acceptodds
Under review as a conference paper at ICLR 2027

Routing Under Fire? Toward Safety-Aware Multi-LLM Serving Systems

Abstract

Large language model (LLM) routers in multi-LLM serving systems balance response quality and inference cost, yet these systems also face adversarial queries that elicit harmful responses. In existing systems, routing and safety are handled separately: routers select models based on quality and cost, while defenses such as screening, translation, and regeneration are applied as fixed pipelines outside the routing decision. However, the safety and cost of a response depend jointly on the query, the model, and the defense, so a fixed defense adds unnecessary cost and quality loss on benign traffic while under-protecting against complex attacks. We therefore study safety-aware routing, which extends the routing action space from models to model–defense configurations. We introduce S²-Bench (Safety-aware Serving Benchmark), comprising 1.4 million labeled adversarial outcomes across 10 models, 16 attack settings, and 16 composite defense configurations, and show that no single defense achieves the best safety–cost trade-off across requests and models. We then propose SafeRouter, which uses risk gating to assess the risk of each query and selects a model–defense configuration accordingly, prioritizing quality for low-risk queries and the lowest-cost predicted-safe configuration for high-risk ones. On held-out attack methods, SafeRouter achieves a 0.34% attack success rate, comparable to an S²-Bench-configured Semantic Router (0.48%) at 4.5× lower adversarial cost and 2.8× lower than the strongest evaluated single-defense pipeline, while matching the benign quality of undefended IRT-Router at comparable serving cost. Code is available at https://anonymous.4open.science/r/SafeRouter-85FC.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-LLM Routing, LLM Safety, Security-Aware Routing, Jailbreak Defense, Adversarial Robustness, Cost-Efficient LLM Serving

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9sxUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies safety-aware multi-LLM routing and proposes SafeRouter, which jointly selects both a target model and a defense strategy for each query. The authors also introduce S²-Bench, containing 1.4M labeled outcomes across 10 models, 16 attack settings, and 16 composite defense configurations. SafeRouter uses risk gating, safety prediction, and utility estimation to route low-risk queries for quality and high-risk queries toward the lowest-cost predicted-safe configuration

Critical strengths
  1. Important and well-motivated problem formulation. The paper correctly identifies that model routing and safety defense should be optimized jointly rather than independently. This is consequential because defense effectiveness and cost depend on the specific query–model–defense combination.
  2. Strong benchmark and empirical evaluation. S²-Bench is large and systematically evaluates 1.4M model–defense outcomes. SafeRouter achieves 0.34% ASR while substantially reducing adversarial serving cost, and the paper also evaluates unseen goals/attack methods and statistical significance.
Critical weaknesses
  1. No adaptive attack against the full routing pipeline. The evaluation holds out existing attack methods but does not test attacks that explicitly optimize against SafeRouter’s risk gate and safety predictor. This limits confidence in the robustness of the system under an informed adversary. The paper explicitly acknowledges this limitation.
  2. The model pool is still relatively concentrated. Although the evaluation includes ten models, most are from the Qwen3 family, with only two Nemotron models and one code-specialized variant. This leaves some uncertainty about how well the routing policy transfers to more heterogeneous commercial and open-source model pools.
Questions for authors
  1. How robust is SafeRouter to adaptive attacks targeting the risk gate or safety predictor?
  2. How sensitive are the results to the composition of the model pool, especially with more heterogeneous model families?
  3. Could the authors clarify whether the safety–cost advantage remains similar when local routing/embedding compute is included in the end-to-end cost?
1 comment
AuthorsOpenReview ↗

# Response to Reviewer 9sxU

We thank the reviewer for recognizing that the intersection of routing and safety is an important and well-motivated problem.

W1 / Q1 — Is SafeRouter robust to attacks that target the router itself?

Even when the attacker targets the risk gate and the safety predictor, SafeRouter is more robust than a single-defense pipeline and comparably robust to the Semantic Router at 4.4× lower cost on S²-Bench. We attack the deployed pipeline with PAIR [12] + DAGR operators [7] (Qwen3-235B attacker, 12 queries/goal, 40 goals). The white-box attacker additionally sees the router's risk score and threshold, the chosen model–defense pair and its predicted safety, and is told that low-risk queries are served undefended.

SystemGoals broken (of 40)
SafeRouter, black-box13 (32.5%)
SafeRouter, white-box11 (27.5%)
Semantic Router (S²-configured)16 (40.0%)
Fixed S424 (60.0%)

Knowing the gate gave the attacker no measurable advantage on these 40 goals (11 vs 13 goals): prompts that still carry the harmful goal keep a high risk score, so only 0.47% of black-box and 1.59% of white-box queries reached the undefended low-risk path.

