acceptodds
Withdrawn from ICLR 2027

FedPTR: Federated Soft-Prompt Routing under Corrupted Evaluation Feedback

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) are deployed with libraries of reusable prompts, creating an online routing problem in which each query must be assigned to an appropriate prompt. In collaborative deployments, federated sharing improves routing by pooling cross-client evidence, but also creates a new vulnerability. Manipulated feedback received by honest clients can propagate through protocol-compliant updates and affect clients with clean evaluation channels. We call this cross-client evaluation contamination. This creates a tension between preserving useful sharing and limiting contamination, while standard uncertainty-based influence control cannot distinguish confidence derived locally from that gained through federation. To address this tension, we introduce FedPTR, a corruption-robust framework for federated online soft-prompt routing that controls feedback influence before evaluation scores are revealed. FedPTR contrasts local and federated uncertainty to quantify dependence on cross-client evidence and augments standard uncertainty clipping with federation-aware attenuation. We distinguish manipulated evaluator scores from trusted response utility and establish affected-to-unaffected prediction influence bounds and trusted-utility regret guarantees for both the full federation and unaffected clients. Extensive experiments show that FedPTR limits cross-client contamination while preserving the benefits of federated routing.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
federated soft-prompting routing, corrupted evaluation feedback, online robust detector

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

This paper proposes FedPTR, a framework for federated online soft-prompt routing under corrupted evaluation feedback. The method aims to reduce the influence of manipulated feedback before evaluation scores are observed. It compares local and federated uncertainty to measure dependence on cross-client information and adds a federation-aware attenuation term to standard uncertainty clipping. The paper also provides theoretical guarantees and experiments showing how corrupted feedback can affect clients with clean evaluation channels.

Critical strengths

The paper is well-written, and the proposed method is easy to follow. The problem is interesting and relevant to federated online learning systems that use automatic evaluators.

The paper clearly demonstrates how corrupted evaluation feedback can affect other clients through shared federated updates. The matched local and federated comparisons help illustrate this issue.

The theoretical analysis is clearly presented, and the assumptions are explicitly stated.

I appreciate that the authors separate trusted utility from observed evaluator scores. This makes the experimental results more meaningful under corrupted feedback.

Critical weaknesses
  1. Novelty of the proposed method: The proposed method is interesting, but I have some concerns about the extent of its methodological novelty. FedPTR builds on the uncertainty-weighted estimator of He et al. (2022), with the main modification being an additional denominator based on the difference between local and federated uncertainty.

While this extension is reasonable and the theoretical analysis provides additional insights, I would like to better understand how much of the improvement comes specifically from this new component.

Also, the connection to LLMs and soft-prompt routing seems mainly to come from the application setting. The routing algorithm operates on frozen query–prompt representations, and it is not entirely clear which parts of the method are specific to LLM prompt routing.

  1. Possible inconsistency between Figure 6 and Table 1: Based on my understanding, there seems to be an inconsistency between Figure 6 and Table 1.

Section C.7 mentions that the utility results in Figure 6 are for T5-Flan-base on AG News under attack, which appears to be the same model and dataset as the first column of Table 1. However, the reported values are quite different. For example, FedPTR achieves around 0.847 in Figure 6 compared with 0.785 in Table 1, while Local Weighted achieves around 0.828 in Figure 6 compared with 0.613 in Table 1. These differences seem significant, especially for Local Weighted. It would be helpful if the authors could clarify whether the two results use different corruption budgets, evaluation settings, or other experimental conditions.

  1. Missing federated uncertainty-weighted baseline in the main comparison: The paper describes a 2×2 comparison between local and federated methods and weighted and unweighted methods. However, Table 1 does not separately include a federated version of the uncertainty-weighted method of He et al. (2022) without the proposed attenuation term. Figure 6 includes a comparison with γ = 0, which removes the federation-aware denominator while keeping standard uncertainty clipping. However, this comparison is not reported across the datasets and models in Table 1. Including this baseline in the main results would help distinguish the benefit of applying existing robust weighting in a federated setting from the additional improvement provided by the proposed denominator.

  2. Minor presentation issues: There are some typos in the figures, including "rount," "AItered," and "Local–federatedpre-feedback influence contro" in Figures 1 and 3.

