASR (6.1% → 11.8%). A structure-based jailbreak detector has an analogous vulnerability, namely attacks that keep the graph attached to a benign-looking goal.

  3. The jailbreak evaluation is not goal-oriented. Detection is reported only as AUROC on labeled conversations. Training positives are attack attempts, not verified successes (Table 12). There is no end-to-end evaluation of whether the guard actually prevents harmful completions, and at what cost to benign users. This contrasts with the live ASR and utility evaluation done for agents.

Questions for authors
  1. With the reader architecture held fixed, how does the frozen state compare against other structured representations, such as G-Guard’s entity graph or an open-schema LLM-extracted graph?

  2. How robust is the jailbreak reader to adaptive attackers? This covers a learned attacker such as SEMA, and a variant whose reward includes evading the guard. Does training the reader on such attacks recover performance?

  3. Can you report a goal-oriented jailbreak evaluation? That means deploying the guard in front of a target model and measuring ASR on verified harmful completions, together with the benign refusal rate.

Reviewer_cfZROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a reusable semantic-relational representation for multi-turn conversations. The framework extracts entities, actions, goals, and constraints into a dynamically updated graph with persistent node identities and typed relations. The learned representation is frozen and reused across three downstream tasks: multi-turn jailbreak detection, agent action authorization, and topic-shift detection. Experiments across multiple datasets demonstrate improvements over dense representation baselines, particularly under distribution shifts.

Critical strengths
  1. Interesting research direction. The paper explores a meaningful problem of learning reusable conversational representations across multiple safety tasks, potentially reducing the need for task-specific representation training.

  2. Comprehensive experimental evaluation. The authors conduct extensive experiments across different datasets and distribution shifts. The controlled comparisons and ablation studies provide useful evidence for the effectiveness of structured representations.

Critical weaknesses
  1. Limited methodological novelty and insufficient comparison with recent graph-based methods. The proposed framework mainly combines established techniques, including semantic extraction, entity tracking, and graph induction. Its methodological novelty remains unclear. Moreover, several recent graph-based LLM/agent safety studies, such as G-Safeguard, GUARDIAN, IPIGuard, and XG-Guard, are not sufficiently discussed. The limited discussion and lack of comparisons with closely related structured safety methods, particularly for agent action authorization, makes it difficult to assess the proposed framework's advantages over existing approaches.

  2. Limited evidence for the general effectiveness of the shared structured representation. Although the paper emphasizes a reusable semantic-relational representation, the benefits of graph structure vary considerably across tasks. Relational information is crucial for action authorization but contributes less to jailbreak detection, while the improvement on topic-shift detection is relatively modest (0.612 vs. 0.582 AUROC). These results raise questions about the extent to which the proposed structured representation offers substantial advantages across diverse downstream tasks.

Questions for authors
  1. What are the key methodological innovations of the proposed framework compared with recent graph-based LLM safety methods, such as G-Safeguard, GUARDIAN, IPIGuard, and XG-Guard? Could the authors clarify these differences and provide more relevant comparisons?

  2. The contribution of graph structure varies considerably across the three tasks. Why does relational information provide substantial improvements for action authorization but relatively limited benefits for jailbreak and topic-shift detection? How does this affect the claimed generalizability of the proposed representation?

Reviewer_6EN8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether a unified dialogue representation can be reused across different tasks. The authors propose a 419M parameter model that extracts entities, actions, goals, constraints, and their relations from multi-turn dialogues. After freezing this representation, lightweight readers are trained for jailbreak detection, agent action authorization, and topic-shift prediction. Experiments show that the method exhibits strong transfer performance across attack families, datasets, and agent environments, particularly on jailbreak detection and action authorization.

