acceptodds
Withdrawn from ICLR 2027

PROOF: Mixed-Traffic Jailbreak Defense via Profile Routing in Frozen LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Safety controls must prevent harmful completions while preserving legitimate assistance. We introduce , which learns to select among fixed safety profiles of a frozen language model. During training, each input is evaluated under every profile using a cost appropriate to its type: unsafe completion, unnecessary refusal, or capability-probe loss. When several profiles share the lowest cost, a tie-aware target retains a preferred action and assigns probability to the other minima. This preserves equally rated alternatives that hard-label imitation would penalize. We study both routing before response generation and selection among profile-generated candidates. On a fixed 200-example outcome table, ten tie-breaking assignments and five paired controller seeds yield a 7.69% reduction in sensitivity to tie-breaking relative to hard-label training, with no statistically resolved change in mean selection regret. Across eight profile-menu conditions, safety differences depend on the available actions. Serving measurements further show that candidate conditioning costs more than routing from prompt features. The results identify a benefit of tie-aware supervision for label sensitivity and distinguish it from menu-dependent safety gains and inference cost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large language model, Jailbreak defense, LLM safety

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_X967OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes PROOF, a profile-routing framework for balancing safety and utility in frozen LLMs. PROOF composes a fixed set of LoRA adapters into multiple safety profiles and trains an MLP controller to route each input to the most appropriate one. The central contribution is a tie-aware supervision objective: rather than assigning a single hard label, it spreads probability mass over all profiles that tie for the lowest observed cost.

Critical strengths

The paper is well structured, addresses a critical problem in LLM safety, the trade-off between over-refusal and harmful behavior. The distinction between prompt-only routing and candidate-conditioned selection is clearly explained, with inference costs measured for both settings. The experimental analysis is properly designed and analyzes both positive and negative results.

Critical weaknesses

The proposed soft-label like objective has limited novelty beyond existing soft-label methods.

The comparison against hard-label cross-entropy alone is insufficient to establish its effectiveness. Additional baselines are needed to determine the effectiveness of the proposed method.

As the method designed to pre-combine the profile, numbers of K files limits the combinations or flexibility of the profile. Why the author did not choose to let the MLP predict a continuous value as the mixing weight and flexiblely combining both LoRA for each request? In table 3, the advantage of the proposed method gradually fall behind the CE-based method, which increasingly raise my concern about the efficacy of the propose PROOF.

Questions for authors

Though AI usage is disclosed, current presentation is not clear or formal enough for ICLR paper. Specifically for Figure 1, at the bottom of the figure, there is "select from existing candidates ", the presentation is even not clear to identify "a" or "e".

Reviewer_6zqsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper treats jailbreak defense as per-request selection from a fixed menu of “profiles.” Each profile is a weighted composition of a safety LoRA and a preservation LoRA on a frozen Qwen2.5-7B-Instruct. For each training input, every profile is scored with a type-specific cost: unsafe completion for harmful requests, refusal for benign ones, and teacher-forced loss for capability probes. A lightweight MLP is then trained to pick the lowest-cost profile and is used at inference, either before generation (prompt-only or prompt-forward features) or by choosing among responses pre-generated under every profile (candidate-conditioned). The main methodological contribution is a tie-aware soft target. When several profiles share the minimum cost, it keeps a preferred representative (0.8) and spreads the remaining mass over the other tied profiles. Experiments cover sensitivity to tie-breaking, the effect of menu composition, serving cost, and robustness to fixed request rewrites. All are single-turn.

Critical strengths
  1. The mixed-traffic framing is sensible. Harmful, benign-edge and capability requests are scored under one selection policy with type-specific costs, and building the menu is cleanly separated from learning to select from it.

  2. The reporting is honest. The authors state that their pre-specified success criterion was not met (7.69% vs. the targeted 20% reduction, with no resolved regret gain), report menus where tie-aware training lowers DSR, and show the prompt-only router failing under rewrites.