W2 / Q2 — Does SafeRouter work with a more diverse model pool?

SafeRouter generalizes to a more heterogeneous pool without changing the method. We added Gemma-4-26B-A4B, Gemma-4-31B (Google) and MiMo-V2.6-Flash (Xiaomi), fully labelled with the same pipeline, and retrained the router on 13 models. Adversarial ASR stays below 1% in every seed, and the router uses the new models where they help: 15.8% of adversarial traffic goes to them and 69.7% of low-risk benign traffic:

System (13 models)ASRUSD/1k
SafeRouter (5 seeds)0.49 ± 0.25%0.087
Semantic Router (S²-configured)0.54%0.186
Fixed S41.41%0.249

Q3 — Does SafeRouter stay cheaper once routing compute is counted?

Including routing overhead, SafeRouter remains over 3× cheaper than the Semantic Router and Fixed S4. Routing is almost entirely the Qwen3-Embedding-0.6B encoder; the 18-net router adds 0.1–0.4 ms per query. Measured on one RTX 3090 over 1,024 S²-Bench probes (median 171 tokens), routing takes 45–56 GPU-seconds per 1k queries (batch 1 / 32): USD 0.006–0.008 per 1k at the 3090's rental rate (USD 0.50/GPU-h), or at most USD 0.03 per 1k even when billed at an H100 rate (USD 1.99/GPU-h) with the same timing; both rates are RunPod's November 2026 listings [11]. Every learned router (IRT-Router, RouteLLM, CARROT) also embeds each query. SafeRouter's adversarial cost becomes USD 0.099–0.124/1k, versus USD 0.415 for the Semantic Router (excluding its own classifiers) and USD 0.401 for Fixed S4: 3.2–4.2× cheaper.

References

[7] Diversity helps jailbreak large language models. NAACL 2025.

[11] RunPod GPU cloud pricing, https://www.runpod.io/pricing, November 2026: RTX 3090 USD 0.50/h; H100 PCIe USD 1.99/h (Community Cloud).

[12] Jailbreaking black box large language models in twenty queries. SaTML 2025.

Reviewer_T4BhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new routing framework for multi-LLM systems that jointly considers safety, quality, and cost. The paper presents a dataset spanning malicious and benign response behavior across numerous base-model/defense configurations and uses this dataset to train SafeRouter. Experiments demonstrate that this router is able to achieve lower ASR with reduced adversarial query cost while maintaining quality and cost on benign queries relative to an established baseline.

Critical strengths
  • The demonstration that using robust data over a broad pre-tested configuration grid either for training or to inform defense selection can potentially reduce both ASR and adversarial cost in a multi-model setting is novel and interesting
  • The presentation is generally quite clear
Critical weaknesses
  • The paper's proposed method rests on the assumptions that 1) generalization from attacks seen during training will be strong toward attacks seen at test time and 2) prioritizing cost over safety when multiple predicted-safe options are available will not significantly increase ASR. The current evaluation is not sufficiently robust to support these assumptions, as the tested attacks are not sufficiently diverse or strong to accurately represent current real-world settings. In particular, strong directly-applied iterative attacks (e.g. [1]) or adaptive attacks may be able to exploit the proposed configuration. Further experimentation is needed.
  • Several relevant prior works are either not cited or not discussed in sufficient detail [2][3][4]. In particular, the paper would benefit from additional discussion of how it relates to existing safety-focused routers (that may not consider a multi-model setting but still adaptively select guardrails).

[1] Zhao, W., Ben-Levi, D., Hao, W., Yang, J., & Mao, C. (2025, April). Diversity helps jailbreak large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) (pp. 4647-4680).

[2] Manczak, B., Zemour, E., Lin, E., & Mugunthan, V. (2024). Primeguard: Safe and helpful llms through tuning-free routing. arXiv preprint arXiv:2407.16318.

[3] Fang, X., Luo, X., Wang, Y., Chen, X., Tian, Y., Sun, Z., ... & Huang, S. J. (2026). SafeThinker: Reasoning about Risk to Deepen Safety Beyond Shallow Alignment. arXiv preprint arXiv:2601.16506.

[4] Cunningham, H., Wei, J., Wang, Z., Persic, A., Peng, A., Abderrachid, J., ... & Sharma, M. (2026). Constitutional classifiers++: Efficient production-grade defenses against universal jailbreaks. arXiv preprint arXiv:2601.04603.

Questions for authors

See the weaknesses.

2 comments
AuthorsOpenReview ↗

# Response to Reviewer T4Bh

We thank the reviewer for recognizing that a broad pre-tested configuration grid can reduce both ASR and adversarial cost.

W1 — Are the attacks strong enough, and does SafeRouter resist adaptive ones?