Critical strengths
  1. The paper addresses a meaningful question: whether different safety tasks can share a unified dialogue state. Freezing the representation and training lightweight task-specific readers is practically appealing.
  2. The hard-negative evaluation is well designed. On topic-matched benign twin samples, the first-turn topic probe drops from 0.871 to 0.531 AUROC, while the proposed method reaches 0.841 (dense baseline: 0.709). This provides solid evidence that the approach goes beyond simple topic recognition.
  3. The experimental evaluation is relatively comprehensive, covering cross-family transfer, edge ablation, position matching, and online comparison on AgentDojo. In the online evaluation, the method reduces the attack success rate (ASR) from 52.1% to 6.09%, while benign utility only drops from 77.33% to 74.22%.
Critical weaknesses
  1. Novelty. Prior work (e.g., G-Guard, DeepContext) has already explored graph-based dialogue representations and stateful multi-turn safety detection. The main contribution here appears to be the cross-task reuse of a shared representation. Moreover, the dense baseline uses mean-pooling of turn embeddings, whereas the proposed method retains multiple semantic nodes and explicit relations. This makes it difficult to isolate the benefit of the graph structure. Although the authors provide edge ablations, the edge-free Action Guard only achieves 0.528 AUROC, while Dense+GRU reaches 0.764. A stronger graph-free baseline that uses the same semantic nodes would better demonstrate the necessity of relation modeling.
  2. Certain aspects of the generalization evaluation. The authors carefully control several shortcuts, yet position matching still reduces Action Guard’s AUROC from 0.951 to 0.825, and results of the dense baseline under the same setting are not reported. In addition, Table 12 contains inconsistent descriptions of the Crescendo training corpus, and the overlap between training and evaluation seed targets is not quantified. These issues make it hard to fully assess the claimed generalization to unseen attack goals.
  3. The practical effectiveness and robustness. On the standard jailbreak split, TPR is only 26.1% at 1% FPR and 47.4% at 5% FPR (although this split itself suffers from topic confounding). Low-FPR detection performance on hard negatives deserves more attention. Furthermore, the extraction F1 scores for GOAL and CONSTRAINT are only 0.46 and 0.55, respectively, yet the impact of these errors on downstream authorization is not clarified. The improvement on TIAGE is also modest (0.612 vs. 0.582 AUROC), providing limited additional evidence for broader task generalization.
Questions for authors
  1. What is the primary methodological novelty relative to existing graph-based or stateful dialogue models? How does the proposed method compare to a stronger graph-free reader that uses the same extracted semantic nodes?
  2. Can the authors clarify the Crescendo training data and quantify the overlap between training and evaluation seed targets? Can the dense baseline be evaluated on the position-matched Action Guard set?
  3. How does the method perform at low false-positive rates under the hard-negative distribution? What is the impact of GOAL and CONSTRAINT extraction errors on downstream action authorization?
  4. Can the authors provide measured inference latency and memory footprint, including how the computational overhead scales with dialogue length and graph size?
7 comments

Part 1/5

We thank the reviewer for the careful review and for recognizing the hard-negative evaluation, the breadth of the transfer tests and the practical appeal of a frozen state read by small readers. New experiments are marked [new]. Parts 1–2 address W1/Q1, Part 3 W2/Q2, Part 4 W3/Q3 and Part 5 Q4, followed by the revisions. Table numbers are those of the revised PDF.

I. Novelty relative to G-Guard and DeepContext (W1, Q1)

What is new. We agree that neither graph-based nor stateful multi-turn detection is new. §6 says so. Three things are new.

  • (i) The state. One learned inducer turns raw dialogue, causally and turn by turn, into open-vocabulary entity, action, goal and constraint nodes with persistent identity and nine typed relations. It makes no LLM calls, is trained once without safety labels, and is then frozen.
  • (ii) Reuse. The same frozen state serves three decisions with different prediction units: a trajectory, a proposed action and a turn boundary. Each task trains only a reader of 24.8K–929K parameters.
  • (iii) The test. Under two levels of exposure control, shortcut probes and matched retrains measure which part of the state each decision uses. The first-turn topic probe is one result of this protocol.
G-GuardDeepContextOurs
StateGNN over a graph of the conversation's queries and harmful keywords, plus the most similar labeled reference queryrecurrent state over turn embeddings; no objectstyped nodes with persistent identity, 9 typed relations
Built bykeyword and query graph from the text; reference retrievalencoder fine-tuned for the task419M inducer, trained once without safety labels, frozen
Decisionsjailbreakjailbreakjailbreak, action authorization, topic shift
Trained per taskGNNwhole modelreader only

Head to head [new]. We re-implemented both from their papers and trained them on the same Crescendo data (also in our response to Reviewer zmpN, W1/Q1).