Critical weaknesses
  1. Missing baselines. The unmodified backbone is never evaluated. Static profiles appear only in the timing study (no safety or utility numbers) and, for Safety-first alone, in the harmful-only rewrite study, where it outperforms the learned routers (68.1% vs. 34.7% worst-of-8 DSR). The tie-resolution and menu studies, the only mixed-traffic evaluations, include no static or base-model rows and no oracle upper bound. Direct CE is an ablation of the training target, not a defense baseline, and no external defenses (e.g., SafeDecoding, Llama Guard as a filter) are compared. As a result, the paper does not show that learned routing improves on simply using one profile.

  2. The gains do not justify the added complexity and serving cost. i. The headline result is a 7.69% reduction in tie-breaking disagreement with unchanged regret, which falls short of the authors’ own prespecified 20% criterion. ii. Safety effects across menus are mixed (−3.4 to +3.4 DSR points). iii. Capability is measured only by teacher-forced probe loss (about 0.75 to 0.78 across menus), with no downstream accuracy. iv. The evaluation sets are small (30 benign and 20 probe inputs in the tie study; BRR in the menu study appears to rest on about 22 benign examples). v. Against these modest and hard-to-interpret gains, the method adds a menu of adapter compositions, an offline scoring pipeline, and a learned selector. At batch size 8, prompt-only routing takes about 2× the per-prompt time of static profiles, and candidate-conditioned selection about 6×.

  3. Only static jailbreak prompts; no adaptive attacker. All attack inputs are pre-generated: benchmark prompts plus a fixed set of eight rewrites, with no feedback-driven or adaptive attack (e.g., GCG, PAIR, TAP). This matters especially here, because the router adds a new decision point an attacker can target, for example by steering a request toward the least safe profile. Even the fixed rewrites already cut the prompt-only router’s worst-case DSR roughly in half relative to static Safety-first, and the prompt-forward and candidate-conditioned selectors are not tested under rewrites at all.

Questions for authors
  1. Can you report DSR, BRR and capability for the unmodified backbone, each static profile, and the oracle selector on the menu-study validation set? Does learned routing improve on the best static profile on mixed traffic?

  2. Given that fixed rewrites already halve the prompt-only router’s worst-case DSR, how do the routers hold up against an adaptive attacker targeting the router itself, for example optimizing prompts to be routed to Preservation-first? Do the prompt-forward and candidate-conditioned selectors share this vulnerability?

  3. All evaluations are single-turn. How would per-request profile selection extend to multi-turn conversations, where harmful intent may be spread across turns?

Reviewer_BECxOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PROOF, which selects among fixed safety profiles of a frozen language model. The profiles combine safety and preservation adapters with different weights. The main change is a training target that assigns some probability to other profiles with the same minimum cost. The paper studies three selectors with different input features. Experiments on Qwen2.5-7B-Instruct examine sensitivity to tie-breaking, profile choices, serving time, and request rewrites. Tie-aware training reduces sensitivity to tie-breaking, but does not show a significant reduction in mean selection regret.

Critical strengths
  1. The paper reports both positive and negative results clearly. It shows confidence intervals, reports safety losses under some profile menus, and acknowledges that lower disagreement does not imply better responses.
Critical weaknesses
  1. The training labels and actual actions do not always match. The study merges two profiles even though their costs can differ. A label for Utility-leaning leads to running Preservation-first, which may have a higher cost. This does not follow the stated rule of training only on the lowest-cost choices. Each profile should be tested as a separate action.
  2. The experiments do not show what benefit comes with the lower safety. Under eight rewrites, PROOF achieves 34.72% defense success. The static safety profile reaches 68.06%, and LlamaGuard routing reaches 66.67%. This test does not report benign refusal or task accuracy, so it is unclear what PROOF gains in return. The mixed-traffic test uses only 192 validation examples. Its safety gains also vary across profile menus.
Questions for authors
  1. Why are Utility-leaning and Preservation-first merged when their costs can differ? Do the main results hold when each profile remains a separate action?
  2. How does PROOF compare with assigning equal probability to all minimum-cost profiles? Why is a preferred representative still needed?
  3. On an independent mixed-traffic test set, does PROOF offer a better safety–utility trade-off than static profiles or LlamaGuard routing? Please report safety, benign refusal, and capability results together.
  4. Does PROOF preserve actual task accuracy? The current capability loss includes prompt tokens and does not directly measure answer correctness.

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

Related papers

Discussion (0)

Sign in to comment.