The attacks are diverse and strong, every test attack is unseen in training, and SafeRouter stays robust under the adaptive attack the reviewer suggests.

Diverse and strong attacks. S²-Bench applies 15 attack methods from six families (role-play, LLM rewriting, context manipulation, iterative optimization, random search, encoding) plus a baseline to 559 goals from three benchmarks, and labels all 8,943 probes on 10 models × 16 defense compositions. Undefended, they jailbreak 10.8–26.0% of probes per model (GPT4Cipher 45.1%).

Generalization to unseen attacks. Each fold's test attack methods are excluded from training, validation and calibration, so every reported ASR is on unseen methods: 0.40 ± 0.10% over 5 seeds, close to the 0.3% validation target. With whole attack families held out, ASR is 0.58 ± 0.15% at $0.122 ± 0.017/1k (5 seeds), versus 0.49% for the Semantic Router at 3.4× the cost; the adaptive prompts below are also unseen, as DAGR generates them against the deployed system.

Preferring cheaper predicted-safe cells. τ* is chosen so that selecting the cheapest predicted-safe cell itself meets a Clopper–Pearson bound on validation ASR. It matches the cheapest fixed cell meeting the same target on training folds (0.40% vs 0.43%) at 2.5× lower cost. Serving the safest predicted cell instead gives 0.17% ASR at $0.164/1k, versus 0.35% at $0.094 for the cheapest predicted-safe cell on the same router (the submitted seed): the target sets an explicit cost–ASR trade-off.

Iterative adaptive attacks (Sec. 5.1, Table 3). We directly attack the full system with PAIR [12] extended with the DAGR operators of [7] (the reviewer's [1]), with a Qwen3-235B attacker, 12 queries per goal on 40 HarmBench / SORRY-Bench goals; the white-box attacker also sees the risk score, threshold, chosen cell and its predicted safety. SafeRouter is more robust than a single-defense pipeline and comparably robust to the Semantic Router at lower cost:

SystemGoals broken (of 40)
SafeRouter, black-box13 (32.5%)
SafeRouter, white-box11 (27.5%)
Semantic Router (S²-configured)16 (40.0%)
Fixed S424 (60.0%)

The attacks did not push SafeRouter onto weak cells: 25 of the 28 successful queries landed on two-defense cells (mostly S1+S4), one on a single defense, and two on the low-risk path.

W2 — How does SafeRouter relate to adaptive guardrails like PrimeGuard?

These frameworks adapt the guardrail of a single model; SafeRouter instead selects a (model, defense) pair across a multi-LLM pool to minimize serving cost under an ASR target.

PrimeGuard [8]SafeThinker [9]Constitutional Classifiers++ [10]SafeRouter (ours)
Models servedone LLMone LLMone protected LLMa pool of LLMs
Per-query decisionanswer / re-evaluate / refuse (+ guidance tip)refusal / LoRA-hardened twin / guided decodingescalate from a linear probe to an exchange classifierwhich (model, defense composition) serves the query
Objectivesafety vs. helpfulnesssafety vs. utility; latencyclassifier computeend-to-end serving cost ($)
Safety calibrationempiricalempiricalempirical (red-teaming)CP-bounded validation ASR target

The revision adds this comparison to Sec. 2.2.

Comparison with PrimeGuard [8]. We ran PrimeGuard's official pipeline on Qwen3-32B on all S²-Bench probes with our judges and prices, and compare it with SafeRouter restricted to the same model, which, like PrimeGuard, adapts the defense per query. On the same model, SafeRouter matches PrimeGuard's safety at 43% of its cost (45–54% including routing overhead of $0.006–0.03/1k), because PrimeGuard adds a ~1,850-token LLM call to every request; with the full pool the gap is 3.0×:

SystemModelsASR ↓Adv. $/1k ↓
Qwen3-32B, no defense112.1%0.104
Qwen3-32B + S411.51%0.171
PrimeGuard (Qwen3-32B)10.37%0.282
SafeRouter, Qwen3-32B only10.40%0.121
SafeRouter100.35%0.094

References

[7] Diversity helps jailbreak large language models. NAACL 2025.

[8] PrimeGuard: Safe and helpful LLMs through tuning-free routing. arXiv:2407.16318, 2024.

[9] SafeThinker: Reasoning about risk to deepen safety beyond shallow alignment. arXiv:2601.16506, 2026.

[10] Constitutional Classifiers++: Efficient production-grade defenses against universal jailbreaks. arXiv:2601.04603, 2026.

[12] Jailbreaking black box large language models in twenty queries. SaTML 2025.