System (AUROC)GenericTwinsWildChat-sens.Over-refusal
Dense reader0.8410.7090.740.70
G-Guard [new]0.8770.6680.6900.641
DeepContext-style [new]0.8530.7140.7300.698
Ours0.8940.8410.8470.831

G-Guard comes close to our state on the generic split but falls below the dense reader on every hard-negative set. Its retrieval of the nearest labeled reference query acts as a topic signal, the same shortcut that takes the first-turn probe from 0.871 to 0.531. Without the retrieval, its twins AUROC rises from 0.668 to 0.693. Read by our own reader, G-Guard's entity graph reaches 0.776 on twins and 0.814 on AgentDojo, against 0.841 and 0.951 for the induced state.

DeepContext-style stays within 0.012 of our dense reader on every set. A recurrent state over turn embeddings, fine-tuned or frozen, does not supply what the induced nodes add.

Part 2 takes up the sharper question in Q1: what do the same nodes give without edges?

Part 2/5

II. A stronger graph-free reader over the same nodes (W1, Q1) [new]

The reviewer is right that 0.528 is not a fair graph-free comparator. That reader (Table 28) differs from the dense control in two ways at once. It has no edges, and it also has no history GRU (App. E.3, K.3). It sees the action node and the mean of all goal and constraint nodes, but not which links to which.

We therefore built three stronger readers over the same frozen nodes. All keep M2's persistent identity, use no induced edges, and have at most the full reader's parameters.

  • (A) Nodes + history GRU (Action Guard). The node reader, plus the dense control's causal GRU over per-turn node pools. The jailbreak reader already has this GRU.
  • (B) Node-set Transformer. 2 self-attention layers over all nodes, with type, turn and speaker embeddings, so any node can attend to any other.
  • (C) Action-to-node attention (Action Guard). The action node attends over every goal, constraint and entity node, with pair features [a; v; a⊙v; |a−v|]. This models relations, but learns them from InjecAgent instead of reading the induced ones.
Reader (no induced edges unless stated)TwinsInjecAgent val.AgentDojo
Dense + GRU (Table 5)0.7090.9680.764
Node readout, no edges (Tables 28, 48)0.8100.9410.528
(A) Nodes + history GRU [new]0.810*0.9750.779
(B) Node-set Transformer [new]0.8180.9790.836
(C) Action-to-node attention [new]n/a†0.9830.874
Full state (induced typed edges)0.8410.9870.951

* The jailbreak reader already contains the temporal GRU, so (A) on twins coincides with the no-edge node readout above (same number, not a separate run).

† (C) needs a proposed-action node to attend from; the jailbreak task has none, so (C) is defined only for the Action Guard.

Stronger node readers close part of the gap, as the reviewer suggested. On AgentDojo, (C) recovers 59% of the full state's gain over dense (0.764 → 0.874). On twins, (B) reaches 0.818, 0.023 below the full state. For jailbreak, most of the gain is already in the nodes and their persistent identity (0.810 without edges), as we also reported to Reviewers cfZR and zmpN.

Induced relations still add, and the gap appears under shift. On InjecAgent validation the three new readers are within 0.012 of the full state, so a graph-free reader with enough capacity fits the source corpus. On AgentDojo, the full state leads (C) by 0.077, and by 0.084 after position matching (Part 3). Like the raw action embedding (0.817 within AgentDojo, 0.518 across corpora; Table 37), the action-to-node matching that (C) learns is fitted to InjecAgent. The induced edges come from an inducer trained on general dialogue and do not depend on the agent corpus.

Live harness [new]. At the same InjecAgent-selected threshold, (C) reaches 12.4% ASR at 73.0% benign utility, against 6.09% at 74.22% for the full state. Its realized false-block rate is 6.8% against the 3% target (ours: 3.5%), so its calibration also transfers less well.

Part 3/5

III. Generalization evaluation (W2, Q2)

(a) Dense baseline under position matching [new]. In AgentDojo, injections arrive in tool outputs, so an unauthorized call can only come after one. Calls made before the first tool output are authorized and easy to classify. Matching drops strata without both classes, which removes those calls. Part of the 0.951 → 0.825 drop is therefore the removal of easy negatives, and part is reliance on position. Under the same matching:

ReaderFull held-outPosition-matchedChange
Position only0.7120.50−0.212
Dense + GRU0.7640.651−0.113
(C) Action-to-node attention, no edges0.8740.741−0.133
Untyped graph0.8600.780−0.080
Full state0.9510.825−0.126

The full state stays first and dense last. Reader (C) beats the untyped induced graph on the full set but falls below it under matching, so part of what it learns is position. After matching, the full state leads dense by +0.174 (+0.187 before). In separation above chance, the dense reader keeps 57% under matching and the full state 72%. The dense reader relies more on position, because its history GRU sees how many turns have passed.