Also, some important algorithmic and implementation details are provided only in the appendix. Including a more complete description of the algorithm in the main paper would improve clarity.

Questions for authors

Figure 6 and Table 1: Could the authors clarify why the attacked-utility values for FedPTR and Local Weighted differ substantially between Figure 6 and Table 1 for T5-Flan-base on AG News? Are the corruption budgets, evaluation metrics, and experimental settings the same?

Federated weighted baseline: Could the authors report the performance of federated uncertainty weighting with γ = 0 across the settings in Table 1? This would help clarify how much improvement comes from standard uncertainty clipping and how much comes from the proposed federation-aware attenuation term.

Methodological contribution: Could the authors explain more clearly how the proposed uncertainty-gap term differs from existing uncertainty-weighted approaches beyond its use in a federated setting?

Connection to LLMs: Which parts of FedPTR specifically depend on LLMs or soft-prompt routing? Would the same method apply directly to general federated contextual-bandit problems with frozen feature representations?

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

This paper studies federated online soft-prompt routing. Clients share a library of frozen soft prompts and a frozen LLM; routing is a contextual bandit over a frozen query–prompt feature map, with the server maintaining a shared weighted ridge estimator pooled across clients. The paper identifies a new threat: because the router learns from automatic evaluator scores, an adversary corrupting the evaluation channel of some clients can degrade routing at clients whose own evaluators are clean. Those clients stay honest and their statistics remain protocol-compliant, so corruption enters through legitimate updates. The paper calls this cross-client evaluation contamination. It proposes FedPTR, which gives each observation a weight ω = min{1, α/κ^fed} / (1 + γ·Δ^fed), committed before the score is revealed, where Δ^fed is the uncertainty reduction attributable to cross-client evidence. It states an influence bound and trusted-utility regret guarantees, and evaluates on six datasets across five backbones.

Critical strengths

1. The threat is novel and the defense is well matched to it. Corruption entering through the evaluator channel at honest, protocol-compliant clients differs genuinely from federated poisoning: there is no malformed update to detect, so Byzantine-robust aggregation has nothing to filter. Section 3 is disciplined about scope, excluding Byzantine clients, malicious servers, prompt-library poisoning and response tampering. Because ω is committed before the score is revealed, the defense cannot be gamed by the corrupted score itself and needs neither attack detection nor knowledge of the corruption budget. Section 4's motivating study is the right design: matched local and federated routers on identical query streams, prompt pools and corruption schedules, so the extra degradation at unaffected clients can only come from sharing.

2. The paper states theoretical guarantees and supports them with experiments. It gives an influence bound and trusted-utility regret guarantees for both the full federation and the unaffected clients, and the empirical work is broad: six datasets, five backbones, a 2 × 2 local/federated × weighted/unweighted design separating sharing from weighting, and a real LLM-evaluator attack alongside synthetic corruption. Reporting on seeds {71, 89, 107} while choosing every configuration on a disjoint seed is good practice.

Critical weaknesses

1. The motivation for the setting is not established. The paper assumes pooling cross-client evidence improves routing, but never argues why evidence about which frozen prompt suits one client's queries should transfer to a client whose query distribution and prompt subset differ — and Section 3 explicitly allows both to be heterogeneous, which makes the premise less obvious, not more. Nor is any concrete deployment described in which several organizations share a prompt library, route online, and rely on automatic evaluators. Every contribution is conditional on that setting being real, so the absence of either argument or example weakens the whole paper.

2. The paper's own contribution is not ablated where it matters, and where it is measured the gain sits inside the error bar. Section 5.2 states that γ = 0 "recovers standard uncertainty-based clipping," so the federation-aware denominator is what is new. Table 1, which carries every headline comparison, has no γ = 0 row — so a reader cannot tell how much of FedPTR's 0.15–0.22 advantage over Fed LinUCB comes from the new term rather than from standard clipping applied under federation. The only place γ = 0 appears is Figure 6a, on a different configuration, where attack utility moves from 0.826 to 0.847 at γ = 0.25 — about 0.02 — with γ = 0 at or below the Local Weighted baseline (0.828) and a band spanning roughly 0.78–0.86, so the effect is unresolved at three seeds. Figure 6b compounds this: mean weight moves 0.003 → 0.078 → 0.001 across γ and "Attenuation" ~0.05 → 0.95 → ~0.05, yet under Equation 6 the factor 1/(1 + γΔ) is exactly 1 at γ = 0 and decreases monotonically. Neither quantity is defined in the main text, so the diagnostics meant to explain the mechanism appear inconsistent with it.