I thank the authors for their clearly presented response. I think these experiments, in particular the study on iterative attackers, are a good step in the right direction. However, the presented results do not alleviate my concerns regarding the potential for the proposed routing mechanism to decrease safety against more realistic attacks. If the fixed defense was selected as the strongest one available to SafeRouter, it would seem that the system would perform worse given the meaningful rate of overly weak defenses being selected. More generally, in order to build confidence in the performance of such a routing system and accurately present tradeoffs, the presentation should include more substantial experimentation against such iterative/adaptive attacks and safety routing alternatives. Accordingly, I will maintain my original score.

Reviewer_8xfLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that routing and safety should be jointly optimized in multi-LLM serving systems, extending the routing space from individual models to model–defense configurations. It makes two main contributions. First, S²-Bench systematically evaluates 10 open models under 16 combinations of pre- and post-generation defenses, using 8,943 adversarial probes derived from 559 harmful goals and 15 attack methods plus a no-attack baseline. The benchmark contains 1.43M outcomes labeled by a PAIR-rubric judge, along with realized token costs. Benign quality labels are drawn from R2Bench. Second, SafeRouter uses a frozen Qwen3-Embedding encoder and a risk gate to select configurations: low-risk queries are routed to the model with the highest predicted quality without additional defenses, while high-risk queries are assigned to the cheapest model–defense configuration whose predicted safety exceeds a calibrated threshold.

Under 7-fold cross-validation across attack methods, SafeRouter achieves an attack success rate (ASR) of 0.34% at a cost of $0.093 per 1,000 adversarial queries. Its ASR is statistically indistinguishable from that of an S²-Bench-configured Semantic Router (0.48%), while incurring only 22% of the adversarial cost, and is lower than that of the best single-defense pipeline, Self-Defense (0.95%). On benign R2Bench test queries, SafeRouter matches the undefended IRT-Router in quality and cost without introducing defense-induced refusals. Additional experiments examine generalization to disjoint attack goals, judge robustness, component ablations, and worst-case performance across attack methods.

Critical strengths

S1. The reframing is useful, and the benchmark enables systematic safety–cost analysis. The paper extends routing decisions to model–defense configurations and evaluates all 8,943 adversarial probes across ten models and sixteen defense compositions, recording the realized cost of each outcome. This comprehensive evaluation makes it possible to examine safety–cost trade-offs across configurations rather than relying on isolated defense comparisons. It also reveals findings of independent interest: Nemotron Super-120B is the most vulnerable model in the evaluated pool, and GPT4Cipher succeeds on 28% of its probes even with the input guard enabled. The benchmark is a substantial contribution and provides the strongest basis for the paper's value.

S2. The adversarial evaluation is carefully designed and supported by robustness checks. The evaluation holds out attack methods, harmful goals, and their combination from training, calibration, and threshold selection, reducing the risk of overly optimistic results from data leakage across these dimensions. Statistical significance is assessed using cluster bootstrapping at the probe, goal, and attack-method levels. In addition, the ASR judge is independent of the judges used by the defenses, and the findings are re-evaluated with two alternative judges and relaxed PAIR thresholds. The authors also appropriately acknowledge that the ASR difference between SafeRouter and the configured Semantic Router is not statistically significant.

Critical weaknesses

W1. The claim that SafeRouter costs benign users nothing is never tested on the requests where it could fail. The benign set consists of R2Bench math, code, knowledge, and multiple-choice questions, while every adversarial probe is a harmful goal, usually wrapped in a jailbreak template, so the two are easy to tell apart. The risk gate accordingly sends 4,201 of 4,203 benign test queries to the quality path, and 0.00% over-refusal and quality parity with IRT-Router then hold almost by construction. The requests that would actually test the claim are benign ones that look adversarial, and the paper itself relies on them, citing XSTest, to justify escalating flagged queries instead of refusing them, yet none appear in the evaluation. Such a request is also poorly served once flagged, because SafeRouter predicts quality only under the default defense and, by Eq. 3, its safety path selects on predicted safety and cost alone, so a flagged benign query goes to the cheapest predicted-safe cell. On adversarial traffic these cells place 67.9% of queries behind Qwen3Guard and 69.8% on Qwen3-0.6B, so a flagged benign request is likely to be refused by the guard or answered by the smallest model. Running XSTest or OR-Bench through the full pipeline and reporting the flag rate, the selected models and defenses, and the resulting quality and refusal rate would show whether the claim holds where it matters, and is the experiment most likely to change my assessment in either direction.