(b) Table 12 (Table 13 in the revision). The reviewer is right that two rows contradict each other. "Training conversations: Full public released dataset" is wrong. The corpus is generated with an automated Crescendo implementation and is not a released split (App. F.1). Corrected:

PropertyValue
Provenancegenerated by an automated Crescendo implementation; not a released split
Attacker / targetnon-Claude open models, 7–20B (e.g. GPT-OSS-20B, Llama family)
Seed objectives200 harmful goals, 6 attack runs each
Conversations2,308: 1,200 attack attempts (52%), 1,108 generic benign
Verified successes among positives58% (dual judge; response to zmpN, W3/Q3)
Turns per conversation3–10, mean 5.6
Train / validation splitconversation-level; seed-level re-split below

(c) Seed overlap [new]. We matched the 200 Crescendo seed objectives against the seed objectives of every evaluation family. Matching was exact first. Then sentence-embedding candidates (cosine ≥ 0.80) were kept if an LLM judge rated them the same harmful goal, and two authors audited every kept pair. We report AUROC with the affected conversations removed (from the twin set, together with their twins).

FamilySame-goal seedsEval. conversations affectedAUROC, allAUROC, removed
ActorAttack vs. twins1954 of 1,673 (3.2%)0.8410.838
ActorAttack, generic split1954 of 1,6730.8940.892
MHJ / CoSafe / RedQueen11 / 7 / 92.9 / 1.6 / 2.2%0.821 / 0.800 / 0.8100.817 / 0.800 / 0.806

No seed matches exactly (0). None of the 159 HarmBench behaviors used for the adaptive attacks in our response to zmpN overlaps either. We also retrained the reader with a seed-level train/validation split, dropping the 19 matched seeds from training. It reaches 0.836 on twins, 0.889 on the generic split and 0.805 on the unseen families (mean). Removing the shared goals, at evaluation or in training, changes AUROC by at most 0.005, so shared attack goals do not explain the transfer. App. F.1 now reports these numbers in place of the "unquantified" paragraph.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Authors Comment

Dear Area Chair and Reviewers,

We sincerely thank all four reviewers for their constructive feedback. Throughout the discussion period, we have substantially revised the manuscript and conducted extensive additional experiments addressing the following concerns:

  • Novelty and stronger baselines (cfZR, zmpN, 6EN8, sn7d): Comparisons against recent graph-based defenses, alternative structured representations, stronger graph-free readers, and frozen-backbone controls.
  • Robustness and practical effectiveness (zmpN, 6EN8, cfZR): Adaptive attackers, end-to-end jailbreak prevention, low-FPR evaluation, and additional live agent evaluations, including IPIGuard.
  • Generalization and representation analysis (6EN8, cfZR, zmpN): Training-seed overlap, position-matched controls, extraction-error sensitivity, task-specific matched retrains, and measured computational efficiency.
  • Readability and reproducibility (sn7d, 6EN8): Substantial manuscript reorganization, new architectural illustrations, expanded training objectives, teacher prompts, dataset construction details, corrected data descriptions, and consolidated experimental findings.

We appreciate that Reviewers cfZR and zmpN have increased their scores following the additional experiments and clarifications. We look forward to hearing from Reviewers 6EN8 and sn7d. In particular, given the extensive revisions following sn7d's proposed outline and detailed suggestions, we would be grateful if they could reassess the manuscript, time permitting.

All changes are incorporated into the latest uploaded manuscript. Given the reorganization, a few table numbers mentioned in our earlier rebuttals may have changed. Regardless, all the experimental results in all teh rebuttals below are incoporated in the uploaded paper PDF.

We respectfully ask the Area Chair to consider the revised manuscript and the full discussion in the final assessment. We believe the paper is substantially stronger as a result of the reviewers' feedback and sincerely appreciate their time and engagement. We remain ever grateful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.