3. The benchmarks are old and simple, and do not exercise the threat the paper motivates. The newest dataset is EuroSAT (2019); CIFAR-10 dates to 2009, TREC to 2002, AG News to 2015. Age aside, most are four- and ten-way classification, which matters more than it first appears: there the trusted utility is just label correctness, directly computable, so no genuine LLM judge exists whose biases an adversary would exploit. Corruption must therefore be injected synthetically rather than arising from the phenomenon motivating the paper, and the one setting with a real LLM evaluator under attack is a single dataset–backbone pair (AG News on T5-Flan-large). A threat model about manipulable evaluators should be demonstrated mainly on tasks whose quality genuinely requires a model-based judge.

4. Minor: presentation and formatting. The submission does not follow the ICLR template — it is A4 (595 × 842 pt) against the style file's US Letter (612 × 792 pt), with text stored as vector outlines rather than selectable characters, so the PDF cannot be searched or read by a screen reader. Table 1 is too small to read, Figure 6's panels are inconsistently sized, and captions are too thin for a figure to stand alone. Separately, Local LinUCB and Local Weighted agree to three decimals, standard deviation included, in three of Table 1's twelve columns.

Questions for authors
  1. Why should evidence about query–prompt fit transfer across clients with heterogeneous query distributions and prompt subsets, and is there a deployment in which organizations share a frozen prompt library, route online, and rely on automatic evaluators?
  2. Can Table 1 be extended with a γ = 0 row, separating federation-aware attenuation from standard uncertainty clipping on the main benchmarks?
  3. How are "Attenuation" and "Effective influence" in Figure 6b defined, and why are they non-monotonic in γ when Equation 6 decreases monotonically for fixed κ and Δ?
  4. Can the evaluation be extended to tasks whose quality genuinely requires a model-based judge, so that evaluator manipulation arises rather than being injected? And in the three columns where Local LinUCB and Local Weighted agree exactly, did the weighting never bind?
Reviewer_zBMpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

FedPTR studies how corrupted evaluator feedback received by honest clients can spread through a shared prompt router and harm clients with clean feedback. It learns which existing soft prompt to use while keeping the model, prompts, and query-prompt features frozen. Before observing each evaluator score, it sets an update weight that combines uncertainty-based clipping with an extra discount based on the difference between local and shared uncertainty. This aims to limit harmful feedback while retaining useful sharing. Experiments cover six language and vision datasets across twelve model/dataset settings, usually with 20 clients, 500 rounds, and three paired seeds. FedPTR reports the highest mean trusted utility in all twelve settings at the primary corruption budget; for example, it scores 0.785 versus Fed LinUCB's 0.591 on T5-Flan-base / AG News. Validation includes propagation diagnostics, corruption and client-count sweeps, an LLM-evaluator attack, and an ablation of the additional discount. The main text also states influence and regret bounds.

Critical strengths
  1. The paper studies a relevant and insteresting risk: clients can spread unreliable evaluator feedback through shared updates. Separating corrupted learning feedback from trusted evaluation utility makes this risk easier to study.
  2. The evaluation covers twelve language and vision settings, with three paired seeds and a separate development split.
Critical weaknesses
  1. The supplied PDF don't include appendix which is referenced for the prediction-influence bound, implementation details and additional experiments (Section 6, pp. 8–9). This limits verification and reproducibility.
  2. Earlier work already studies corrupted rewards in federated or cooperative bandits (Fan et al., 2023, Section 7.2, Ghaffari et al., 2025, Hu et al., 2026). Please cite these works and clarify what FedPTR adds beyond applying robust bandit methods to prompt routing.
  3. The paper claims that the new term "regulates, rather than eliminates, cross-client sharing", but it's not obvious from current results without a baseline of γ = 0 the standard uncertainty clipping.
Questions for authors
  1. Why are generation experiments limited to FLAN-T5-base? Does the method also work with decoder-only LLMs on generation tasks? Qwen2.5-7B is tested only on classification tasks.
  2. Three points in weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.