W2. The evidence does not show that per-query selection is what makes SafeRouter cheap, because a simpler explanation fits the same numbers and the paper does not rule it out. The simpler explanation is that SafeRouter mostly sends flagged traffic to the cheapest model behind an input guard, while the baselines it is credited with beating, Self-Defense and the configured Semantic Router, serve attacks with the models that IRT-Router or the domain policy picks for quality. SafeRouter routes 69.8% of attacks to Qwen3-0.6B, and Table 11 shows that one fixed cell of this kind, Qwen3-0.6B with Qwen3Guard and Self-Defense, applied to every adversarial query already reaches 0.25% ASR at $0.112 per 1000 queries against 0.34% at $0.093 for SafeRouter. That cell was picked on test outcomes, but it is already 3.7 times cheaper than the configured Semantic Router, which covers most of the reported 4.5× gap. Because the baselines differ from SafeRouter both in which model receives an attack and in how the defense is chosen, the 4.5× cost gap cannot be credited to selection. The control that separates the two is SafeRouter's own gate followed by one fixed cell. Table 12 reports such a control, but its 3.69% ASR at the operating point γ = 0.5 cannot be reconciled with the headline, as I explain in Q1, so this control should be rerun consistently and reported with variation across training runs, which currently span 0.22% to 0.67% ASR.

Questions for authors

Q1. Why does Table 12 report an ASR that is inconsistent with the headline result? Table 12 appears to apply a fixed model–defense configuration whenever SafeRouter's risk gate fires at γ = 0.5, while leaving unflagged queries on the low-risk path. Yet this configuration achieves 0.21% ASR when applied to all queries in Table 11, whereas the gated variant reports 3.69% ASR in Table 12. Figure 5 appears to indicate that only a small fraction of adversarial queries pass through the gate. Since the low-risk path uses the default composition and only 0.5% of adversarial queries reportedly receive it, these results seem difficult to reconcile. Could the authors report the gate's pass rate and the ASR of unflagged adversarial probes, clarify how unflagged queries are served in Table 12, and explain how the fixed configuration is selected? This discrepancy directly affects the interpretation of the control experiment in W2.

Q2. Which quality head or routing policy actually serves low-risk queries? Table 2 reports exactly the same benign quality for SafeRouter and IRT-Router (0.842), with nearly identical costs ($0.61 versus $0.608). However, Eq. 3 indicates that SafeRouter selects the model with the highest predicted quality, without explicitly accounting for cost. Could the authors clarify whether the low-risk path uses SafeRouter's own quality predictions or delegates routing to IRT-Router, and explain how the reported quality and cost results follow from the stated algorithm? This distinction is important for interpreting the benign-side comparison.

Q3. How robust are the headline results and ablations to training variability? The reported ASR of 0.34% is below the ten-run mean of 0.43 ± 0.14% and close to the configured Semantic Router's 0.48%. Which run produces the headline result, and are the ablations in Table 3 evaluated using the same run or averaged across multiple runs? The ASRs of the static-cost and fixed-model variants (0.66% and 0.41%) fall within the reported 0.22%–0.67% range of the full model across runs, making it difficult to attribute the differences to the corresponding design choices based on single-run results. Moreover, the focal-loss variant achieves the same reported ASR at lower cost, which raises questions about the necessity of the inverse-focal design. Could the authors report means and variability for the ablations across multiple runs to establish which design choices yield consistent benefits?

Q4. Do the held-out attack-method folds test generalization to genuinely distinct attack strategies? How are the 16 attack methods assigned to the seven folds? If closely related methods, such as the five role-play templates or three LLM-based rewrites, are distributed across training and test folds, the evaluation may overestimate generalization to unseen attacks. Could the authors clarify the grouping procedure and, if related methods are split across folds, evaluate a family-level split that holds out entire attack families?

Q5. How sensitive is the reported cost advantage to the assumed prices of locally served models and defenses? As discussed in W2, much of the adversarial cost reduction appears to come from applying Qwen3Guard before generation or routing requests to Qwen3-0.6B. The paper assigns these components prices of $0.02 and $0.01 per million input tokens, respectively. Since these components appear to be served locally, the cost comparison depends on the authors' pricing assumptions. Could the authors explain how these prices are derived and report a sensitivity analysis under alternative cost assumptions, including a model pool without an exceptionally cheap model? This would help establish whether the claimed cost advantage is robust beyond the specific deployment assumptions.

Q6. Can SafeRouter generalize to models or defenses that were not included in training? The safety head scores a fixed set of 160 model–defense configurations using measured S²-Bench labels. Adding a model or a pre-generation defense therefore appears to require substantial additional evaluation, followed by recalibration of the corresponding safety predictions and threshold. However, the paper does not evaluate whether safety predictions transfer to previously unseen models or defenses. Could the authors conduct a leave-one-model-out experiment using the existing ten models, or otherwise quantify the evaluation and recalibration overhead of extending the configuration pool? This would clarify how readily SafeRouter can adapt to evolving production serving systems.

4 comments
AuthorsOpenReview ↗

# Response to Reviewer 8xfL (1/3)

We thank the reviewer for recognizing that S²-Bench enables systematic safety–cost analysis and that our evaluation is carefully designed.

W1 — Does SafeRouter over-refuse benign requests that look harmful?

For safe requests that resemble harmful ones, S²-Bench's per-cell measurements locate the source of refusals, and adding such requests to the gate's training removes most of them.

W1.1 Flag rate, selected cells, refusal and quality. R2Bench reflects the ordinary traffic a serving system mostly sees, where SafeRouter adds 0.01% refusals; for the harder case, we ran XSTest and OR-Bench-Hard-1K through the full pipeline with XSTest's 3-class scoring. The gate flags 1.4% of R2Bench but 88.8% of XSTest-safe and 100% of OR-Bench-Hard. Flagged queries go to S1+S4 (51%) or S3-guard cells (46%), on Qwen3-0.6B or 1.7B for 81% of them:

SystemASR (S²-Bench)XSTest-safe refusal / qualityOR-Bench-Hard refusal / quality
Fixed S40.95%6.4% / 0.8237.4% / 0.36
Semantic Router (S²-configured)0.48%11.2% / 0.7737.9% / 0.31
SafeRouter (default gate)0.35%38.0% / 0.4281.5% / 0.08
SafeRouter (gate retrained with borderline-benign data)0.46%14.0% / 0.7446.6% / 0.32
SafeRouter (gate retrained, IRT-Router on the low-risk path)0.45%9.2% / 0.7941.0% / 0.34

Quality: the paper's judge against Gemini 2.5 Flash reference answers, excluding the 11% of prompts whose reference refuses.

W1.2 How flagged requests are served. Our measurements identify the S1 and guard cells of W1.1 as the ones that refuse such requests. As the gate never saw such requests, we add 30k OR-Bench-80K and 7.7k synthetic XSTest-style prompts to its benign class only. These prompts are disjoint from all test sets (near-duplicates of XSTest removed), following standard practice of training safety models on benign prompts that resemble harmful ones and evaluating on held-out sets [1–3]. The retrained gate flags only 10–13% of these sets. Over 5 seeds, refusal is 13.4 ± 2.0% and 43.2 ± 5.6%, while ASR is 0.51 ± 0.12%. Unflagged queries take the low-risk path, which can use another quality router in place of SafeRouter's own quality head (Q2); with IRT-Router there, the retrained SafeRouter refuses less and answers better than the Semantic Router on XSTest-safe, at lower ASR. None of this changes the main results; it shows why per-cell measurements and per-query model–defense choice matter. This experiment will be added to the appendix. We thank the reviewer for suggesting this test.

W2 — Do the savings come from per-query selection or from a cheap model?

The savings come from per-query selection: a fixed cell chosen on training folds, as a deployer would choose it, costs 2.5× at equal safety.

W2.1 The cheapest-model explanation. We ablate SafeRouter's model and defense choices and compare against fixed cells chosen on training folds; all rows use the same folds, and the fixed cell meeting the ASR target is rerun under each of the 5 seeds' gates. Fixing the model to Coder-Next costs 3.4×, so SafeRouter's savings require per-query model selection:

PolicyASR$/1k
SafeRouter (5 seeds)0.40 ± 0.10%0.092
SafeRouter, defense fixed to S3+S4 (5 seeds)0.32 ± 0.06%0.106
SafeRouter, model fixed to Coder-Next (5 seeds)0.42 ± 0.18%0.311
Fixed cell, cheapest meeting the ASR target on training folds (5 seeds' gates)0.43 ± 0.01%0.232
Fixed cell, lowest training ASR within SafeRouter's cost (seed 42)4.71%0.052

W2.2 The Table 11 cell. That cell was chosen on test outcomes. Chosen on training folds, the best fixed cell differs by fold: Qwen3-0.6B + S3+S4 in two folds, Super-120B + S1+S4 in three, Coder-Next + S1+S4 in two. At SafeRouter's cost, the best fixed cell is 12× less safe (4.71%). This control will be added to the appendix. SafeRouter also does not rely on a cheap model: without models of ≤4B parameters, it stays 3.7× cheaper than the Semantic Router (Q5).

References

[1] WildTeaming at scale: From in-the-wild jailbreaks to (adversarially) safer language models. arXiv:2406.18510, 2024.

[2] WildGuard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of LLMs. arXiv:2406.18495, 2024.

[3] FalseReject: A resource for improving contextual safety and mitigating over-refusals in LLMs via structured reasoning. arXiv:2505.08054, 2025.

AuthorsOpenReview ↗

# Response to Reviewer 8xfL (2/3)

Q1 — Why did Table 12 disagree with the headline ASR?

We apologize for the oversight: the submission did not state the calibration rule, and Table 7 only listed γ = 0.5 (the sigmoid midpoint).

Q1.1 Why the numbers differ. With the calibrated γ, only 0.44% of adversarial probes pass the gate (39 of 8,943); they are served undefended by the quality model, and 1 of them is jailbroken. The paper's 0.34% (30 of 8,943) scores every probe on the safety path; including the gate, the full pipeline reaches 0.35% (31 of 8,943), so the gate costs a single jailbreak. The revision reports this end-to-end figure.

Q1.2 Calibration rule and fixed cell. γ is the 99th percentile of the risk score on benign training traffic (a 1% false-positive rate), the same rule used for the Semantic Router baseline. The submitted Table 12 selected its fixed cell on test outcomes; the control in W2 selects it on training folds only. The revision states the rule in Sec. 3.3 and Table 8 (Table 7 of the submission) and removes the submitted Table 12.

Q2 — Which router serves low-risk queries?

Q2.1 The quality head and the submitted numbers. Eq. 3 serves them with SafeRouter's own quality head. The submitted SafeRouter entry ($0.61) was IRT-Router's figure ($0.608): we apologize for an oversight; the submitted benign column was filled in during writing from an earlier experiment, and SafeRouter's row reused the IRT-Router low-risk-path value instead of its own quality head. The revision completes the column: every row is scored per query on the full benign judgements, with each system's own routing and pipeline cost. This moves IRT-Router from $0.608 to $0.710 and SafeRouter to $0.749; quality moves by more than 0.02 only for Paraphrase (S2), the configured Semantic Router, and the two LLM-as-router rows, now scored on the models they select. Refusal rates move by under one point, the adversarial columns are untouched, and no conclusion of the paper changes.

Q2.2 Revised numbers and other quality routers. Since the safety path only handles the 1.4% of benign queries the gate flags, another quality router can be plugged in. Evaluated per query on the 4,203 R2Bench test queries (blocks score 0), the quality head matches IRT-Router's quality at a slightly higher cost, because it sends 74% of unflagged queries to Coder-Next where IRT-Router sends 37% to the cheaper Super-120B; plugging IRT-Router in inherits its cost and leaves the adversarial side unchanged (0.35% at $0.094):

Low-risk pathQualityOver-refusal$/1k
IRT-Router alone (reference)0.8360.00%0.710
SafeRouter, own quality head0.8360.01%0.749
SafeRouter + IRT-Router0.8360.01%0.707

The revision adds this configuration as a row of Table 2. The same holds with the retrained gate of W1: on R2Bench, 0.837 quality, 0.00% over-refusal and $0.710/1k (cells recorded for 99.4% of queries), with 0.45% ASR at $0.111/1k against 0.46% at $0.108 with the quality head.

Q3 — Are the results and ablations stable across training seeds?

Q3.1 The headline run. The submitted run is seed 42, whose full pipeline gives 0.35% (Q1.1); the paper's 0.34% scores every probe on the safety path. Over 5 retrained seeds SafeRouter gives 0.40 ± 0.10%, consistent with the paper's 10-run mean (0.43 ± 0.14%) and comparable to the Semantic Router (0.48%) at 4.4× lower cost.

Q3.2 Ablations over seeds. The submitted ablations were single runs (seed 42). We retrain every variant with the same 5 seeds. Standard focal loss doubles ASR (0.86% vs 0.40%) and is far less stable (sd 0.46 vs 0.10) for a slightly lower cost; fixing the model keeps ASR but costs 3.4×; the cost head lowers cost by 13% relative to no cost head, and neither it nor static costs at selection change ASR beyond noise. This supports the inverse-focal loss and per-query model selection, with cost prediction mainly saving cost; Table 4 (Table 3 of the submission) will report these means:

VariantASR$/1k
SafeRouter0.40 ± 0.10%0.092 ± 0.012
static cost at selection0.35 ± 0.05%0.097 ± 0.018
no cost head0.37 ± 0.07%0.106 ± 0.017
standard focal loss0.86 ± 0.46%0.087 ± 0.011
no risk head3.30 ± 0.32%0.057 ± 0.007
fixed model (Coder-Next)0.42 ± 0.18%0.311 ± 0.002
AuthorsOpenReview ↗

# Response to Reviewer 8xfL (3/3)

Q4 — Does SafeRouter generalize to unseen attack families?

Q4.1 Fold assignment. The 16 methods are split into 7 folds of 2–3 at random (fixed seed). Holding out attack methods is a standard test of generalization to unseen attacks [14, 16], since new jailbreaks are largely recombinations of techniques from earlier attacks [16]. Related methods therefore fall on both sides: the five role-play templates span four folds, and PastTense and FutureTense are in different folds.

Q4.2 Family-level split. Because a held-out method can still share a family with training methods, and holding out whole groups is stricter [15], we also hold out one of 7 attack families per fold (5 seeds, as in all other experiments). Even under this stricter split, ASR is 0.58 ± 0.15% at $0.122 ± 0.017/1k, 3.4× lower cost than the Semantic Router (0.49%, $0.412/1k), and encoding is the hardest unseen family for every system (Semantic Router 2.15%, Fixed S4 2.24%):

Held-out familyrole-playLLM rewritecontextiterativerandom searchencodingno attack
ASR (mean of 5 seeds)0.53%0.88%0.47%0.45%0.00%2.08%0.18%

The family-level split will be added to the appendix.

Q5 — Does the cost advantage depend on the assumed prices?

Our prices follow standard practice, and the advantage holds when the cheap components are priced higher.

Q5.1 How prices are derived. Following R2-Router [4] and LLMRouterBench [5], per-token prices are those of the cheapest OpenRouter provider for each checkpoint (April–May 2026). Qwen3-0.6B/1.7B and Qwen3Guard-0.6B had no reliable listing; as in RouterArena [6], which prices unlisted models by size-based tiers, we set their input prices below that of the cheapest listed small Qwen3 (Qwen3-4B, $0.03/M tokens).

Q5.2 Alternative cost assumptions. S²-Bench records per-call token counts, so costs can be re-priced under each sheet without new generations; we re-select τ* and re-configure the Semantic Router, without retraining SafeRouter. Pricing the small models and the guard higher leaves SafeRouter below 0.75% ASR and 3.0–4.4× cheaper than the Semantic Router:

Price sheetSafeRouter ASR, $/1kAdvantage over Semantic Router
paper prices0.35%, 0.0944.4×
current listings (0.6B/1.7B at 10×)0.40%, 0.0954.4×
0.6B/1.7B/4B at the Qwen3-8B rate0.74%, 0.1093.8×
Qwen3Guard at 10×0.49%, 0.1433.0×

Q5.3 A pool without an exceptionally cheap model. Even without any model of 4B or fewer parameters, SafeRouter stays 3.7× cheaper than the Semantic Router at lower ASR (0.41% at $0.111/1k vs 0.48% at $0.415). The re-pricing analysis will be added to the appendix.

Q6 — How costly is adding a new model or defense?

As with quality routers trained on per-model outcomes [5, 13], a new model or defense needs labels and one retrain, and both are cheap. With any one model removed, the 9-model routers reach 0.28–0.92% ASR, so no single model carries SafeRouter's safety. Safety predictions for a new model do not transfer from a few labelled probes (median AUROC 0.53 vs 0.76 when fully trained). Labelling one model on all probes and compositions costs $10–31, and one defense for all 10 models at most $64; either is followed by one retrain of about 22 GPU-minutes. We added three new models this way (Response to 9sxU, W2/Q2). The leave-one-model-out results will be added to the appendix.

References

[4] R2-Router: A new paradigm for LLM routing with reasoning. ICML 2026 (arXiv:2602.02823).

[5] LLMRouterBench: A massive benchmark and unified framework for LLM routing. arXiv:2601.07206, 2026.

[6] RouterArena: An open platform for comprehensive comparison of LLM routers. ICLR 2026 (arXiv:2510.00202).

[13] RouterBench: A benchmark for multi-LLM routing system. arXiv:2403.12031, 2024.

[14] Robust prompt optimization for defending language models against jailbreaking attacks. NeurIPS 2024.

[15] When benchmarks lie: Evaluating malicious prompt classifiers under true distribution shift. AIWILD Workshop @ ICLR 2026 (arXiv:2602.14161).

[16] Adversarial Déjà Vu: Jailbreak dictionary learning for stronger generalization to unseen attacks. ICLR 2026.

Thank you for the detailed response. It helps clarify two specific issues I raised: the discrepancy in Table 12 resulted from an omitted calibration detail, and the benign cost reported for SafeRouter was inadvertently taken from IRT-Router. I also appreciate the additional experiments on borderline-benign prompts, multiple random seeds, and fixed-cell baselines, which provide useful further evidence.

These additional results are helpful, although some concerns regarding benign behavior remain. In particular, the original gate still exhibits relatively high refusal rates on XSTest-safe and OR-Bench-Hard, while the improved results rely on retraining the gate with additional borderline-benign data and modifying the quality router. I also remain somewhat uncertain about the extent to which the reported cost savings can be attributed specifically to per-query selection. Although the fixed-cell comparison on the training fold provides some additional support, I believe further evidence in the main evaluation would help establish this point more clearly.

Overall, the rebuttal has improved my understanding of the paper and addressed several of my questions. However, I am not yet fully convinced that the remaining concerns regarding benign behavior and the attribution of cost savings have been sufficiently addressed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.